Right-Sizing AI Operations: A Decision Framework for 2026

How $1M to $5B businesses should approach the buy decision when 95% of pilots fail and 5% return 30×.

§0 — TL;DR

§1 — The 2026 Reality

Two numbers frame everything that follows. They are not in tension; they describe the same market from opposite ends.

95% of enterprise GenAI pilots deliver zero measurable P&L impact. MIT Project NANDA's GenAI Divide report (July 2025), based on 52 executive interviews, 153 leader surveys, and analysis of 300 public deployments, found that the vast majority of enterprise AI initiatives produced no measurable bottom-line outcome. The NANDA methodology has drawn pushback since publication — on sample composition, the definition of "measurable P&L impact," and survivorship in public reporting — but the directional finding is independently corroborated by McKinsey's 2026 State of AI: of the 39% of organisations attributing any EBIT impact to AI, most report under 5% of EBIT attributable. The exact failure share is debated; the order of magnitude is not. (MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025," July 2025; McKinsey "State of AI" 2026 release.)

The surviving 5% are running away with the category. Klarna's CS agent deployment alone is reported at $60M saved and the workload equivalent of 853 employees by Q3 2025. JPMorgan runs 450+ AI use cases in production daily — including agentic generation of investment-banking presentations in 30 seconds (vs. hours for a junior analyst) and a wealth-management AI assist credited with a 20% gross-sales increase during market volatility. Gartner's market estimate moved from $7.6B in 2025 to $10.8B in 2026, with a best-case 2035 projection of $450B. (Multiple sources — see Appendix B.)

The decision question is not whether the category is real. It is whether your organisation is structurally positioned to land in the 5%.

Table 1.1 — The Two-Group Market, Stated Honestly

Source: MIT Project NANDA (July 2025); McKinsey State of AI 2026.

Table 1.2 — Where the 5% Concentrate

Source: Composite case study analysis from Klarna (FintechWeekly, 2025), JPMorgan public disclosures (Emerj, 2026), AI Monk enterprise case studies (2025–26).

§2 — Why the Buy Decision Is Different in 2026

Three structural shifts changed the buying landscape between 2024 and 2026. Each one moved a piece of the decision out of "should we?" and into "which vehicle?"

The 2026 buyer effectively chooses among three vehicles: a legacy per-seat SaaS product retrofitted with AI features; a specialised agent platform with published pricing and a governance substrate; or a custom build on direct foundation-model APIs (Anthropic, OpenAI, etc.). The third option is currently the most common pattern at enterprise scale — and is also where the 95% failure rate documented in §1 most heavily concentrates, primarily because the workflow redesign, the audit substrate, and the governance posture all have to be assembled in parallel with the AI itself. The three shifts below describe why the first option is fading, why the second is emerging, and why the third is structurally harder to land than its lower published cost suggests.

Shift one: per-seat pricing has structurally collapsed for agent-density categories. When a single AI agent does the work that previously required 10, 20, or 50 human seats, per-seat pricing inverts — the customer with more agents pays less in relative terms. The shift applies specifically to categories where the unit of work is an event, not a user (agent platforms, AI-assisted CS, document processing); per-seat remains the dominant model for CRM, HRIS, ERP and the rest of the seat-anchored stack. Within the agent-platform category, Gartner predicts at least 40% of enterprise SaaS spend will shift to usage-, agent-, or outcome-based pricing by 2030, with seat-based revenue share declining from 21% to 15%. Hybrid pricing (fixed base + variable consumption) is the dominant transition state, used by 43% of SaaS firms in 2025–26, projected to hit 61% by end of 2026. Firms using hybrid pricing report 38% higher revenue growth and 38% higher net revenue retention than pure-subscription firms. (Bessemer Venture Partners AI Pricing Playbook; Valueships AI Pricing 2026 analysis.)

Shift two: predictable scaling has become the procurement gate. Gartner reports 90% of CIOs cite cost forecasting as their top AI deployment challenge, and 78% of IT leaders report unexpected charges from consumption-based or AI pricing models. Cancellation rate among agentic AI projects is now forecast at 40%+ by end of 2027, primarily for scope-and-cost reasons rather than capability reasons. A published rate that scales linearly is the only model that survives a procurement-led review at mid-market and enterprise scale.

