The argument
AI adoption is not gated by capability or by budget. It is gated by proof.
Companies buy AI they never put to work. Half of midmarket finance teams buy it; 12% run it in production. A production result from a peer is what moves them.
Smaller companies buy AI but rarely run it.
50% of finance teams at companies under 500 employees buy AI tools instead of building their own. Only 12% have any AI live in day-to-day work.
Large companies deploy because their boards push.
37% of large-company finance teams already run AI in daily work, and 77% say their board is pressing them to.
Doubts about the payoff fade once AI is live.
58% of large-company finance leaders without live AI say they cannot justify the cost. Among those already using it, that drops to 36%.
Buyers pick vendors on accuracy, not AI claims.
Finance leaders weigh nine factors when choosing a vendor. Accuracy comes first; the sophistication of the AI comes last.
Buyers pay more only after the AI has worked.
95% of finance chiefs will pay extra once the AI has proven itself, and 65% of Fortune 1000 buyers would pay at least 20% more.
Investors see the payoff; their companies do not.
95% of fund leaders say AI delivered what they expected, yet only 7% say it is built into the companies they own.
Where the studies converge
- Proof beats claims at selectionBuyers choose vendors on demonstrated results, not on how advanced the AI sounds.Measured directly in every study5 of 5
- The premium is paid after proofBuyers will pay more for AI, but only once it has proven itself in their own business.Indirect in AI security buyers5 of 5
- References at own size and domainBuyers trust proof from companies like their own: the same size and the same industry.Indirect in accounts payable automation and compliance software5 of 5
- Layer on the system of recordBuyers want AI added to the systems they already run, not a replacement for them.Not measured by AI services buyers; indirect in AI security buyers4 of 5
- Human stays on the exception pathBuyers want routine work automated and a person deciding the exceptions.Not measured by compliance software; indirect in AI services buyers4 of 5
- Per-seat pricing loses to fixed or outcomeBuyers prefer a fixed fee or paying for results over paying per user.Not measured by accounts payable automation; indirect in chief financial officers4 of 5
- Chief financial officers
- CFOs and senior finance leaders, $50M to $5B+ revenue
- Jan 2026
- Accounts payable automation
- Accounts payable automation software customers
- 2025
- AI services buyers
- US Fortune 1000 class buyers of AI and digital engineering services
- Jul 2026
- Compliance software
- Regulated mid-size buyers of compliance and quality software
- 2026
- AI security buyers
- Enterprise AI safety, security and trust-and-safety buyers
- Jun 2026
The gate
One gate, every size band. Each band needs a different production result to clear it.
| Measure | Under 500 employeesn=60 | 500 to 2,499 employeesn=39 | 2,500 or more employeesn=30 |
|---|---|---|---|
| Where each band stands | |||
| Position at the gate | Waits for a peer's result | Pilots, rarely finishes | Deploys first, then owes the number |
| Already in production17% across all 129 | 12% | 10% | 37% |
| Do no pilots37% across all 129 | 50% | 21% | 30% |
| Buy only, no internal AI36% across all 129 | 50% | 28% | 17% |
| Board pressure, moderate or strong57% across all 129 | 53% | 49% | 77% |
| What clears the gate | |||
| Proof before purchase |
|
|
|
| Proof after purchase |
|
|
|
A vendor can measure and publish every one of these today. None of them needs a better model.
Crossover CFO study, finance leaders, by employee count. Bold bars carry the argument; 2,500+ is directional (n=30).
What invisible AI costs a vendor
Proof wins the deal. Visible AI wins the renewal, at a higher price.
AI ranks last of nine criteria when buyers choose. After it works, 95% will pay more.
of the vendor's customers have never noticed AI in its product, even though the product ships it.
would still accept a higher price for AI they can see working.
customers who left the vendor moved to a competitor with stronger AI.
At 5.9, the vendor scores lower than every one of the 20 companies in our benchmark. Its own peers score 6.6, level with the weakest of them.
What changes in 2027
Four shifts, each driven by a number in the data.
How to think about the AI trade
Read every AI number as bought or activated.
| Rule | In practice | Evidence |
|---|---|---|
| 1Separate intelligence bought from intelligence activated | Ask which one an adoption number counts | 50%Midmarket buy only |
| 2Ask who pays for being wrong | Whoever bears the mistake decides go-live | 77%Enterprise board pressure |
| 3AI is last at selection and first at renewal | Model AI as a renewal uplift in the NRR bridge | 95%Pay a premium after proof |
| 4Benchmark activation, not adoption | Report production results by size band each quarter | 12%Midmarket in production |
| Ask at every portfolio review | ||
| Q1Proof density | What share of our AI customers are in production, by size band? | |
| Q2Price realization | What AI price did we realize at renewal, against list? | |
One vendor: 10% of customers had seen its AI work, 73% would pay for it. The gap is measurement, not engineering.
One gap, three altitudes
Same gap at every level: the evidence exists, but it is scattered.
| Who | Symptom | Cause | What clears it |
|---|---|---|---|
| The AI buyer | Buys AI, never runs it | Proof scattered across vendors | A graded production result |
| The portfolio | Return the companies cannot show | Learning split across report, CRM, deck | One number, entry to exit |
| Diligence itself | Answers IC once, then sits unused | Same scatter, one level down | Research and action in one system |
One production number, with base, source and grade, where the next decision is made.
The session
The Proof Gate
Open the presentationIntelligence bought is not intelligence activated. A graded production number activates it, on either side of the deal.