| name | ai-agent-business-case-and-roi |
| description | Use when building a task-based business case for an AI agent, including intervention-adjusted benefits, the full agent cost stack, downside scenarios, and autonomy-ramp payback. |
| metadata | {"portable":true,"compatible_with":["claude-code","codex"]} |
AI-Agent Business Case and ROI
Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178.
Use When
- The proposal must justify the price of an agent or a multi-agent system to a CFO, head of operations, or investment committee.
- The buyer has been promised "AI does the work" and the CFO now wants the numbers.
- The pricing rests on per-resolution, per-outcome, per-step, or success-based commercials and the case must show that the model can pay for itself at the proposed rate.
- The buyer's CFO will challenge: intervention overhead, irreversibility cost, scope-creep cost, model-call cost per task, supervisor cost, escalation cost, and the cost of being wrong.
Do Not Use When
- The case is for a non-agent AI feature (use
ai-on-saas-business-case-and-roi).
- The case is for generic AI strategy with no measurable outcome workflow (use
ai-transformation-proposal).
Domain Inputs
- Baseline cost per outcome today (cost per resolved ticket, per triaged claim, per reconciled batch, per drafted document).
- Volume per month per tenant.
- Headcount cost (loaded), ramp curve, attrition.
- Average handling time (AHT) and time-to-resolution baseline, with P50 / P95 distribution.
- Model-call cost per task at the candidate provider(s) — agents make many calls per task, not one.
- Target autonomy level by quarter, and the intervention rate at each level.
- Supervisor cost (oversight queue staffing, on-call, escalation engineer).
- Eval, red-team, drift-watch cost.
- Insurance / liability cost differential for agentic actions.
- Buyer regulator exposure (downside).
- FX assumptions (USD model costs, local revenue).
Domain Method
-
Build the Agent Value Stack for each agent or use case: cost per outcome saved (volume × baseline cost − agent cost), AHT reduction monetised, deflection-rate monetised (work that never reached a human), time-to-resolution monetised (revenue cycle effect or SLA-breach avoidance), capacity unlock (volume the buyer can now handle without hiring), revenue from agent-tier upsell where the agent is a product feature.
-
Build the Agent Cost Stack: model calls per task × tasks per month (this is the biggest line and the most overlooked), embedding cost, tool-call cost (search, third-party API), oversight cost (supervisor headcount or on-call), eval cost (golden set + LLM-as-judge + human review), red-team cost, drift-watch cost, engineering cost, integration cost, vendor liability / insurance differential, cloud compute, vector store / memory.
-
Compute Tasks-per-FTE-Saved with Intervention Discount:
FTE saved = (tasks completed by agent × (1 − intervention rate)) / (FTE annual task capacity)
Intervention rate must come from the pilot, not from the brochure. An agent with 60 % autonomy at go-live saves less than half what a 90 % autonomy agent saves — and supervisor cost rises with intervention rate.
-
Compute Cost per Outcome at three intervention rates (P50 / P90 / P99 pessimistic), so the proposal is honest about the curve as autonomy improves over time.
-
Compute Three Scenarios by autonomy ramp (downside, base, upside):
- Downside — autonomy ramp stalls; intervention rate stays high; one minor irreversible incident; one model-price increase.
- Base — autonomy ramp tracks the pilot; intervention rate falls month over month within plan; zero irreversible incidents.
- Upside — autonomy ramp accelerates; intervention rate halves at month six; volume expansion across tenants.
-
Compute Payback in months, accounting for model-call cost per task, supervisor cost, and the autonomy ramp. Payback that ignores model-call cost is fiction.
-
Compute NPV over 36 months with the buyer's discount rate and sensitivity on (a) model-call price (±30 %), (b) intervention rate (±10 percentage points), (c) volume (±20 %).
-
Write the Downside Scenarios:
- (a) Irreversible-action incident — one wrong agent action sized at the buyer's worst-case (refund, fine, public statement, lost customer). Mitigation links to L1 / L2 gating and Responsible-AI commitment.
- (b) Regulator action on agentic system — pause, conformity assessment, ban in market. Mitigation links to compliance and procurement.
- (c) Intervention overhead overshoots — supervisor headcount exceeds plan; net savings collapse. Mitigation links to eval and prompt regression cadence.
- (d) Scope-creep cost — buyer adds use cases without adding pricing; engineering and ops drown. Mitigation links to the action catalogue gate and change-control clause.
- (e) Model-price shock — provider raises prices or restricts use. Mitigation links to model-gateway and fallback.
- (f) Adoption stall — staff resist; volume routed to agent stays low; payback delays. Mitigation links to change management.
- (g) Liability event — a court / regulator / contract holds the buyer or agency liable. Mitigation links to insurance, contract, audit log.
- (h) Service-credit cost — credits issued under the SLA Class credit schedule and intervention-credit clause reduce realised revenue. Model the worst-credit-month at P90 / P99 intervention, P90 / P99 task-success miss, and a kill-switch breach. Stress-test margin floor against the aggregate credit cap. See ai-agent-sla-and-credit-schedule and ai-agent-intervention-credit-and-abort-refund.
- (i) Refund-trigger event — model the engagement-level loss if an Abort-and-Refund trigger fires (Irreversible-Action Incident at agency fault; intervention overshoot for 60 consecutive days; regulator action). The contingency reserve should cover at least the worst single-trigger refund scenario.
-
Output the Agent Business Case Memo in the buyer's shape (CFO one-pager, board memo, investment-committee deck).
