Agentic AI: From Better Answers to Entirely New Industry Economics
Why agents create value that dashboards, rules engines, RPA, predictive models, and copilots cannot create alone—and how that value compounds across 16 major industries.
Agentic AI is not important because it makes chat more intelligent. It is important because it can convert a goal into a governed sequence of decisions and actions across fragmented data, software, organizations, and time. That turns AI from a content tool into an operating capacity—and changes which services can be profitably delivered, at what volume, and to whom.
This research playbook extends the strategic thesis introduced in Agentic AI: Transforming Industries, Unleashing Explosive Growth, and Shaping a Prosperous Future. It moves from the macro opportunity into the concrete workflows, architecture, economics, business models, controls, and measures required to create value.
Explore the playbook
- Why systems of action matter
- What the research says
- Six levels of transformation
- Shared enterprise architecture
- Opportunity selection
- Healthcare & pharmacy
- Banking & financial services
- Payments & fintech
- Logistics & supply chain
- Manufacturing & industrial operations
- Agriculture & agritech
- Energy & utilities
- Real estate & construction
- E-commerce & retail
- Technology services & IT
- Customer service
- Legal services
- Education
- Government & public services
- HR & talent
- Insurance
- Agent-ready data
- Implementation model
- Value and ROI
- Risk and accountability
- Workforce redesign
- One-, three-, and five-year roadmap
- Decision framework
- Research library
From systems that record work to systems that complete work
Enterprises have spent decades building systems of record: EHRs, core banking systems, ERPs, transportation management platforms, claims systems, student systems, CRMs, and government case platforms. They preserve transactions, enforce schemas, and expose approved functions. They are essential—but they generally wait for a person to understand the situation, decide what matters, move between applications, and close the loop.
A system of action sits above those records. It can notice an event, assemble relevant context, interpret documents and policy, plan the work, call approved tools, ask for human authority when required, verify the resulting state, and continue monitoring until the outcome is complete. The authoritative record does not disappear; the agent becomes the adaptive coordination layer around it.
What Agentic AI contributes that prior technologies do not contribute alone
| Technology | What it does extremely well | Where it reaches a ceiling | Best role in an agentic system |
|---|---|---|---|
| BI and dashboards | Aggregate history, expose trends, and support human analysis. | Insight stops at the screen; a person still investigates and executes. | Supply measures, exceptions, and outcome evidence. |
| Rules engines | Execute stable policy and calculations consistently. | Rules become brittle when inputs are ambiguous or exceptions multiply. | Remain the authority for eligibility, limits, pricing, safety, and compliance logic. |
| RPA / workflow automation | Move structured data through a known path at high volume. | Interface changes, missing fields, novel documents, and branching cases create maintenance-heavy path explosion. | Perform tightly bounded, deterministic actions selected by an agent. |
| Predictive machine learning | Score risk, demand, failure, propensity, and anomalies. | A score does not gather evidence, negotiate constraints, or resolve the case. | Provide specialized predictions to a broader plan. |
| Generative copilot | Summarize, draft, search, and improve an individual’s task. | The human remains the workflow engine and moves every case forward. | Support human judgment and create reviewable artifacts. |
| Agentic AI | Plan and adapt across unstructured evidence, tools, people, and long-running state. | Probabilistic behavior, cost, latency, and failure risk make unconstrained autonomy unsafe. | Coordinate the system; delegate stable work to deterministic components and consequential judgment to accountable people. |
1. It turns unstructured evidence into action
Contracts, notes, images, calls, regulations, and email contain operating facts that traditional workflow software cannot reliably consume without templates and manual entry.
2. It bridges disconnected systems
An agent can choose among APIs, search, documents, and governed computer use, reducing the human “swivel-chair” work between systems that were never designed to cooperate.
3. It handles the long tail
Traditional automation is strongest on the happy path. Agents can interpret why a case deviates, gather missing context, select a recovery path, and escalate when confidence or authority is insufficient.
4. It closes the loop
Prediction becomes observation → decision → action → verification → learning. Value is measured in resolved cases, not the quality of an answer in isolation.
5. It makes personalization economical
Every patient, customer, farm, claim, student, contract, or asset can receive a contextual plan without creating a separate hard-coded workflow for each one.
6. It separates throughput from headcount
Once the workflow is governed and evaluated, capacity can scale by software economics—allowing more cases, smaller transactions, longer service hours, and offerings that human-only delivery could not profitably support.
A concrete example: denied medical authorization
A dashboard can display the denial. A rules engine can check a code. RPA can copy fields. A predictive model can estimate approval likelihood. A copilot can draft an appeal. An agentic workflow can detect the denial, interpret the reason, retrieve the payer’s current rule, assemble the clinical evidence, identify what is missing, ask the clinician to approve a source-grounded resubmission, transmit it, monitor the response, notify the patient, and update the EHR. Every predecessor technology still contributes—but the agent coordinates them around the outcome.
What current research says about adoption, impact, and the scale gap
The evidence does not support a story of universal autonomous enterprises. It supports a more valuable conclusion: agentic systems are beginning to deliver material workflow gains, but the organizations capturing enterprise value are those redesigning work, building proprietary evaluation loops, and treating governance as operating infrastructure.
The investor and operator lenses—combined
Andreessen Horowitz: automate the glue work
Computer-use agents expand the reachable workflow surface by operating legacy software as people do. The same analysis cautions that present systems still resemble advanced RPA more than universally capable autonomous coworkers.
Bessemer: vertical systems of action
Vertical AI is reaching language-heavy, multimodal, service-intensive work that traditional SaaS did not solve. Durable advantage comes from domain workflow, private evaluation data, permissions, and traceable outcomes.
General Catalyst: transformation in the operating environment
Its healthcare work with AWS emphasizes cloud and data modernization, real clinical deployment, and outcome-driven ecosystem collaboration—not isolated demonstrations.
Khosla Ventures: scalable intelligence as capacity
Vinod Khosla frames AI as an intellectual analogue to engines multiplying muscle power. This is the abundance thesis: dramatically lower the marginal cost of expertise while making explicit policy choices about safety and distribution.
Blackstone and KKR: value in the real economy
Private-market research directs attention beyond model vendors toward physical businesses, proprietary data embedded in workflows, operational leverage, and the infrastructure constraints of power, compute, and connectivity.
The combined conclusion
The model is not the moat. The compounding assets are workflow access, trusted context, outcome feedback, exception history, tool integrations, permission design, and the distribution channel through which completed work is delivered.
