The governing rule
The client supplies the money; we supply the engineering deltas and the method. No currency figure appears in a deliverable unless the client's finance function sourced or ratified the input. That is not caution for its own sake — it is what makes the number survive contact with the CFO.
The model shape
- Net value
- Σ value drivers − Σ costs − instability tax
- First-year ROI
- net value ÷ total first-year investment
- Payback
- months until cumulative net value crosses zero, after the dip
Value drivers
| Driver | What we measure | Claim discipline |
|---|---|---|
| Capacity released | Lead-time and wait reduction | Only value if reinvested — state explicitly what it was spent on, or finance will |
| Rework reduction | Rework share, revert rate, churn | The most defensible driver: measurable without heroic assumptions |
| Incident cost avoidance | Change failure rate, recovery time | Client's incident cost model only, never benchmark-derived |
| Time to market | One named flagship commitment | Specific beats comprehensive |
| Onboarding and knowledge access | Time to first meaningful contribution | Supported by the evidence that AI helps most on unfamiliar code |
| − Instability tax | Increase in change failure rate, recovery time, unplanned work | A mandatory negative line, even when it is zero |
Costs
- Licences and seats — including unused seats; they are real cost.
- Inference and infrastructure where usage-priced.
- The verification tax: reviewer minutes × loaded cost. The cost everyone forgets; naming it is a credibility moment.
- Enablement and change time, including deliberate unassisted practice.
- Platform and pipeline investment — often the largest and most durable item.
- Governance, audit and security: the real burden of adoption has shifted here.
- Consulting fees. Omitting them invites exactly one question.
- The J-curve dip, shown as an explicit line and never buried.
The sensitivity table is mandatory
| Assumption | Low | Base | High |
|---|---|---|---|
| J-curve dip depth | 10% | 15% | 25% |
| J-curve duration | 2 months | 3 months | 12 months |
| Share of released capacity reinvested | 25% | 50% | 80% |
| Verification tax per AI-assisted PR | measured −30% | measured | measured +50% |
| Change-failure-rate movement | improves 10% | flat | worsens 20% |
Present the range before the point estimate
A CFO who sees the range first trusts the estimate; one who sees the estimate first interrogates it. The twelve-month duration column is the one that changes the conversation — in DORA's own calculator it flips an illustrative positive case to a negative first-year return.
When the client has no cost data
- Physical units onlyEngineer-days released, incident hours avoided, rework share reduced. Many boards are satisfied by this and it is fully defensible.
- A client-chosen placeholder rateApplied transparently with the arithmetic shown, so finance can substitute their own figure. Their number, their model, our method.
- Defer the money modelTo the quarter-two review, and use quarter one to establish the physical deltas. This is the usual recommendation.
Never
Fill the gap with an industry average. A single unsourced currency figure contaminates every number beside it.