The Tasks-per-FTE-Saved Formula (worked illustration)
Assume:
- Volume: 100,000 tickets / month
- Baseline AHT: 12 minutes; FTE handles 8,000 / month
- Baseline FTE need: 12.5
- Pilot intervention rate: 25 % at month 1; target 10 % at month 6
- Tickets the agent attempts: 80 % of volume = 80,000
At month 1:
- Agent-completed (no intervention) = 80,000 × (1 − 0.25) = 60,000
- FTE saved = 60,000 / 8,000 = 7.5 FTE
At month 6:
- Agent-completed = 80,000 × (1 − 0.10) = 72,000
- FTE saved = 72,000 / 8,000 = 9.0 FTE
The CFO is buying a curve, not a step change. The proposal commits to the curve.
Quality Standards
- Value stack and cost stack are presented side by side; intervention rate is named.
- Model-call cost per task is computed, not omitted as "cloud cost".
- Supervisor cost is computed; the agent does not get free oversight in the model.
- Cost per outcome is given at three intervention rates.
- Three scenarios reflect autonomy ramp, not just demand variation.
- Downside scenarios include an irreversible-incident dollar figure.
- FX is named where revenue is local and model costs are USD.
- The discount rate is the buyer's stated rate.
- The proposal does not quote tasks-per-FTE without subtracting intervention.
Domain Risks
- "The agent saves 10 FTE" with no intervention rate.
- "Cost per ticket falls 80 %" with no model-call cost subtracted.
- Three scenarios that vary only volume, not autonomy.
- Irreversible-action incident treated as "we will be careful".
- Supervisor cost zeroed because "the agent runs itself".
- Payback computed at month 1 autonomy assumed to last forever.
- Scope-creep cost ignored; the case assumes the catalogue does not grow.
- Liability event excluded because "the buyer accepts the risk".
Domain Outputs
- Agent Value Stack table.
- Agent Cost Stack table.
- Tasks-per-FTE-Saved with Intervention Discount (worked at month 1 and month 6).
- Cost per Outcome at P50 / P90 / P99 intervention.
- Three-Scenario ROI (downside / base / upside) keyed to autonomy ramp.
- Payback in months and NPV at 36 months.
- Sensitivity on model-call price, intervention rate, volume.
- Downside Scenarios with mitigations.
- Agent Business Case Memo in the buyer's preferred shape.
Anti-Patterns
- Inventing a metric, credential, constraint, or buyer position. Fix: cite the supplied source or mark the item as an assumption requiring confirmation.
- Treating an unavailable check as passed. Fix: mark it not assessed and state the evidence needed to resume.
- Advancing autonomy without a named gate owner. Fix: require observable evidence, accountable acceptance, and a rollback path.
- Reusing another sector or use case without reassessment. Fix: retest affected parties, action scope, reversibility, and jurisdiction.
- Writing acceptance as “satisfactory” or “appropriate”. Fix: define an observable measure, threshold, evidence record, and decision owner.
Inputs
| Artefact | Source/provider | Required? | Missing-input behaviour |
|---|
| task baseline, loaded labour cost, intervention data, agent cost stack, and buyer discount rate | Buyer evidence, ToR, approved discovery record, system owner, or measured operating data | Yes | Stop the affected decision; list the missing source and return only a qualified outline or assumption register. |
Outputs
| Artefact | Consumer | Acceptance condition |
|---|
| Agent business-case model and decision memo | Buyer CFO and proposal evaluator | Scope, assumptions, exclusions, owners, decision logic, and observable acceptance tests are explicit and traceable to supplied evidence. |
Evidence Produced
| Evidence | Consumer | Acceptance condition |
|---|
| agent business-case model and decision memo | Buyer CFO and proposal evaluator | Every load-bearing claim traces to supplied evidence; assumptions, owners, gates, exclusions, and observable acceptance conditions are explicit. |
Capability Contract
Default to read-only for discovery, analysis, review, and planning. Minimum capability is access to the supplied artefacts and permission to calculate or inspect evidence. Edit only the requested proposal working copy. Do not change production systems, contact affected parties, publish, spend, certify compliance, or approve autonomous action without explicit authority from the accountable owner.
Degraded Mode
If files, interviews, telemetry, specialist review, network access, or calculation tools are unavailable, produce the narrowest useful qualified result. Mark each unavailable check as not assessed, separate facts from assumptions, lower confidence, and state the evidence needed to resume. An unassessed gate is never a pass.
Decision Rules
| Choice | Action | Failure or risk avoided |
|---|
| Select investment case | Model downside, base, and upside using measured task and intervention inputs. | A persuasive but uneconomic ROI. |
| Required evidence, authority, or accountable owner is missing | Stop the affected recommendation or commitment and record the gap. | Invented evidence or unauthorised autonomy. |
| Gate evidence is complete and accepted | Advance only within the approved scope and retain the evidence trace. | Scope drift and irreproducible approval. |
Workflow
- Confirm the consumer, authority, neighbouring-skill route, and required inputs; stop when a mandatory source or accountable owner is missing.
- Inspect the evidence and record facts, assumptions, conflicts, and unavailable checks; stop on a failed safety, finance, regulatory, or acceptance gate.
- Apply the domain method and decision rules within the qualified scope, retaining an evidence trace.
- Draft the contracted output and reconcile it with methodology, work plan, staffing, pricing, risk, and governance; recover by revising the affected scope or control and rerunning the failed gate.
- Verify acceptance conditions, permission boundaries, direct references, and anti-slop controls; block release until failed checks are corrected.
Worked Example
A service team handles 40,000 tasks monthly. Model benefit after supervisor intervention, model calls, evaluation, operations, and credits; show payback at the buyer's discount rate and flag any unverified baseline.
References