The six levels of Agentic AI transformation
Autonomy is not a feature toggle. An agent should earn a wider operating license only when its performance, observability, failure containment, and business economics meet predefined thresholds. The right destination differs by workflow.
| Level | Operating form | What the system does | Human authority | Value unlocked | Evidence required to advance |
|---|---|---|---|---|---|
| Level 0 | Insight | Searches, summarizes, predicts, or displays; no case ownership. | Human interprets and acts. | Faster information access. | Grounded accuracy, retrieval quality, user adoption. |
| Level 1 | Task agent | Completes one bounded unit such as extraction, triage, reconciliation, or draft preparation. | Human owns the broader workflow. | Reduced task time and queue pressure. | Task precision, stable latency/cost, known failure modes. |
| Level 2 | Domain agent | Applies specialized policy and tools within one function; prepares recommendations and actions. | Human approves consequential output. | Expert capacity and consistent preparation. | Policy adherence, source traceability, exception coverage. |
| Level 3 | Workflow agent | Owns a bounded case across multiple systems from trigger through verified completion. | Human approves defined gates and novel exceptions. | Cycle-time compression and higher throughput. | End-to-end completion, idempotency, recovery, customer outcome. |
| Level 4 | Multi-agent operation | Coordinates specialist agents, teams, counterparties, approvals, and long-running state. | Humans supervise risk, policy, and escalations. | Cross-functional operating leverage and service scale. | System-level evaluation, conflict control, security, workload economics. |
| Level 5 | Agentic business model | Runs a material value stream and enables a service, market, or product that was previously uneconomical. | Humans set objectives, law, risk appetite, and exception authority. | New revenue pools, granular personalization, machine-speed markets. | Durable unit economics, stakeholder trust, regulatory fit, controllable externalities. |
Practical rule: deploy the simplest level that can create the outcome. A Level 1 agent resolving a costly, frequent task can outperform a complex multi-agent demonstration. BCG’s graduated-autonomy concept similarly begins with shadow mode, moves through supervised execution, and grants guided autonomy only after demonstrated performance.
The shared enterprise Agentic AI architecture
The diagram comes first because architecture is a system of relationships, not a shopping list. The agentic layer coordinates reasoning and action; deterministic services enforce hard boundaries; systems of record remain authoritative; identity, evaluation, human control, and audit cross every layer.
Plan with models; constrain with software
Use models to interpret ambiguity, retrieve context, generate options, and adapt the plan. Use code, policy engines, optimizers, schemas, and transaction controls for calculations and boundaries that must be repeatable.
Give each agent a real identity
Every tool call should carry a short-lived, least-privilege identity, delegated purpose, case ID, data scope, transaction limit, and accountable owner. Read, prepare, approve, and execute are different permissions.
Persist case state—not unlimited memory
Long-running work needs checkpoints, provenance, retention, correction, and deletion. Memory should store what the workflow is allowed to remember, not everything the model happened to encounter.
Verify the resulting world
A successful tool call is not the same as a successful outcome. The verifier checks source evidence, calculations, permissions, policy, completeness, and the authoritative system state after the action.
How to identify a genuinely agentic opportunity
A workflow is a strong candidate when value is meaningful, volume is sufficient, the work has a recognizable completion state, inputs are substantially digital, judgment repeats, exceptions are expensive, and an accountable owner already exists. Agentic AI is a poor fit when the process is undefined, the outcome cannot be evaluated, the action is irreversible without strong controls, or a simple rule can solve it.
| Dimension | High-potential signal | Warning sign | Question to quantify |
|---|---|---|---|
| Coordination burden | Several systems, teams, documents, and follow-ups per case. | One stable screen and one deterministic decision. | How many handoffs and wait states exist? |
| Variation | A common objective with a long tail of legitimate exceptions. | No agreed normal path or policy. | What share of cases leaves the happy path? |
| Evidence | Outcome can be checked against source records and resulting state. | Correctness is subjective and consequences surface months later. | What would an independent verifier inspect? |
| Economics | High volume, delay cost, rework, loss, or unmet demand. | Low frequency and cheaper manual completion. | What is the fully loaded cost and value per completed case? |
| Control | Actions can be staged, limited, monitored, reversed, or appealed. | Agent would receive broad credentials or direct safety-critical control. | What is the worst plausible action and its containment? |
Healthcare and pharmacy
Healthcare fragmentation is structural: providers, payers, pharmacies, labs, imaging centers, post-acute organizations, life-sciences companies, and patients operate different systems and incentives. A clinical episode may be medically straightforward while its administrative pathway crosses dozens of messages, portals, documents, benefits rules, and handoffs. That coordination burden consumes scarce clinical capacity and makes small gaps become delays, denials, nonadherence, and avoidable deterioration.
Disconnected records, phone and fax queues, manual chart review, episodic outreach, repeated identity and benefit checks.
A patient-specific case layer assembles evidence, coordinates organizations, maintains longitudinal state, and routes decisions to the licensed professional.
More episodes per team, faster access, fewer preventable denials, continuous adherence support, and earlier intervention without proportional administrative hiring.
Where agents can act
- Access and revenue: eligibility, referral, scheduling, waitlists, prior authorization, claim preparation, denial recovery, and payment reconciliation.
- Care continuity: assemble the longitudinal record, track results, coordinate discharge, pharmacy, transportation, home services, and multilingual follow-up.
- Pharmacy: medication reconciliation, refill barriers, formulary alternatives, adherence, safety-event routing, and pharmacist escalation.
- Life sciences: trial matching, site operations, protocol-deviation monitoring, safety-report reconciliation, and regulatory evidence assembly.
New businesses enabled
Continuous-care operations as a service
Sell a per-member or outcome-based service that continuously closes screenings, medication, referral, and post-discharge gaps across payers and providers—economically serving populations too costly for nurse-only outreach.
Virtual specialty access network
Combine asynchronous specialist oversight with agents that prepare records, manage protocols, coordinate diagnostics, and monitor follow-through—expanding scarce expertise to rural and underserved markets without pretending the agent is the clinician.
Evidence and constraint: the CMS interoperability and prior-authorization rule reinforces structured exchange, reasons, and status transparency. General Catalyst and AWS emphasize data modernization and real clinical deployment as the path from demonstration to outcome.
Banking and financial services
Banking combines high-volume information work with strict obligations to identify customers, understand risk, explain decisions, protect assets, and preserve evidence. The operational problem is not a shortage of databases; it is the labor of reconciling documents, entities, transactions, policies, external data, exceptions, and approvals across front, middle, and back offices.
Siloed onboarding, periodic reviews, analyst-built credit packages, alert queues, duplicate evidence requests, and manual closing conditions.
A persistent customer or deal case coordinates identity, evidence, policy, risk specialists, approvals, documents, and servicing events.