Shift three: for EU-exposed operators, the EU AI Act enforcement deadline (August 2, 2026) is real. High-risk AI systems (employment, credit, education, law enforcement) now require Article 11 technical documentation, Article 12 event logging, and Article 9 risk management. Non-compliance penalties reach €35M or 7% of global turnover. "EU-exposed" includes any operator selling into the EU market, processing EU customer data, or running AI-mediated decisions that touch EU subjects — not just EU-headquartered firms. For organisations outside that perimeter, the deadline is a leading indicator rather than a hard cliff: the US, UK, and APAC regulatory direction is following the same per-decision audit substrate, on a slower timeline. The audit substrate is no longer optional infrastructure — it is a regulatory requirement that arrives whether the buyer is ready or not.

Table 2.1 — How the Three Shifts Change the Buy Vehicle

Source: Composite. Bessemer 2026 AI Pricing Playbook; Gartner Strategic Predictions 2026; EU AI Act Regulation 2024/1689.

§3 — The Three Inputs That Determine Everything

The recommended deployment shape is determined by three inputs the buyer already knows: industry, annual revenue, and ambition. No CRM data, no operational telemetry, no internal benchmarking required.

Table 3.1 — The Three-Input Diagnostic Model

Table 3.2 — Invocation Density, Industry-Calibrated

Invocation density reflects multi-agent chains per business event. A typical insurance claim routes through 8–10 agents (intake, classification, document analysis, fraud, liability, communication, audit); a support ticket 4–6; a recruiting funnel 5–8. Density translates to a per-$M-revenue, per-month invocation rate. The figures below are calibration choices grounded in agent-chain modelling against representative workflows in each category — they are not measured industry averages, since no comparable public dataset exists. They are the values the diagnostic uses; a buyer with internal telemetry should override them with their own measured rate.

For businesses that don't sit cleanly in one category: hybrid models default to the dominant cost-pattern. Fintech and insurtech behave structurally like SaaS (event-rich, software-led decisioning); healthtech and edtech behave like Services (people-and-decision-intensive, regulated visibility); marketplaces and DTC commerce split along the operations-vs-tech-stack divide. Pick the category your dominant cost line resembles, and the diagnostic remains directionally correct. Above ~$100M revenue, the right move is to re-anchor against your own internal telemetry.

Table 3.3 — Ambition Multiplier

Ambition has two effects in the model, not one. It drives invocation volume (above), and it drives where in each function's lo–hi rate band the per-function impact is computed (Conservative = floor of the band, Base = midpoint, Aggressive = ceiling). The two effects compound: an Aggressive setting picks up both more invocations and richer per-function impact.

The recommended invocation volume falls out of:

recommended_invocations_per_month
  = annual_revenue ($M) × invocation_density × ambition_volume_multiplier

ambition_volume_multiplier  — Conservative 0.6, Base 1.0, Aggressive 1.4
ambition_rate_position      — Conservative 0.0, Base 0.5, Aggressive 1.0
                              (position within each function's published lo–hi band)

Table 3.4 — Worked Examples

§4 — Volume Drives the Vehicle

The 150,000 invocations/month ceiling is not a sales preference. It is the operational point at which a flat published rate stops being the right contractual instrument.

Below the ceiling, the buyer's bill is predictable to the dollar (90% of CIOs' #1 concern, addressed structurally). Above the ceiling, the buyer should be negotiating volume discount, dedicated ops engineering, and custom SLA — not paying a published per-invocation rate that compounds linearly without discount.

Table 4.1 — Self-Serve vs Custom: The 150K Line

The crucial property of this model is continuity. A self-serve customer who grows past 150K invocations/month is not re-platforming, re-implementing, or re-training — only the contract changes. Same agents, same audit trail, same data, same governance policies.

Table 4.2 — Market Pricing Context (2026)

For comparison, here is where the $99 + $0.06/invocation model sits relative to published peer pricing in the agentic AI category.

Source: Awesome Agents Pricing Comparison 2026; LangChain published pricing; Microsoft Copilot Studio published pricing (2026); IBM watsonx published tiers; Lyzr & OneReach published positioning; Anthropic / OpenAI published API pricing; Agno public statements (2026).

§5 — Year-1 ROI Is a Band, Not a Number

A single ROI multiple — "this investment returns 47.3× in Year 1" — is the fastest way to lose a CFO. The number reads as marketing math regardless of how rigorous the underlying model is. Three things go wrong simultaneously: the decimal precision implies false certainty, the magnitude triggers anchoring bias against unrealistic returns, and the buyer instinctively discounts everything else in the pitch.

