AI ROI Model
Every dollar spent on AI is a dollar not spent elsewhere. Compared to what?
Only 7% of enterprises see significant ROI from AI. Not because AI doesn't work — because they calculate two cost inputs and skip five. A pilot that looks profitable on build cost plus model usage fails commercially when integration, maintenance, change management, data migration, and operational complexity are added. This model captures all seven inputs so the capital allocation decision is honest.
The metric that matters is not time saved. Time saved only matters when it maps to a P&L line. The North Star is revenue per employee — and whether AI moves it toward software-company margins.
0. Decision Frame
Before building the model, answer the framing questions.
| Question | Answer |
|---|---|
| Which workflow is being evaluated? | [specific workflow from Constraint Map] |
| Is the workflow classified as Artifact or Hybrid? | Artifact / Hybrid — if Real, stop here |
| Volume: how many times per month? | [number — if < 25/month, flag for review] |
| Current cost: who does it, at what hourly rate, for how long per unit? | [$X/hr × Y hrs × Z units/month] |
| What business metric does this workflow directly affect? | [revenue, margin, cycle time, capacity — not "time saved"] |
| What is the opportunity cost of capital? (what else could this money do) | [name the alternative investment] |
Low-frequency flag: workflows occurring fewer than ~25 times per month rarely justify a custom AI system. The build cost does not pay back. State the threshold before proceeding.
1. Complete Cost Model
The two inputs most pilots use — and the five they skip.
Always Counted
| Cost Input | One-Time | Monthly Ongoing | Conviction | Notes |
|---|---|---|---|---|
| Build cost (design, development, testing) | $[X] | — | H/M/L | Internal or vendor time |
| Model usage (API calls, tokens, inference) | — | $[X] | H/M/L | Scale with volume |
Commonly Skipped
| Cost Input | One-Time | Monthly Ongoing | Conviction | Notes |
|---|---|---|---|---|
| Integration costs (API access, data connectors, existing system hooks) | $[X] | $[X] | H/M/L | Can reach $60K/yr — verify before assuming cheap |
| Maintenance (prompt updates, model version changes, monitoring) | — | $[X] | H/M/L | Ongoing, not zero |
| Change management (repositioning people, retraining, communication) | $[X] | — | H/M/L | Often exceeds build cost in mature orgs |
| Data migration (cleaning, formatting, loading historical context) | $[X] | — | H/M/L | Underestimated in 90% of projects |
| Operational complexity (new failure modes, monitoring, incident response) | — | $[X] | H/M/L | Inverse of reliability |
Total Cost Summary
| Year 1 | Year 2 | Year 3 | |
|---|---|---|---|
| One-time costs | $[X] | — | — |
| Monthly ongoing × 12 | $[X] | $[X] | $[X] |
| Total | $[X] | $[X] | $[X] |
2. Benefit Model
Map every benefit to a P&L line. If it does not connect to revenue, margin, or cost — name it as a soft benefit and do not count it in the primary calculation.
Hard Benefits (P&L connected)
| Benefit | Monthly Value | How Calculated | P&L Line | Conviction |
|---|---|---|---|---|
| Headcount reduction or redeployment | $[X] | [roles × fully-loaded cost] | Payroll | H/M/L |
| Cycle time reduction (maps to revenue if it accelerates billing or deal close) | $[X] | [days saved × deals/month × deal value × close rate] | Revenue | H/M/L |
| Capacity increase (more units handled with same headcount) | $[X] | [additional units × margin per unit] | Revenue / Margin | H/M/L |
| Error reduction (rework cost, compliance penalty avoided) | $[X] | [error rate reduction × cost per error] | Ops cost | H/M/L |
Soft Benefits (do not count in primary calculation)
| Benefit | Why it is soft | When it might become hard |
|---|---|---|
| "Time saved" (unspecified) | Time saved only matters if it is redirected to a revenue-generating activity | When redirected hours are tracked and linked to new output |
| Employee satisfaction | Real but not measurable in ROI terms | When turnover cost is calculated and the link is documented |
| "Better decisions" | Requires before/after decision quality measurement to quantify | When decision quality is tracked as a KPI |
3. Revenue Per Employee Benchmark
The North Star metric. Tracks whether AI is moving the business toward software-company margins.
| Metric | Current | 12-Month Target | 24-Month Target |
|---|---|---|---|
| Annual revenue | $[X] | $[X] | $[X] |
| Full-time equivalent headcount | [X] | [X] | [X] |
| Revenue per employee | $[X] | $[X] | $[X] |
| Industry benchmark (if known) | $[X] | — | — |
Reference points:
- Labor-intensive service business: $100–150K/employee
- Consulting or agency: $150–250K/employee
- AI-augmented service business (target): $400K+/employee
- Software company: $500K–$1M+/employee
The transformation goal is not to reach software margins. It is to move the ratio meaningfully with a defined capital investment.
4. Return Calculation
| Metric | Year 1 | Year 2 | Year 3 | 3-Year Total |
|---|---|---|---|---|
| Total hard benefits | $[X] | $[X] | $[X] | $[X] |
| Total costs | $[X] | $[X] | $[X] | $[X] |
| Net return | $[X] | $[X] | $[X] | $[X] |
| ROI % | [%] | [%] | [%] | [%] |
| Payback period | [months] | — | — | — |
5. Build / Don't Build Decision
Apply the checklist before committing capital.
| Gate | Check | Pass? |
|---|---|---|
| Volume ≥ 25 units/month | [volume] | YES / NO |
| Integration cost < annual benefit | [$X vs $X] | YES / NO |
| Business logic is defined | [documented or not] | YES / NO |
| Success criteria are measurable | [KPI named] | YES / NO |
| Change management cost included | [$X budgeted] | YES / NO |
| Net 3-year return is positive | [$X] | YES / NO |
| Capital allocation beats alternatives | [vs $X in alternative] | YES / NO |
Decision: BUILD / DON'T BUILD / INVESTIGATE FURTHER
If any gate fails, document why before proceeding. A gate failure is not a blocker — it is a prerequisite to resolve.
6. When AI Does Not Make Sense
State these explicitly before the model is presented to leadership.
| Condition | Why it fails |
|---|---|
| Workflow < 25/month | Build cost doesn't pay back at low frequency |
| Integration cost > annual benefit | Math fails regardless of model quality |
| Business logic undefined | You end up designing the logic and the AI simultaneously — scope creep, no ROI |
| Efficacy drop from 100% → 60% costs more than the saving | Client-facing judgment tasks where partial AI quality destroys trust |
| Change management cost exceeds automation benefit | Mature organisations with entrenched roles — cultural cost is real |
The consultants who say "don't build an AI system here" build more trust than those who say yes to everything.
Context
- Constraint Map — Identifies which workflows to model here
- AI-Native Future State — What the benefit side of the model is buying
- Transformation Roadmap — How modules sequence to compound ROI
- Unit Economics — The per-unit math that underpins the capacity benefit calculation
- Cash Flow Projection — Where the build cost hits the P&L
Links
- Theory of Constraints — Goldratt — The compounding flywheel: unlock one constraint, reinvest into the next
- Return on Investment — Standard ROI calculation base
- Capital Allocation — The frame this model sits inside
Questions
Which of the seven cost inputs does your current AI pilot skip — and does the math still work when all seven are counted?
- If "time saved" is your primary benefit metric, what P&L line does it connect to?
- What is your current revenue per employee — and what would it need to be in 24 months for this investment to have been worth making?
- When integration costs more than the annual benefit, what is the decision — and who makes it?