Faster decisions, lower case cost, higher small-business coverage, better evidence consistency, more timely risk response, and growth without equal operations growth.
Where agents can act
- Onboarding and financial crime: collect evidence, resolve entity mismatches, enrich alerts, link parties, prepare narratives, and route effective challenge.
- Lending: assemble applications, verify figures, calculate covenants, identify policy exceptions, manage conditions, and coordinate closing.
- Servicing and collections: interpret hardship, identify eligible options, manage promises, prepare compliant communications, and escalate vulnerable customers.
- Treasury, markets, and wealth: investigate breaks, monitor limits, explain liquidity, prepare advisor views, and track actions within mandates.
New businesses enabled
Always-on SME finance desk
Offer small businesses continuous cash-flow forecasting, covenant monitoring, receivables intervention, credit readiness, and funding-option coordination at a price point previously reserved for larger companies with finance teams.
Continuous underwriting relationships
Replace episodic applications with permissioned, ongoing evidence refresh and proactive facility adjustment—allowing lenders to serve thinner-file and rapidly changing businesses while preserving human credit authority.
Control context: the Federal Reserve, OCC, and FDIC’s 2026 model-risk guidance centers risk-based development, validation, monitoring, governance, and third-party oversight. Agent inventory, intended-use boundaries, change control, effective challenge, and traceable explanations are operating requirements, not documentation afterthoughts.
Payments and fintech
Payments are becoming the execution layer for both people and software agents. The opportunity extends beyond cheaper fraud review: future commerce requires verifiable agent identity, delegated authority, intent-bound credentials, transaction limits, dispute rights, and proof that an action matched the user’s objective. At the same time, merchants still reconcile fragmented gateways, processors, bank accounts, ledgers, fees, refunds, and settlement files.
Batch reconciliation, static routing, manual dispute evidence, merchant onboarding queues, brittle AP/AR processes, and fragmented consent.
Intent and authority travel with the transaction while risk, routing, evidence, ledger, and customer-resolution agents coordinate in real time.
Higher authorization, lower fraud and operations cost, faster settlement resolution, machine-speed purchasing, and financial operations for smaller merchants.
Where agents can act
- Onboard merchants, validate beneficial ownership, classify business activity, and resolve missing evidence.
- Select routes using authorization probability, fees, fraud, settlement speed, customer preference, and merchant economics.
- Assemble disputes and chargebacks from network rules, receipts, device history, communications, and fulfillment evidence.
- Operate accounts payable, receivable, cash application, collections, refunds, ledger reconciliation, and cross-border exceptions.
New businesses enabled
Intent-bound agent wallets
Consumers and businesses delegate a goal, budget, category, duration, and approval rule to an agent. The wallet pays only when machine-verifiable terms are met and carries evidence for cancellation or dispute.
Autonomous finance operations for microbusinesses
Deliver invoice intake, approval, payment timing, collections, cash application, reconciliation, tax evidence, and cash-flow intervention as a per-transaction service—work too fragmented for traditional small-business software.
Security boundary: PCI DSS keeps cardholder-data scope, segmentation, credentials, logging, and access control central. An agent should receive a tokenized, purpose-limited payment capability—not unrestricted access to card data or treasury accounts.
Logistics and supply chain
A supply chain is a network of promises under continuous change. An order, container, truck, labor shift, customs filing, supplier commitment, weather pattern, and customer SLA may each live in a different system or company. Control towers made disruption visible; agentic systems can evaluate recovery choices and coordinate the actions that follow.
Planner spreadsheets, EDI and email, fragmented TMS/WMS/ERP data, manual carrier calls, static plans, and exception queues.
Event-driven case agents maintain order-to-delivery context, simulate alternatives, negotiate constraints, request authority, and execute recovery across partners.
More shipments per planner, shorter exception latency, lower expedite and inventory cost, better asset use, reliable commitments, and new service tiers.
Where agents can act
- Convert demand shifts into governed sourcing, inventory, capacity, and replenishment proposals.
- Tender loads, manage appointments, optimize routes, resolve carrier rejections, and coordinate warehouse and yard operations.
- Prepare customs evidence, product classification, restricted-party checks, filings, and broker review.
- Investigate freight invoices, damage, shortages, accessorial charges, claims, maintenance, and vendor recovery.
New businesses enabled
Guaranteed-outcome logistics
Price and sell verified delivery outcomes—not merely transportation capacity—because agents can continuously manage routes, inventory substitutes, customs, appointments, and exception recovery within an insured risk envelope.
Shared operations network for small shippers
Pool planning intelligence, procurement, compliance, and exception handling across thousands of small companies, giving them a virtual control tower and negotiating capacity previously limited to global enterprises.
Operating context: U.S. Customs and Border Protection’s Automated Commercial Environment demonstrates the structured filings, documents, manifests, partner-agency data, and review steps an agent must coordinate. Blackstone identifies logistics and the physical economy as a major downstream AI value pool.
Manufacturing and industrial operations
Manufacturing has spent years connecting design, PLM, ERP, MES, quality, maintenance, warehouse, supplier, and field-service systems, yet the digital thread remains incomplete. Engineers, supervisors, planners, technicians, and operators still reconcile drawings, specifications, work instructions, alarms, inspection results, deviations, parts, supplier messages, and customer commitments by hand. Agentic AI can become the context and coordination layer across the product and asset lifecycle without bypassing deterministic process control or safety systems.
One-way data flows, tribal knowledge, manual shift handovers, reactive maintenance, isolated quality systems, change propagation delays, and service disconnected from design.
Product, production, quality, maintenance, supply, safety, and service agents operate on a traceable digital thread and coordinate bounded work around verified plant state.
Higher throughput and uptime, faster engineering change, lower scrap and rework, more technicians per expert, resilient scheduling, and outcome-based aftermarket revenue.
Where agents can act
- Translate demand and engineering changes into capacity, material, tooling, labor, qualification, and schedule implications.
- Prepare work instructions and shift handovers from current specifications, deviations, machine state, quality history, and skill requirements.
- Correlate alarms, sensor trends, maintenance history, manuals, parts, warranties, and production impact to plan safe intervention.
- Coordinate nonconformance, root-cause evidence, containment, supplier corrective action, regulatory documentation, field service, and design feedback.
New businesses enabled
Outcome-based industrial equipment
Sell uptime, output quality, energy performance, or cost per unit rather than only a machine and service contract because agents can continuously coordinate telemetry, diagnostics, parts, technicians, production windows, warranties, and verified performance.