A band, by contrast — "12×–18× in Year 1, where real enterprise AI deployments actually sit" — preserves three things at once: the rough magnitude (which is what matters), the inherent uncertainty (which the buyer already knows is there), and the comparison anchor (which makes the band defensible).

Table 5.1 — The 5×–60× Envelope, with Public Benchmarks

Source: Public Klarna disclosures (2024–25), JPMorgan AI roadmap (Emerj 2026), McKinsey State of AI 2026, AI Monk SMB case studies. Public benchmark ROIs above are point estimates derived from each company's own disclosures, not bands computed from the formula in Table 5.2 — they serve as the empirical anchor for where the 5×–60× envelope actually sits.

Table 5.2 — Diagnostic-Output ROI Bands by Profile

Each band below is computed from the diagnostic's underlying model and clamped to the 5×–60× envelope:

y1_roi_lo = max(5,  floor(y1_roi_multiple × 0.8))
y1_roi_hi = min(60, ceil (y1_roi_multiple × 1.1))

Each row is computed strictly from the formulas above — y1_roi_lo = max(5, floor(roi × 0.8)), y1_roi_hi = min(60, ceil(roi × 1.1)). A buyer can reproduce these numbers from the three inputs (industry, revenue, scenario) and the methodology in Appendix A. ROI bands shown for self-serve fits only. Above the 150K/mo ceiling, both pricing and ROI shift into the custom-contract regime where the band reflects negotiated rates.

§6 — Decision Matrix by Company Size

The right buying motion is determined more by company size than by industry. Below is the matrix that should drive your decision in 2026.

Table 6.1 — Recommended Motion by Revenue Band

Note: bands above are computed strictly from §5's methodology against representative profiles at the midpoint of each revenue band, Base scenario. The Conservative-vs-Aggressive setting within a band shifts the ROI by ~30% in either direction. Diagnostic outputs that fall outside the band — high or low — are not errors; they are either a genuine outlier or an unrealistic scenario setting and should be re-checked.

§7 — Six Worked Scenarios

Each scenario shows how the diagnostic routes the buy for a representative organisation, calibrated against published 2026 benchmarks.

§7.1 — Lean Services Firm ($3M revenue, Conservative)

Profile: 12-person consulting practice. Founder-led. No formal ops function. Want to free senior consultant time for client work.

Diagnostic output:

• Recommended volume: ~1,800 invocations/month (under the 5,000 included)
• Monthly cost: $99 flat (no overage)
• Annual cost: $1,188
• Total opportunity: ~$90K/year (3% of revenue, weighted toward Conservative)
• Y1 net value: ~$22K
• Y1 ROI band: 15×–21×

Decision: Buy. The cost is irrelevant; the only question is which workflow to pilot first.

Recommended pilot: Inbound proposal triage and first-draft response (BD function). Routes through 4–6 agents per inbound, generates ~150–250 invocations/month per active deal. Even at base volume, the consultant team gets ~4 hours/week back. Benchmarks against reported SMB lead-qualification deployments showing response-time compression from hours to 60 seconds.

§7.2 — Growing Services Firm ($25M revenue, Base)

Profile: 80-person services firm. COO function exists, processes mostly manual. Pre-Series B mindset — want to scale without scaling headcount proportionally.

Diagnostic output:

• Recommended volume: ~25,000 invocations/month
• Monthly cost: $99 + 20,000 × $0.06 = $1,299
• Annual cost: $15,588
• Total opportunity: ~$750K/year
• Y1 net value: ~$172K
• Y1 ROI band: 8×–13×

Decision: Buy. Pilot one function for 60 days, expand if the diagnostic's recommendation holds within ±20% of measured impact.

Recommended pilot: Recruiting / HR top-of-funnel. 60–70% of recruiter hours are sourcing and screening; agents can carry that band end-to-end. Volume builds quickly (~3,000 invocations/month per active hiring funnel). Aligns with the McKinsey workflow-redesign finding — the function is redesigned around the agent rather than retrofitted.

§7.3 — Mid-Market SaaS ($100M revenue, Base)

Profile: 350-person SaaS company. Mature support function, growing CS function. Engineering productivity is the largest cost line.

Diagnostic output:

• Recommended volume: ~120,000 invocations/month (just under self-serve ceiling)
• Monthly cost: $99 + 115,000 × $0.06 = $6,999
• Annual cost: $83,988
• Total opportunity: ~$4.0M/year
• Y1 net value: ~$0.95M
• Y1 ROI band: 9×–13×

Decision: Buy self-serve with a 12-month review planned. If sustained volume exceeds 150K invocations/month, transition to custom contract — same platform, negotiated rate.

