VALUE ATTRIBUTION
FIELD NOTE 08 / 09If €100 of AI gets you €1 million in revenue, run.
A spectacular AI return can be manufactured by putting a tiny model bill under revenue that an entire commercial system helped create. The ratio is impressive. The explanation is missing.
Revenue comes from a changed commercial system, not from a token invoice.
SHARE THIS FIELD NOTE
Pass the useful signal on.
A tiny denominator can make almost anything look transformative.
Spend €100 on model usage. Associate the output with a campaign, a list of accounts or a sales sequence. Point to €1 million in pipeline or revenue. Divide the large number by the small one. The slide now contains a return so spectacular that asking questions feels pessimistic.
Run anyway. Not because AI cannot create material revenue. It can. Run because the claim has confused the cost of one ingredient with the cost and contribution of the complete commercial system.
The model did not build the brand, create the product, earn access to the customer, maintain the CRM, hire the seller, fund the media or negotiate the deal. If all of that disappears from the denominator while the full revenue appears in the numerator, the calculation is not bold. It is structurally dishonest.
The smaller the AI bill looks, the more carefully you should inspect everything excluded from it.
Between an AI output and revenue sits an entire chain of human behaviour.
An account-prioritisation system does not create revenue when it produces a ranked list. A seller has to trust the recommendation, contact the account, create a relevant conversation and move a decision that would not have moved otherwise. A campaign system does not create revenue when it generates copy. The work must reach the right customer, change a response and survive the rest of the buying journey.
Every link can break. The data can be stale. The recommendation can arrive too late. The team can ignore it. The customer can act for a different reason. A deal can appear in the influenced column even though it was already likely to close.
A credible value case makes this mechanism visible. It explains what changed in the work, who acted differently and which customer behaviour moved as a result. Without that bridge, the revenue number is decoration.
Pipeline is not revenue. Influenced is not incremental.
AI claims often travel through progressively softer categories. A generated email touched an account, so the account became influenced pipeline. Influenced pipeline later became revenue. The complete contract value was then presented as AI value.
The commercial question is counterfactual: what happened because the system existed that would not otherwise have happened? Perhaps the team reached more qualified accounts, improved conversion, shortened the cycle or protected margin. Those are valuable changes. They also require a baseline, a comparison and enough time to observe the result.
Even incremental revenue is not automatically return. Delivery costs, media, discounts, sales effort, implementation and managed operation still matter. Revenue is a useful outcome. Contribution or margin is closer to the economic value the business retained.
- Separate activity produced from behaviour changed.
- Separate pipeline touched from pipeline created.
- Separate total revenue from incremental revenue.
- Separate incremental revenue from contribution retained.
- Separate correlation from a result the system can reasonably claim.
Make the value claim survive five questions.
An honest AI business case does not need laboratory perfection. It needs a traceable line from the original problem to the economic result, with assumptions marked as assumptions and evidence collected where the system actually operates.
Baseline → Mechanism → Adoption → Increment → Margin
- 01Baseline
Record the prior volume, quality, conversion, time and cost before introducing the system.
- 02Mechanism
State exactly which decision or action the AI system is expected to improve and why that should affect the outcome.
- 03Adoption
Measure whether the intended users receive, trust and act on the system's work in the real process.
- 04Increment
Estimate the change above what would probably have happened without the system, using a credible comparison.
- 05Margin
Subtract implementation, operation, human effort and other commercial costs from the value retained.
A credible return includes the complete cost of producing useful work, not only the final model call.
Frontier models are not cheap. Cheap demos are.↗Big AI returns are possible. That is why weak proof is unnecessary.
Start with a use case where the mechanism is short enough to observe. Measure accepted work, time saved, quality improved or a decision made sooner. Preserve the old baseline. Compare similar teams, accounts or periods where practical. Track whether people used the output and what happened next.
As the evidence strengthens, move from operational measures to customer and commercial outcomes. Say what is observed, what is estimated and what is still too early to know. A modest result with a credible chain is worth more than a spectacular multiple nobody can reproduce.
The purpose of measurement is not to make AI look small. It is to find the conditions under which it genuinely creates value, then scale those conditions. A serious revenue leader should be excited by that work. Anyone selling a miracle ratio without the mechanism should make them leave the room.
Do not buy the multiple. Inspect the mechanism.
SOURCES & FURTHER READING
Follow the thinking.
Practical guidance on preserving baselines, measurement methods and evidence that supports the value claim being made.
↗02Harvard Data Science Review: Toward Causal Field Evaluations of AI SystemsA framework for evaluating whether an AI system changes meaningful outcomes for real users and tasks in the field.
↗03NBER: Generative AI at WorkA large field study that connects a specific AI deployment to measured productivity effects and shows how impact varies across workers.
↗