Virtual industrial engineering for smaller plants
Combine remote experts with agents that preserve plant context, prepare changes, troubleshoot, manage quality evidence, and coordinate vendors—bringing sophisticated operational support to factories that cannot staff every specialty.
Research direction: NIST’s 2026 AI and machine-learning roadmap for smart manufacturing highlights industrial data, sensing, autonomous systems, digital twins, robotics, supply-chain optimization, sustainability, and the need for trustworthy operation. Deloitte’s 2026 Manufacturing Outlook similarly identifies agentic smart operations and aftermarket service as major opportunities.
Agriculture and agritech
Agriculture combines biological uncertainty, local knowledge, thin margins, seasonal urgency, and physical execution. Data arrives as weather, field history, sensor streams, satellite and drone imagery, equipment telemetry, scouting notes, product labels, storage conditions, and commodity contracts. The opportunity is not generic agronomic chat; it is a field-specific operating loop grounded in evidence and constrained by safety.
Scattered farm records, episodic scouting, late detection, vendor-specific tools, limited agronomist capacity, and disconnected input, machinery, insurance, and market decisions.
Field-level agents fuse perception, agronomy, weather, equipment, labor, compliance, and market context into governed work plans and feedback loops.
More acres per advisor, earlier response, lower input waste, higher equipment utilization, reduced post-harvest loss, and specialist access for smaller farms.
Where agents can act
- Build crop plans from soil, rotation, seed, weather, labor, equipment, budget, and market constraints.
- Detect stress, weeds, disease, pests, water, and nutrient issues from multimodal field evidence.
- Schedule irrigation, scouting, application, maintenance, harvest, drying, storage, and transport.
- Coordinate inputs, commodity marketing, contract obligations, insurance evidence, sustainability records, and multilingual extension support.
New businesses enabled
Agronomy-as-a-service per acre
Continuously monitor and coordinate each field, escalating only uncertain or consequential decisions to agronomists. This can profitably extend high-quality support to farms too small or remote for frequent specialist visits.
Verified outcomes marketplace
Package field evidence, practice execution, yield, water, and carbon outcomes into auditable products for lenders, insurers, buyers, and sustainability programs—while agents manage the evidence trail throughout the season.
Research direction: USDA Agricultural Research Service programs already combine multi-source sensing, meteorological observations, computer vision, autonomous field platforms, and AI—the same perception-to-plan-to-action architecture required for grounded agricultural agents.
Energy and utilities
Electric systems must balance reliability, safety, affordability, security, environmental obligations, and increasingly variable supply and demand. Distributed solar, batteries, electric vehicles, flexible loads, extreme weather, aging assets, and interconnection queues create coordination problems at a scale and speed that human-only operations cannot economically manage—but safety-critical control cannot be delegated to unconstrained language models.
Siloed planning and operations, alarm floods, manual interconnection studies, crew and parts coordination, reactive customer communication, and limited DER visibility.
Agents create context-rich plans across grid, markets, assets, weather, crews, customers, and regulators while deterministic control systems enforce electrical limits.
Faster restoration, lower balancing and field cost, more renewable hosting capacity, better asset use, fewer avoidable truck rolls, and monetized flexibility.
Where agents can act
- Forecast load, renewable output, congestion, storage availability, and demand response; prepare dispatch and bidding recommendations.
- Correlate alarms, outages, weather, vegetation, wildfire exposure, topology, crew position, and customer vulnerability.
- Plan inspections, maintenance, switching preparation, parts, field work, permitting, and interconnection evidence.
- Resolve billing and service cases and assemble reliability, emissions, cybersecurity, and regulatory reports with source lineage.
New businesses enabled
Grid-flexibility operating networks
Aggregate thousands of heterogeneous batteries, EVs, buildings, and industrial loads into reliable market products by continuously negotiating customer preferences, device availability, forecasts, grid constraints, telemetry, and settlement.
Resilience as a managed outcome
Coordinate backup assets, maintenance, fuel, weather preparation, load priorities, crews, and post-event recovery for campuses, municipalities, and critical facilities under an outcome-based service contract.
Market and safety context: FERC Order No. 2222 creates a coordination opportunity for distributed-resource aggregations. NERC CIP standards reinforce the need to isolate critical systems and govern cyber access.
Real estate and construction
Real estate is a document-heavy financial asset attached to a physical place, local law, financing, design, construction, and years of service. Deals and buildings accumulate title records, zoning, leases, plans, BIM, inspections, permits, requests for information, submittals, schedules, invoices, telemetry, work orders, and tenant communications—usually across firms and incompatible systems.
Deal rooms, spreadsheets, email approvals, isolated contractor systems, manual lease abstraction, reactive maintenance, and weak lifecycle continuity.
A property digital case connects acquisition, diligence, design, build, lease, operations, capital planning, compliance, and disposition evidence.
More assets per team, shorter diligence and permit cycles, lower schedule and change-order leakage, higher occupancy, faster maintenance, and improved net operating performance.
Where agents can act
- Screen opportunities, normalize offering materials, coordinate title, survey, zoning, environmental, tax, insurance, engineering, and lease diligence.
- Track design decisions, RFIs, submittals, permits, schedules, procurement, progress evidence, invoices, and change orders.
- Abstract leases, monitor obligations, reconcile charges, manage notices, tours, applications, renewals, and service.
- Prioritize maintenance and capital plans using telemetry, work orders, warranties, parts, vendor performance, access, tenant impact, and energy data.
New businesses enabled
Institutional operations for small portfolios
Offer lease administration, maintenance coordination, energy optimization, vendor management, accounting evidence, and tenant service to owners who cannot support a full asset-management organization.
Continuous building-performance contracts
Price services around uptime, comfort, energy, response time, and compliance because agents can coordinate telemetry, vendors, warranties, access, parts, work orders, and verification continuously.
Rights boundary: the Fair Housing Act applies across renting, buying, mortgages, and related services. Advertising, lead routing, screening, and service agents require testing, accessible experiences, accurate notices, data correction, and meaningful human review.
E-commerce and retail
Retail must reconcile customer intent with products, inventory, price, promotion, fulfillment, returns, loyalty, supplier constraints, and trust. Traditional personalization optimizes a click or ranking. Agentic commerce can complete a customer objective and coordinate the operational consequences, provided sponsorship, substitutions, total price, recurring terms, and data use remain transparent.
Channel silos, search that does not understand constraints, forecast handoffs, static promotions, stockouts, manual supplier content, and costly returns.
Customer, merchandising, inventory, supplier, fulfillment, service, and finance agents coordinate around verified intent and contribution margin.
Higher conversion and availability, lower working capital and return cost, long-tail personalization, faster assortment decisions, and always-on service and selling.