Recommended sequence: Support (Tier 0 deflection — the function class where Klarna succeeded before overreaching) → Customer Success (retention signals) → Engineering productivity. Each function added compounds the platform's marginal value. Note: keep human handoff cleanly architected from day one — the Klarna 2025 reversal is the warning.

§7.4 — Established Manufacturer ($500M revenue, Base)

Profile: 1,800-person manufacturer, multi-site. Substantial back-office, complex supply chain. Regulated finance function. Procurement-led buying culture.

Diagnostic output:

• Recommended volume: ~250,000 invocations/month (above self-serve ceiling)
• Vehicle: Custom contract
• Total opportunity: ~$11M/year
• Y1 net value: ~$2.45M (gross opportunity × 25% capture − $0.3M custom-contract services cost)
• Y1 ROI band: contract-dependent (typically 10×–20× at this scale)

Decision: Talk to sales. The published price is informational only; the contract is built around volume discount, dedicated ops engineer, custom SLA, and procurement-aligned terms.

Recommended pilot: Production planning + supply chain coordination first (highest pool size, internal-visibility governance — fastest deployable). AP/AR (finance / back-office) second once governance controls are validated against internal audit requirements. EU AI Act compliance posture is material at this scale — see §11.

§7.5 — Enterprise Services ($1B revenue, Aggressive)

Profile: 4,500-person services firm. Substantial bench, billable utilisation is the controlling metric. Regulated client base. Multi-jurisdiction operations.

Diagnostic output:

• Recommended volume: ~1,400,000 invocations/month (deeply enterprise scale)
• Vehicle: Custom contract
• Total opportunity: ~$45M/year
• Y1 net value: ~$11M
• Y1 ROI band: contract-dependent (typically 15×–30× at this scale, given billable-delivery compounding)

Decision: Strategic engagement. Dedicated ops engineer, data residency in multiple regions, custom SLA aligned to internal change-management cadence. Multi-year contract typical at this scale.

Recommended sequence: Billable delivery (utilisation + hours-per-job) first — the compounding case is structural; this is where the Klarna-style efficiency story meets the JPMorgan-style scale story. Bench reduction and back-office second. Recruiting / BD third.

§7.6 — Pre-IPO Enterprise SaaS ($1.5B revenue, Aggressive)

Profile: 6,000-person SaaS company. Multiple product lines, large enterprise customer base, regulated industries among customers, IPO timeline within 18 months.

Diagnostic output:

• Recommended volume: ~2,500,000 invocations/month
• Vehicle: Custom contract with multi-region residency
• Total opportunity: ~$75M+/year
• Y1 net value: ~$18M+
• Y1 ROI band: 15×–25× (negotiated contract)

Decision: Strategic, multi-stakeholder engagement. The contract structure is procurement-led; the deployment is COO-led; the compliance posture is CISO + GC led. Multi-year, multi-region residency, custom SLA, dedicated ops engineer is table stakes at this scale.

Critical considerations:

• EU AI Act compliance (§11) for any AI affecting employment, credit, education, law enforcement contexts in your customer base.
• Audit substrate must be in place pre-IPO for S-1 disclosure posture.
• Workflow-redesign focus (McKinsey 2026) — this is where the 55% vs 20% redesign-rate gap shows up at scale.

§8 — The Klarna Lesson: When Aggressive Becomes Wrong

Klarna's 2024–2025 trajectory is the public case study every enterprise AI decision should be calibrated against. The arc compresses both the upside and the failure mode into 18 months.

Table 8.1 — The Klarna Timeline

Source: Klarna public statements; FintechWeekly (2025); Internative Klarna AI Reversal Postmortem; Solutions Review; Acefone.

What Klarna got right (and the diagnostic captures):

• Picked one workflow (CS) that was genuinely high-routine.
• Measured impact in dollar terms ($60M saved is independently verifiable).
• Maintained a public deployment posture that the rest of the market could learn from.

What Klarna got wrong (and the diagnostic's structure warns against):

• Over-rotated to AI for interactions where judgment, empathy, or compliance nuance mattered. Confident-but-wrong answers about policy, fees, or payment terms became a regulatory and brand issue simultaneously.
• Treated the AI deployment as a labour-substitution exercise rather than a capability-augmentation exercise. The CEO's later framing — "focused too much on efficiency" — is the diagnostic of the failure mode.

