Key Takeaways
- Most enterprise AI programmes cannot demonstrate P&L impact because they never measured a baseline — ROI starts with unit economics captured before deployment.
- Full cost accounting includes model consumption at scale, evaluation infrastructure, human oversight, integration and rework — not just licence fees.
- Four metrics survive finance scrutiny: fully-loaded cost per completed task, cycle time, first-pass yield and capacity redeployed.
- “Hours saved” only become savings when capacity is reduced or redeployed to revenue work — otherwise they are comfort, not cash.
- Realistic benchmarks: 30–60% process cost reduction and 3–10x throughput on well-scoped workflows within 6–12 months.
- ROI follows a curve — negative in pilot, breakeven in production hardening, compounding at scale — and needs quarterly portfolio review like any capital allocation.
Boardrooms have flipped from “why aren’t we doing AI?” to “where is the money?”. Research groups tracking enterprise generative-AI adoption — from MIT-affiliated studies to the large consultancies — keep landing on the same uncomfortable finding: a large majority of enterprise AI pilots produce no measurable P&L impact, not because the technology fails, but because nobody instrumented the economics. Meanwhile the organisations that did instrument them report 30–60% cost reductions on targeted workflows.
The difference is not model choice. It is measurement discipline. Here is the finance-grade framework we use with CFOs and CIOs.
Why AI Business Cases Fail Finance Scrutiny
Three recurring defects sink credibility:
- No baseline. The “before” state is estimated after deployment, from memory, by the team that owns the project’s success.
- Partial cost accounting. The business case prices licences and ignores consumption, oversight and rework.
- Phantom savings. “10,000 hours saved” that never left the payroll and never moved to revenue work.
Each defect is fixable — before deployment, cheaply; after deployment, painfully.
Step 1: Baseline the Unit Economics First
For each process an agent will touch, capture at least 30 days of:
- Fully-loaded cost per transaction — labour, tooling, error remediation, allocated overhead.
- Cycle time — request to completion, end to end, including queue time.
- Error and rework rate — how often output needs correction, and what correction costs.
- Volume distribution — peak patterns, seasonality, backlog behaviour.
Thirty days of measurement prevents twelve months of argument. It also frequently reveals that the process should be fixed before it is automated — the theme of our piece on the AI process layer.
Step 2: Price the Full Cost Stack
| Cost Line | What It Includes | Typical Share of Year-1 TCO |
|---|---|---|
| Platform & licences | Agent framework, orchestration, vendor fees | 15–25% |
| Model consumption | Per-task inference at production volume | 15–30% |
| Evaluation infrastructure | Test harnesses, golden datasets, regression suites | 10–15% |
| Human oversight | Review queues, exception handling, approvals | 15–25% |
| Integration & data | APIs, RAG indexes, pipeline maintenance | 15–20% |
| Rework from agent errors | Human correction of failed tasks | 5–15% (falls with reliability) |
Two lines deserve special attention. Consumption looks trivial in pilots and compounds brutally at volume — model routing and small-model substitution (see our analysis of small language models) are the levers. Oversight is a feature, not a bug — but it must be priced, and its cost falls as reliability climbs the march of nines.
Step 3: Track the Four Metrics That Survive Audit
- Cost per completed task (fully loaded). The headline number — includes oversight and rework, compared directly against the baseline.
- Cycle time. Hours-to-minutes improvements often carry more business value than labour savings: faster procurement cycles, faster customer resolution, faster closes.
- First-pass yield. The share of tasks completed correctly without human correction — the single best proxy for agent reliability economics.
- Capacity redeployed. The honest version of “hours saved”: headcount-equivalent capacity actually reduced or reassigned to revenue-generating work, signed off by the function head.
What Good Looks Like: Field Benchmarks
| Workflow | Typical Cost Reduction | Typical Payback |
|---|---|---|
| Document processing & extraction | 50–70% | 2–4 months |
| Customer service resolution | 30–50% | 3–6 months |
| Finance operations (AP/AR, reconciliation) | 30–50% | 4–6 months |
| IT service management (L1/L2) | 40–60% | 3–5 months |
| Procurement (drafting & evaluation) | 50–70% effort reduction | 1–2 tender cycles |
The procurement line is not hypothetical — it is what platforms like BidsInsight deliver by compressing RFP drafting ~70% and evaluation effort ~60% on real tenders.
The ROI Curve: Set Expectations by Phase
- Pilot (months 0–3): negative ROI by design — you are buying evidence, not savings. Success metric: validated first-pass yield on real volume.
- Production hardening (months 3–6): approach breakeven as reliability climbs and oversight narrows to exceptions.
- Scale (months 6–12+): compounding returns as fixed costs amortise across workflows and consumption is optimised.
Business cases that promise linear savings from month one get rejected by finance — and deserve to be.
Five Traps That Inflate AI Business Cases
- Claiming saved hours without capacity redeployment sign-off.
- Pricing pilot-scale consumption into production-scale volume.
- Ignoring rework: at 95% reliability, 5% of volume needs human correction — price it.
- Double-counting savings already claimed by an RPA or transformation programme.
- Skipping the counterfactual: some improvement would have happened anyway; attribute only the delta.
Governance: Review AI Like a Portfolio
Agent scope expands, models change, volumes shift. Review quarterly against the original baseline: re-price consumption, re-verify first-pass yield, retire agents that stopped earning their keep, and reallocate to workflows with better curves. AI portfolios need the same discipline as any other capital allocation — and the CFO should chair the review.
We help finance and technology leaders build these business cases with measured baselines through our IT Consulting & Advisory and Agentic AI practices. The best AI investment is the one you can defend in the board meeting after next.
Frequently Asked Questions
How do you calculate ROI on agentic AI?
Capture a 30-day baseline of fully-loaded cost per transaction, cycle time and error rate before deployment. After deployment, compare the same metrics including all AI costs — consumption, oversight, evaluation infrastructure and rework — and attribute only the delta. Review quarterly as scope and volumes change.
What ROI do enterprises actually achieve with AI agents?
Well-scoped deployments typically deliver 30–60% process cost reduction and 3–10x throughput within 6–12 months. Document processing reaches 50–70% savings with 2–4 month payback; procurement drafting and evaluation reach 50–70% effort reduction within one to two tender cycles.
What hidden costs should a CFO include in an AI business case?
Model consumption at production volume, evaluation and testing infrastructure (10–15% of TCO), human oversight and exception handling (15–25%), integration and data pipeline maintenance, and rework from agent errors. Oversight cost is the most commonly omitted line.
Why do most enterprise AI pilots show no P&L impact?
Research consistently attributes it to measurement failure rather than technology failure: no pre-deployment baseline, partial cost accounting, and “hours saved” that were never redeployed or reduced. Pilots also stall when broken processes are automated as-is or reliability stays below production grade.
When should an AI agent be retired or re-scoped?
At quarterly portfolio review: if first-pass yield declines, consumption outgrows value, or the workflow volume no longer justifies fixed costs, re-scope the agent, route tasks to cheaper models, or retire it. Treat AI like any capital allocation — with exit discipline.