Where agents can act
- Guide discovery through needs, fit, compatibility, availability, total cost, delivery, and approved purchase authority.
- Plan assortment, price, promotion, allocation, replenishment, transfer, markdown, and supplier negotiation within guardrails.
- Onboard products and suppliers by validating claims, documentation, attributes, images, safety, and compliance.
- Resolve returns, exchanges, warranties, fraud, restocking, reverse logistics, store tasks, labor, and local exceptions.
New businesses enabled
Delegated household commerce
A permissioned agent manages recurring needs, price ceilings, substitutions, dietary or compatibility rules, timing, loyalty, and returns—creating a trusted replenishment relationship rather than a sequence of search sessions.
On-demand micro-assortments
Agents continuously detect small demand clusters, source compliant products, create localized content, allocate limited inventory, test economics, and retire weak assortments—making niche markets profitable.
Trust boundary: the FTC’s dark-pattern guidance is directly relevant to agentic commerce. Sponsorship, ranking, substitutions, total price, recurring obligations, consent, and cancellation must be visible and controllable.
Technology services and IT
Technology teams are both early builders and early subjects of agentic transformation because their work is already digital, instrumented, testable, and tool-rich. The opportunity is not unrestricted autonomous coding. It is an evidence-based delivery and operations system that can turn intent into tested, observable, reversible change while escalating uncertainty and preserving engineering accountability.
Requirements handoffs, ticket queues, tool fragmentation, manual runbooks, alert overload, repeated environment setup, and knowledge trapped in people and chat.
Product, engineering, testing, security, release, reliability, service desk, and FinOps agents share a governed work graph and verified system state.
More output per team, shorter lead and recovery times, lower operations toil, broader custom-software economics, and continuous security and cost hygiene.
Where agents can act
- Convert product goals into requirements, acceptance tests, dependency maps, bounded code changes, documentation, and review packages.
- Coordinate build, test, security scan, release, migration, feature flag, observability, and rollback.
- Correlate logs, traces, topology, deployments, tickets, security signals, and known problems during incidents.
- Resolve service requests, access, assets, licenses, knowledge updates, cloud commitments, and unit-cost anomalies.
New businesses enabled
Custom software for the long tail
Build and maintain small, company-specific applications and integrations whose lifetime economics previously did not justify a software team—expanding the addressable market from standard SaaS to outcome-specific software.
Reliability as a verified outcome
Sell managed reliability, security, and cost objectives with agent teams continuously investigating, preparing changes, verifying recovery, and learning from incidents under explicit blast-radius limits.
Market and engineering context: Goldman Sachs projects the application-software market could reach $780 billion by 2030, with agents more than 60% of the total. NIST’s Secure Software Development Framework provides the control baseline for provenance, isolation, testing, review, and secure release.
Call centers and customer service
Customer service is the clearest early agentic market because cases are high volume, SOPs are relatively explicit, channels are digital, and outcomes—resolved issue, completed transaction, customer satisfaction, repeat contact—can be measured quickly. The strategic leap is from conversational containment to verified resolution across account, order, billing, field service, identity, product, and policy systems.
Channel-specific bots, repeated authentication, knowledge search, narrow scripts, transfers without context, manual after-call work, and quality sampling.
A unified service case authenticates, reasons, acts, verifies, communicates, and transfers full context to a person when vulnerability, risk, or novelty requires it.
24/7 multilingual capacity, higher first-contact resolution, lower customer effort, proactive prevention, broader sales coverage, and quality monitoring across every interaction.
Where agents can act
- Identify intent, language, urgency, vulnerability, sentiment, identity risk, and the exact outcome requested.
- Resolve appointments, orders, billing, returns, technical support, service changes, payment arrangements, and warranty requests.
- Proactively contact customers about disruptions, incomplete onboarding, renewals, usage, risk, and next-best preventive action.
- Maintain knowledge from verified resolutions, analyze all interactions for quality, and surface training and policy gaps.
New businesses enabled
Resolution as a service
Price customer operations per verified outcome—resolved claim, retained subscriber, scheduled appointment, restored service—rather than per seat or minute, aligning the provider with customer and enterprise value.
Always-on revenue concierge
Give small and global businesses continuous, multilingual product guidance, qualification, onboarding, renewal, and service capacity that learns the customer context across the lifecycle.
Adoption and conduct context: a16z finds support attractive because high-volume work has explicit SOPs and measurable results; its enterprise adoption analysis also cautions that not every industry has such verifiable outputs. The FCC’s AI-voice ruling makes consent, identification, purpose, time, and opt-out policy enforceable requirements.
Legal services
Legal work combines dense language, evidence, deadlines, negotiation, jurisdiction, and professional judgment. Traditional legal software stored matters and templates but did little to accelerate the unstructured reasoning lawyers perform. Agentic systems can make the entire matter more legible and coordinated while keeping the lawyer responsible for advice, filings, confidentiality, strategy, and candor.
Email intake, manual issue spotting, repeated research, document version friction, deadline risk, fragmented obligations, and services priced around human hours.
A matter-level agent team maintains sources, issues, evidence, playbooks, approvals, drafts, negotiations, deadlines, obligations, and client communication.
More matters per professional, faster contract and diligence cycles, broader access, better institutional knowledge, continuous obligation monitoring, and outcome-oriented delivery.
Where agents can act
- Conduct intake, conflicts preparation, engagement, evidence collection, matter opening, and client-status coordination.
- Search authorities, build issue maps, verify citations, prepare drafts, compare arguments, and preserve research trails.
- Review and negotiate contracts across legal, privacy, security, IP, commercial, regulatory, and operational playbooks.
- Coordinate discovery, diligence, dockets, filings, holds, deadlines, regulatory mapping, obligations, budgets, and knowledge reuse.
New businesses enabled
Continuous contract compliance
Move beyond a contract repository to a service that monitors obligations, renewals, notice windows, policy changes, operational evidence, counterparties, and remediation across every agreement.
Legal operations for the underserved middle
Combine lawyer oversight with agents that prepare recurring corporate, employment, privacy, commercial, and regulatory work—making continuous legal hygiene affordable to small and midsize businesses.
Professional boundary: ABA Formal Opinion 512 applies duties of competence, confidentiality, communication, candor, supervision, and reasonable fees. a16z’s adoption research identifies legal as an early vertical because AI reaches text-heavy, nuanced work that static workflow software did not.
Education
Education’s abundance opportunity is not unlimited generated content; it is personalized explanation, practice, feedback, advising, and support coordinated around actual mastery and learner context. Institutions already have LMS, student, assessment, financial-aid, attendance, accessibility, advising, and communication systems—but people must interpret their combined signal and intervene.