Table 8.2 — The Hybrid Architecture That Held Up

§9 — When NOT to Buy

A diagnostic that never recommends not buying is a sales tool, not a decision framework. Below are the conditions under which the diagnostic's output should be discounted or ignored.

Table 9.1 — Conditions That Argue Against Buying Now

§10 — Reading the Diagnostic Like an Analyst

The diagnostic produces five numbers and one band. Below is what each means and how to use it.

Table 10.1 — Field-by-Field Interpretation

§11 — Compliance: The August 2026 Deadline

The EU AI Act's high-risk AI system enforcement begins August 2, 2026. For any organisation deploying AI in employment, credit, education, or law-enforcement contexts — directly or as a vendor — the compliance posture is no longer optional.

Table 11.0 — EU AI Act Enforcement Timeline

The deadline is not a single cliff. It is the largest of several enforcement gates that began rolling out in February 2025 and continue through 2027.

Source: EU AI Act (Regulation 2024/1689) Article 113 (Entry into force and application); European Commission AI Act enforcement schedule.

Table 11.1 — Core EU AI Act Requirements for High-Risk Systems

Table 11.2 — Penalties

Source: EU AI Act (Regulation 2024/1689); Tredence EU AI Act Compliance Guide 2026; Raconteur Technical Audit Guide.

§12 — Recommended Next Steps by Role

Table 12.1 — One-Page Action by Role

Appendix A — Methodology

The diagnostic's invocation count is derived from:

recommended_invocations_per_month
  = annual_revenue ($M)
  × invocation_density_for_industry  (Goods 500, SaaS 1,200, Services 1,000)
  × ambition_multiplier               (Conservative 0.6, Base 1.0, Aggressive 1.4)

The total opportunity is computed by summing per-function impacts:

impact_per_function
  = base_amount                       (rev / cogs / ga / rd × revenue)
  × interpolated_rate                 (between fn.lo and fn.hi based on ambition)
  × (routine_pct / 0.5)               (routine share — caller-overridable)
  × (automation_ceiling / 0.6)        (per error-tier: low 0.78, med 0.60, high 0.40)
  × governance_factor                 (per visibility tier — internal / customer / regulator)

total_opportunity = sum across functions of impact_per_function

Year-1 net value is computed conservatively:

year_1_capture_rate = 0.25
services_cost ($M) =
  0.000  if revenue <   $10M
  0.015  if $10M ≤ revenue <  $50M
  0.050  if $50M ≤ revenue < $250M
  0.100  if revenue ≥ $250M and self-serve
  0.300  if custom contract

year_1_net_value = max(0, total_opportunity × year_1_capture_rate − services_cost)

Year-1 ROI band:

y1_roi_multiple = (year_1_net_value × 1,000,000) / annual_platform_cost
y1_roi_band     = { lo: max( 5, floor(y1_roi_multiple × 0.8)),
                    hi: min(60, ceil (y1_roi_multiple × 1.1)) }

The 5× floor and 60× ceiling are not statistical bounds — they are the empirical range where public 2024–2026 enterprise AI deployments published their actual outcomes. The clamp prevents the diagnostic from emitting numbers that are mathematically possible but not credibly defensible against the public benchmark set.

Appendix B — Sources

Market sizing and adoption:

• Gartner, "40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026" (Aug 2025).
• Gartner, "Over 40% of Agentic AI Projects Will Be Cancelled by End of 2027" (June 2025).
• Gartner, 2026 Hype Cycle for Agentic AI.
• Gartner, Strategic Predictions for 2026.

Enterprise AI ROI and outcomes:

• McKinsey, "The State of AI in 2025: Agents, Innovation, and Transformation."
• McKinsey, "State of AI Trust in 2026: Shifting to the Agentic Era."
• McKinsey, "The State of AI: How Organisations Are Rewiring to Capture Value."
• MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025" (July 2025).
• AI Monk, "12 Agentic AI Examples with Measurable ROI: Enterprise Case Studies 2025–2026."

Case studies:

• Emerj, "Artificial Intelligence at JPMorgan Chase."
• FintechWeekly, "Klarna Reverses Course on AI Customer Support, Resumes Human Hiring" (2025).
• Internative, "Klarna's AI Reversal: A Postmortem in 3 Lessons" (2025).
• Solutions Review, "Klarna's AI Layoffs Exposed the Missing Piece: Empathy" (2025).
• Acefone, "Klarna's AI U-Turn: Key Learnings about Reliable AI Adoption in CS" (2025).

