AI value
How I measure AI value.
Two instruments. The Amplifier Test screens every AI use case before money is spent. Net AI Value estimates the ones that pass, in dollars.
The principle
AI amplifies whatever it touches.
On a sound process it compounds value. On a broken one it scales the errors, at full speed.
The same tool and the same vendor can be an asset in one company and a liability in the next. The process decides. So I test what is already there before anyone spends on AI.
Google's DORA research found the same in software teams: AI acts as an amplifier of a team's existing strengths and weaknesses. I apply it to the P&L and the deal.DORA (Google Cloud), State of AI-assisted Software Development, 2025.
AI high performers, about 6% of respondents, were nearly 3x as likely to have redesigned workflows. A correlation, not proof of cause.McKinsey, The state of AI in 2025: Agents, innovation, and transformation, November 2025.
Instrument 1: the screen
The Amplifier Test.
AI amplifies what is already there. Test what is there first.
The Amplifier Test is five gates for every proposed AI use case, before money is spent: value, process, data, people, and risk.
| Gate | The question |
|---|---|
| 1. Value | Is there a named P&L line this moves, with a baseline today? Is AI the cheapest way to move it? |
| 2. Process | Is the process documented, stable, and owned, with a known error rate? AI on a broken process scales the errors. |
| 3. Data | Is the data available, good enough, and ours to use? |
| 4. People | Will the people who run the process adopt the change, and who owns the outcome? |
| 5. Risk | Is the risk proportionate to the value, and can we switch it off if it goes wrong? |
Each gate is green, amber, or red. Any red gate: do not fund it yet.
In my experience most AI ideas fail at least one gate. I stop the ideas that are unlikely to pay, so the ones that pass get funded properly.
Four verdicts. Every use case leaves with one.
Fund
Every gate is green, or amber with a named fix and an owner. Fund it, baseline it, and report the result in dollars.
Fix the process first
The process is the problem and the fix is known. Fix it, then test again before you automate.
Not yet
The value, the data, or the people are not ready. Set a date and test again.
Kill
No credible path to value: no P&L line, a cheaper fix without AI, or a risk out of proportion to the value. Stop the spend and move the money to an idea that passes.
Instrument 2: the number
Net AI Value.
What it should save, after what it costs to build, run, and change, and after a reserve for what could go wrong.
Net AI Value is the risk-adjusted net present value of an AI use case. It produces two outputs that are never added together.
Output 1: Hold-period value
PV(cash value captured) − PV(cost to capture) − risk reserve.
Output 2: Run-rate EBITDA at exit
The recurring cash benefit minus the recurring run cost, in the year a buyer will underwrite. A buyer pays a multiple of baselined, recurring EBITDA.
- Cash value captured
- Annual run-rate benefit × realization (the share converted to cash: hires avoided, headcount reduced, volume absorbed without adding staff; redeployed time is capacity and does not count) × adoption factor (ramp and partial adoption), discounted at a stated rate over the remaining hold.
- Cost to capture
- One-time (build, process change, training, severance) plus recurring (licenses, model usage, support, rework).
- Risk reserve
- Σ probability × impact for discrete loss events only: data or IP claim, security incident, regulatory action, vendor exit, error at scale. Adoption shortfall lives in the adoption factor, not here. No double counting.
I present it as Low, Base, and High, with the three assumptions that move the answer most and the break-even adoption. Pre-tax. A management estimate, not GAAP, not Adjusted EBITDA, and not a quality-of-earnings opinion.
Worked example: one process, three cases.
Illustrative example. Hypothetical numbers and assumed inputs. Not a client and not a result. Your results will differ.
A $150M-revenue distribution business. 60 people in order-to-cash and customer service at $70K loaded cost ($4.2M). AI could free 25% of the effort ($1.05M a year of capacity). Assume 60% converts to cash through hires avoided as volume grows: $630K a year at full adoption. Year one runs at half speed while the change lands. Three years left in the hold. Discount rate 12% (20% shown as sensitivity). One-time build and change $400K; run cost $120K a year.
| Case | Assumed adoption | Extra cost | Risk reserve | Hold-period value at 12% (20%) | Run-rate EBITDA at exit |
|---|---|---|---|---|---|
| Sound process | 80% | none | $80K (data/IP $20K, security $30K, vendor exit $30K) | +$217K (+$119K) | +$384K a year |
| Broken process, automated as is | 35% | $100K a year rework | $170K (adds error at scale, 15% × $600K = $90K) | −$667K (−$661K) | about $0 |
| Fix the process first, then automate | 80% | $150K process fix; 6-month delay | $80K | −$45K (−$136K) | +$384K a year |
What it shows: automating the broken process loses about $670K and leaves nothing a buyer will pay for. Fixing the process first roughly breaks even in cash inside this hold and leaves $384K a year of recurring EBITDA that a buyer can underwrite at exit. Break-even adoption for the sound process is about 62% at 12%.
FAQ
What Net AI Value is not.
- How precise is Net AI Value?
- Net AI Value is an estimate you can argue with. Every line is an assumption with an owner, and I show you which three assumptions move the answer most.
- Is Net AI Value a valuation?
- No. It is a planning estimate. It does not value a company, and it does not set or recommend a purchase price.
- Is it a forecast of returns?
- No. It is a range with its assumptions written down. Actual results can differ materially.
- Is it the same as Adjusted EBITDA?
- No. It is a pre-tax management estimate. It is not GAAP, not Adjusted EBITDA, and not a quality-of-earnings opinion.
- Who can rely on it?
- Only the client named in the engagement letter. Nobody should use it as the sole basis for an investment, pricing, or financing decision. Items that touch legal or regulatory matters go to your own counsel.
The method underneath
Run on the Seven Pillars.
The Amplifier Test and Net AI Value tell you whether to fund an AI use case. The Seven Pillars of AI Leadership, from Why AI Fails, tell you whether the organization has the discipline to run it.
Investor? See where these fit in the deal lifecycle.
Put a number on your AI plan.
One conversation. I will tell you which gates your top AI ideas would clear.