One-to-many instruction, delayed feedback, disconnected support services, administrative navigation, episodic advising, and educators overwhelmed by preparation and documentation.
A transparent learner case coordinates curriculum, practice, assessment evidence, advising, finance, accessibility, family communication, and human intervention.
More feedback per learner, teacher capacity, earlier support, better completion, individualized pathways, multilingual access, and affordable lifelong learning.
Where agents can act
- Adapt explanation, examples, guided practice, pacing, and formative feedback to demonstrated understanding.
- Help educators plan, differentiate, build rubrics, identify misconceptions, and review learner evidence.
- Coordinate degree planning, enrollment, financial aid, tutoring, accessibility, attendance, and student-success outreach.
- Support research discovery, reproducibility, scheduling, procurement, facilities, compliance, and multilingual family communication.
New businesses enabled
Mastery subscription for every learner
Provide continuous, curriculum-aligned tutoring and feedback with educator escalation and a portable, correctable mastery record—at a marginal cost that makes one-to-one support far more accessible.
Skills-to-work pathway operator
Continuously map a learner’s evidence to changing job requirements, generate projects and practice, verify skills, coordinate mentors and employers, and adjust the pathway based on outcomes rather than course completion alone.
Student boundary: the U.S. Department of Education’s safe, ethical, and equitable AI toolkit spans privacy, security, civil rights, equity, accessibility, misinformation, and educational purpose. A tutoring model becomes a system only when it is embedded responsibly in that broader environment.
Government and public services
Government processes combine complex rules, old systems, paper evidence, huge case volumes, accessibility, public records, procurement constraints, and due process. Residents do not experience government as departments; they experience life events—losing a job, starting a business, recovering from disaster, moving, aging, or caring for a family. Agents can coordinate that journey, but public authority must remain grounded in law and accountable decision rights.
Program silos, repeated forms, long backlogs, inaccessible language, manual evidence checks, call-center handoffs, and citizens acting as integration middleware.
A consented life-event or case layer coordinates programs, evidence, rules engines, officials, inspections, notices, payments, appeals, and records.
Shorter backlogs, higher access and completion, fewer errors and improper payments, multilingual service, better emergency coordination, and more capacity per caseworker.
Where agents can act
- Guide benefits, permits, licenses, taxes, grants, procurement, and public-information requests.
- Collect and validate evidence, calculate rules deterministically, assemble cases, prepare notices, and route authorized decisions.
- Coordinate inspections, remediation, enforcement, appeals, public records, redaction, retention, and program integrity.
- Support emergency operations with weather, infrastructure, shelter, supply, responder, vulnerability, and communication context.
New businesses and public capabilities enabled
Life-event public concierge
With explicit consent, coordinate every relevant program and step around a resident’s objective, reducing the navigation burden without creating an opaque universal citizen profile.
Regulation as a machine-actionable service
Publish official rules, definitions, evidence requirements, forms, changes, and appeal paths in agent-readable form so businesses and residents can comply by design and agencies can process cases consistently.
Public accountability: U.S. federal AI direction is available through OMB memoranda. Agencies must also satisfy program law, records, procurement, accessibility, security, privacy, civil-rights, and administrative-law obligations. Every consequential case needs durable inputs, rule versions, actions, approvals, notice, and correction history.
HR and talent
HR agents influence access to work, pay, benefits, development, mobility, and support. The value opportunity is large because workforce plans, recruiting, interviews, onboarding, payroll, access, learning, internal projects, leave, and employee service are fragmented across systems and teams. The risk is equally large: opaque inference and historical data can scale discrimination or surveillance.
Keyword matching, scheduling friction, unstructured interviews, slow onboarding, ticket-based employee service, static job architecture, and incomplete skills visibility.
A permissioned talent and employee case coordinates job evidence, accommodations, structured decisions, onboarding dependencies, skills, opportunities, and service resolution.
Faster hiring and productivity, better candidate and employee experience, more internal mobility, scalable support, reduced vacancy, and more precise workforce planning.
Where agents can act
- Translate workforce plans into job-related skills and requirements; draft roles; source and schedule without inferring protected attributes.
- Organize applicant evidence against explicit criteria, support accommodations, and coordinate structured interviews and decision records.
- Manage offers, background steps, equipment, access, payroll, benefits, orientation, mentors, and early support.
- Resolve leave, policy, payroll, benefits, mobility, and manager requests; connect employees to projects, learning, mentors, and internal roles.
New businesses enabled
Internal opportunity market
Continuously match verified employee skills and aspirations with projects, mentors, learning, short-term assignments, and roles—then coordinate approvals, capacity, access, and development evidence.
Full-service people operations for small employers
Combine professional oversight with agents that coordinate recruiting, onboarding, policy, payroll evidence, benefits, learning, and employee requests, making mature people operations affordable below enterprise scale.
Employment boundary: the EEOC’s selection-procedure guidance explains why a tool can create unlawful discrimination through intentional treatment or unjustified disproportionate exclusion. Avoid emotion, personality, or trustworthiness inference from faces, voices, or private communications.
Insurance
Insurance turns uncertain events into contracts, pricing, reserves, prevention, service, and recovery. Submissions, policy forms, endorsements, inspections, images, weather, medical records, estimates, telematics, IoT, vendor networks, and state requirements create a multimodal coordination problem across the policy lifecycle. Agents can reduce friction and make prevention a continuous service, but actuarial integrity, fair treatment, licensed authority, and explainability remain central.
Manual submission ingestion, periodic underwriting, document-heavy claims, adjuster and vendor coordination, catastrophe queues, leakage, and reactive risk management.
A policy or claim case coordinates exposure evidence, coverage, licensed decisions, inspections, vendors, fraud review, settlement, recovery, prevention, and communication.
Faster quote and claim cycles, more small-account coverage, lower expense and leakage, better catastrophe response, earlier mitigation, and products tailored to changing risk.
Where agents can act
- Ingest submissions, normalize exposure data, retrieve external evidence, prepare underwriting files, and manage quote-to-bind subjectivities.
- Receive first notice of loss, verify coverage, triage severity, coordinate immediate help, collect evidence, prepare estimates, and route authorized settlement.
- Connect fraud, duplicate claims, subrogation, salvage, vendor, catastrophe, and recovery workstreams without making unsupported accusations.
- Monitor telematics, sensors, property conditions, weather, maintenance, and behavior to coordinate risk reduction and dynamic service.
New businesses enabled
Prevention as part of the policy
Continuously detect risk, recommend and coordinate mitigation, verify completion, and adjust service—shifting the insurer from payer after loss to a partner that reduces loss frequency.