Pricing and market structure:

• Bessemer Venture Partners, "The AI Pricing and Monetisation Playbook" (2026).
• Valueships, "AI Pricing in 2026: SaaS Pricing Models That Actually Work."
• MindStudio, "SaaS Pricing Is Breaking: Why Per-Seat Models Don't Survive the AI Agent Era."
• Awesome Agents, "Agent Platform Pricing Compared 2026."
• LangChain published pricing (2026).

Regulatory:

• European Union, AI Act Regulation 2024/1689 (Annex III, Articles 9, 11, 12).
• Tredence, "EU AI Act 2026 Compliance Guide for US Companies."
• Raconteur, "EU AI Act Compliance: A Technical Audit Guide for the 2026 Deadline."
• SecurePrivacy, "EU AI Act 2026: Key Compliance Requirements for Enterprises."

SMB-specific findings:

• Reinventing.ai, "AI Agent Trends: ROI Pressure Pushes Enterprises to Orchestration, While SMB Adoption Accelerates" (March 2026).
• Glivera, "Agentic AI for Small Business" (2026 guide).
• Aalpha, "AI Agents for Small Businesses — In-Depth Guide" (2026).

Appendix C — Glossary

• Invocation — One agent receiving an event, completing its task, and publishing its output. A single business workflow typically routes through 5–10 invocations end-to-end.
• Invocation density — Industry-specific multiplier expressed as invocations per $M of annual revenue per month. Reflects multi-agent chains per business event in that industry.
• Routine share — The portion of a function's pool that is rule-bound or pattern-bound enough for an agent to carry. Differs by function within an industry.
• Automation ceiling — The maximum portion of routine work an agent can carry hands-off before human-in-the-loop is required. Depends on the cost of an error in that function (low / medium / high).
• Governance factor — The portion of eligible routine work that actually moves to agents in a given workflow, given the controls that workflow needs and the controls the buyer can put in place.
• Self-serve ceiling — 150,000 invocations / month. Below this, the published rate applies. Above this, a custom contract is required.
• Year-1 capture rate — The portion of annual opportunity realistically captured in the first 12 months. Calibrated at 25% based on enterprise deployment ramp curves.
• Year-1 ROI band — Year-1 net value divided by annual platform cost, expressed as a range clamped to 5×–60× (the range where real enterprise AI deployments actually land).
• Services cost — One-time implementation engagement cost. Scales with company size: $0 (<$10M revenue) → $300K (custom contract).
• Self-serve plan — The published $99/month + $0.06/invocation subscription. Available below the 150K invocations/month ceiling.
• Custom contract — Negotiated annual contract above the self-serve ceiling. Volume-discounted invocation pricing, dedicated ops engineer, custom SLA, data residency.
• The 95% — MIT Project NANDA's finding that 95% of enterprise GenAI pilots show zero measurable P&L impact, primarily due to integration and workflow-redesign failures (84%) rather than technology failures.
• The 5% — The minority of organisations achieving rapid, measurable impact. Characterised by buying from specialised vendors (67% success rate vs 33% for in-house build), redesigning workflows around the agent (55% vs 20% for non-top-performers), and starting with one painful, well-defined function rather than horizontal deployment.
• Human-in-the-loop (HITL) — A workflow design where an agent's output is reviewed, approved, or modified by a human before it takes effect downstream. The complement of an "autonomous" workflow. Required wherever the cost of an agent error exceeds the cost of human review time.
• Guardrail — A runtime constraint applied to an agent's output (or input) that blocks, rewrites, or escalates the action when the output violates a policy. Examples: PII redaction, output-class restriction, regulatory disclosure language, refusal to make policy interpretation statements. Guardrails are the operational implementation of Article 14 (human oversight) and Article 15 (accuracy/robustness) of the EU AI Act.
• Policy artefact — A machine-readable document defining what an agent may and may not do in a specific workflow context, who can override which decisions, what audit information is captured per decision, and what fallback path applies when a decision exceeds the agent's authority. Policy artefacts are the substrate that makes Article 11 technical documentation and Article 12 event logging produceable at scale.
• Audit substrate — The persistent, per-decision record of every agent action: input, output, policy version, model version, intervening human approvals, downstream effect. The audit substrate is the layer that distinguishes a governed agent platform from a bolt-on monitoring tool — it is captured at runtime, not retrofitted from logs.
• Agent-density category — A software category in which the unit of work is an event (a claim, a ticket, a deal, a query) rather than a user. Per-seat pricing inverts in agent-density categories because one agent does the work of many seats; usage- or invocation-based pricing dominates.