Micro-duration and embedded coverage
Agents assemble context, explain terms, bind under deterministic rules, monitor exposure, trigger evidence, and coordinate claims for small or temporary risks whose manual servicing cost made them uneconomical.
Regulatory and operating evidence: the NAIC Model Bulletin emphasizes governance, accuracy, fairness, validation, third-party oversight, drift, and examination evidence. BCG reports examples of agentic claim handling reducing cycle time and improving NPS, while stressing that outcomes vary and controls must be designed from the beginning.
From disaggregated data to an agent-ready operating fabric
Agentic AI does not require every source to be copied into one giant database. It requires a governed way to resolve identity, discover relevant information, understand meaning, retrieve only permitted context, recognize freshness, and write verified outcomes back to authoritative systems. A dashboard can tolerate a stale field; an action-taking system cannot.
| Layer | Examples | What agents need | Failure if missing |
|---|---|---|---|
| Identity and master data | Customer, patient, employee, supplier, product, asset, organization. | Authoritative resolution, aliases, relationships, consent, purpose, permissions. | Correct action on the wrong entity. |
| Transactions and cases | Orders, claims, policies, loans, visits, filings, contracts, tickets. | Current state, owners, dependencies, history, completion condition. | Duplicate or out-of-sequence action. |
| Unstructured knowledge | Notes, email, manuals, regulations, transcripts, plans, images. | Parsing, permissions, provenance, version, jurisdiction, semantic retrieval. | Unsupported conclusion or policy mismatch. |
| Operational events | Telemetry, logs, GPS, meters, clicks, calls, deadlines, status changes. | Freshness, ordering, deduplication, confidence, event-to-case correlation. | Late or irrelevant intervention. |
| Decision evidence | Sources, calculations, prompts, policies, approvals, tool results. | Immutable trace with privacy-aware retention and reproducibility. | No explanation, audit, correction, or learning loop. |
Context package
Send the minimum case-specific context, with source, owner, time, sensitivity, jurisdiction, and permission—not an unfiltered enterprise dump.
Tool contract
Every tool declares allowed inputs, outputs, authority, idempotency, limits, expected state change, errors, owner, and rollback path.
Decision record
Store what evidence was used, which policy and model versions applied, who approved, what action occurred, and whether the resulting state matched intent.
Implement around a workflow, not around a model
1. Define and observe
Name the outcome owner. Map normal and exception paths. Establish baseline cost, volume, delay, quality, risk, and demand. Collect representative cases, adversarial examples, and known failures.
2. Prepare and supervise
Build narrow tools and permission boundaries. Run in shadow mode, then allow the agent to stage actions for human approval. Review evidence, uncertainty, policy, reversibility, and the actual resulting state.
3. Bound and scale
Automate only case types that meet outcome and risk thresholds. Keep transaction, iteration, cost, rate, and time limits; monitor drift and overrides; expand scope based on evidence.
The end-to-end case pattern
- Trigger: detect a request, event, deadline, anomaly, or unmet objective.
- Identity and intake: establish who or what is involved; structure the case and permissions.
- Plan: decompose work, dependencies, tools, evidence, approval gates, budget, and completion criteria.
- Specialist work: retrieve data; call predictive models, rules, optimizers, and domain agents; surface contradictions.
- Verify: check sources, calculations, policy, completeness, expected state, and downstream consequences.
- Authorize and execute: obtain the required human or deterministic approval and call the narrowest tool.
- Confirm and communicate: verify the authoritative record, explain status, and preserve appeal or correction paths.
- Learn: connect production traces to outcome metrics, incident review, evaluation data, and controlled change.
The 10/20/70 reality: BCG advises treating algorithms as roughly 10% of an AI transformation, the technology backbone as 20%, and people and process change as 70%. The exact split will vary, but the operating lesson is durable: the most difficult work is redesigning roles, decision rights, data, incentives, and exception handling.
Measure capacity, outcomes, risk, and new revenue—not tokens
Released time is not automatically value. It becomes valuable when it produces more completed service, shorter queues, better quality, lower overtime, growth without equivalent hiring, or an intentional workforce change. Faster cycle time matters when it improves conversion, experience, working capital, reliability, health, learning, or risk. New-business value should be measured separately from efficiency so a transformation does not get reduced to headcount arithmetic.
Outcome
Revenue, completion, quality, loss, reliability, safety, health, learning, or resident result.
Process
Cycle, backlog, throughput, straight-through rate, exception, rework, handoff, and human minutes.
Stakeholder
Customer effort, trust, adoption, override, employee workload, access, and distributional impact.
Risk and system
Unsupported action, policy breach, security incident, fairness, drift, latency, cost, availability, and reversal.
Unit economics that prevent “successful” pilots from failing at scale
- Measure cost per verified completed case, including inference, retrieval, tools, human review, retries, and incident burden.
- Segment by case type. Easy-case averages can hide expensive or unsafe long-tail performance.
- Compare with the true baseline: labor, queue delay, abandonment, leakage, rework, quality failures, and unmet demand.
- Track the autonomy yield: the share of cases completed inside the approved license without unnecessary escalation or silent error.
- Price new services around value or outcome only when measurement, attribution, and reversal rights are credible.
Agentic risk is action risk
Hallucination is only one failure mode. An agent can retrieve malicious instructions, expose protected data, use the wrong identity, call the wrong tool, exceed authority, perform the right action on the wrong account, repeat a transaction, optimize the wrong objective, coordinate with a compromised agent, or announce completion before the system of record changed.
Before action
Inventory and risk-classify every agent. Define intended use, owner, data, tools, prohibited actions, approval gates, testing, identity, transaction limits, and rollback.
During action
Use least privilege, short-lived credentials, untrusted-input isolation, deterministic policy, rate and cost limits, source grounding, independent verification, and human escalation.
After action
Verify authoritative state, retain the trace appropriately, monitor outcomes and drift, enable correction and appeal, investigate incidents, and control model, prompt, data, and tool changes.
The NIST AI Risk Management Framework organizes risk work around govern, map, measure, and manage. OWASP’s agentic guidance adds risks created by tools, memory, identities, and autonomy. For organizations in scope, the EU AI Act applies a risk-based regime with phased and use-specific obligations. Classification must follow the actual use and affected people—not the marketing label on the model.
Work shifts from processing cases to designing and governing outcomes
Agentic AI changes work at the task, workflow, team, and business-model levels. It can remove repetitive effort, create more service capacity, and increase availability. It can also reduce demand for some roles, compress entry-level work, intensify oversight, and concentrate mistakes. Organizations need a workforce strategy before autonomy changes the operating model by accident.
Domain authority
Professionals define acceptable evidence, policy, exceptions, quality, and the cases where human relationships or judgment are the product.
Agent product and operations
Named owners manage outcome, queue, capacity, cost, incidents, drift, permissions, tool health, evaluation, and controlled improvement.
Exception and assurance expertise
People handle novel, emotional, adversarial, rights-affecting, and high-consequence cases while independent teams test reliability, safety, security, fairness, and compliance.
Apprenticeship must be redesigned. If agents produce every first draft, junior professionals may never build the judgment required to supervise them. Training should include critique of agent work, simulated cases, source verification, exception handling, stakeholder communication, and explicit practice in independent reasoning.
Workforce signal: BCG reports that among heavy adopters, 43% expect greater demand for generalists who can manage human-agent teams, while 29% expect fewer traditional entry-level roles. These are expectations, not destiny; they make deliberate job design, transition support, and skill measurement urgent.
The one-year, three-year, and five-year roadmap
One year · Bounded execution
Task and domain agents become standard. Document operations, customer service, coding, compliance preparation, and exception triage lead. Most consequential actions remain human-approved. Evaluation, identity, permissions, and case state become shared platform capabilities.
Three years · Orchestrated functions
Agents coordinate across applications and departments. They receive service levels, limited budgets, transaction policies, and identities. Interfaces evolve from chat to action centers showing plan, evidence, state, exceptions, approvals, cost, and outcomes.
Five years · Agent-readable industries
APIs, contracts, catalogs, credentials, policies, consent, and dispute mechanisms become machine-actionable. Partner agents negotiate schedules, inventory, services, evidence, and payments. Physical systems connect through digital twins and guarded control interfaces.
The likely endpoint is not a fully autonomous company. It is a company where routine coordination happens at machine speed, evidence is continuously assembled, exceptions reach the correct authority, and scarce human attention shifts toward judgment, relationships, strategy, creativity, and novel problems.
A decision framework for founders and enterprise leaders
For founders
- Start with a costly completed outcome owned by a specific buyer.
- Choose work where domain workflow, integrations, permission, and evaluation create durable advantage.
- Use real cases and exceptions to build a proprietary evaluation and improvement loop.
- Design compliance, security, auditability, and human authority into the product.
- Sell measurable work or a new capability—not a generic assistant.
- Price the outcome only after unit economics and attribution are reliable.
For enterprise leaders
- Fund a portfolio of workflow transformations rather than one giant “AI program.”
- Invest in data, tools, identity, evaluation, security, process, and adoption—not only models.
- Assign business ownership and decision rights before production write access.
- Use build, buy, and partner decisions at the component level.
- Measure customer, workforce, risk, and growth outcomes beside efficiency.
- Expand autonomy only when production evidence supports a larger license.
Ask these seven questions before funding an agent
- What exact work will be complete when the agent succeeds?
- Why do coordination, ambiguity, and exceptions make simpler automation insufficient?
- Which authoritative evidence, deterministic rules, systems, and people are required?
- How will an independent verifier know that each step and the final state are correct?
- What is the worst plausible failure, and which boundary prevents or contains it?
- Do cost per verified outcome and latency remain viable at full production volume?
- Which new service, market, transaction, or customer segment becomes possible—not merely cheaper?
Build the system of action, not the wrapper
The first generative-AI wave changed how people create and retrieve information. The agentic wave changes how organizations execute work. Its defensible benefit is not that a model replaces every prior technology; it is that a governed reasoning layer can compose data, rules, predictions, optimizers, tools, people, and counterparties around a completed outcome.
Across all 16 industries, the same advantage recurs: unstructured evidence becomes operational, fragmented systems become one case, exceptions become manageable, human authority becomes explicit, and capacity scales beyond the economics of manual coordination.
The winners will not have the most agents. They will redesign the most valuable workflows, build the deepest evidence and evaluation loops, earn the right level of autonomy, and use lower coordination cost to create services that did not make economic sense before.
Agentic AI’s largest prize is not labor substitution. It is making previously scarce coordination, expertise, and service abundant—without making accountability disappear.
Selected research, investor, regulatory, and industry sources
Research and market estimates are directional, reflect their authors’ methodologies, and are not guarantees of results. Regulatory obligations vary by use, jurisdiction, and date.
Enterprise, investment, and economic research
- McKinsey: The State of AI in 2025
- McKinsey: Seizing the Agentic AI Advantage
- BCG: How Agentic AI Is Transforming Enterprise Platforms
- BCG: Agents and the Next Wave of AI Value Creation
- Bain: AI Insights and the Cross-System Labor Opportunity
- Deloitte: Agentic AI Is Scaling Faster Than Guardrails
- Goldman Sachs Research: Agents, Productivity, and Software Market Growth
- Morgan Stanley Research: AI in the Workplace
- Andreessen Horowitz: Computer Use and Agentic Coworkers
- Andreessen Horowitz: Where Enterprises Are Actually Adopting AI
- Andreessen Horowitz: Anatomy of an Enterprise Platform Company
- Bessemer Venture Partners: The State of AI 2025
- General Catalyst: AI and Healthcare Transformation
- Khosla Ventures: A Roadmap to AI Utopia
- Blackstone: AI at Scale
- KKR: 2026 Global Macro Trends and Productivity
- Insight Partners: The State of the AI Agents Ecosystem
Governance and industry authorities
- NIST AI Risk Management Framework
- OWASP Agentic AI Threats and Mitigations
- European Union Artificial Intelligence Act
- CMS Interoperability and Prior Authorization Final Rule
- Federal Reserve SR 26-2: Revised Model Risk Guidance
- PCI Security Standards Council Document Library
- U.S. Customs and Border Protection: Automated Commercial Environment
- NIST: 2026 AI and ML Roadmap for Smart Manufacturing
- Deloitte: 2026 Manufacturing Industry Outlook
- USDA ARS: Precision Agriculture and Autonomous Systems Research
- FERC Order No. 2222 Explainer
- NERC Critical Infrastructure Protection Standards
- HUD Fair Housing Act Overview
- FTC: Bringing Dark Patterns to Light
- NIST Secure Software Development Framework
- FCC Declaratory Ruling on AI-Generated Voices and the TCPA
- American Bar Association Formal Opinion 512 Overview
- U.S. Department of Education: Safe, Ethical, and Equitable AI Toolkit
- OMB Memoranda on Federal AI Use and Acquisition
- EEOC: Employment Tests and Selection Procedures
- NAIC: Artificial Intelligence and the Model Bulletin for Insurers
