Skip to content

Every number sayshow it was established.

AI reporting often stops at activity. A finance team needs to know what changed, what it cost, and whether the money is real. These are the distinctions we use, with arithmetic you can check.

  1. Modeled

    Calculated from stated assumptions. Useful for choosing what to test; not evidence that anything happened.

  2. Measured

    Observed in your systems, against a baseline, for a stated period.

  3. Verified

    Measured, then reconciled with your financial records and confirmed by the person who owns the budget. Verified by the customer is not the same as independently audited.

  1. AI spend vs. AI usage

    Spend is what you pay for: seats, tokens, platform fees and committed contracts. Usage is what people actually open.

    How it’s calculated

    Seat utilization = active seats ÷ purchased seats

    Worked example, fictional figures

    150 seats at $40 a month cost $6,000. If 105 were active this month, utilization is 105 ÷ 150 = 70%, and the cost per active seat is $6,000 ÷ 105 ≈ $57.

    What it doesn’t tell you

    Active means opened at least once in the period. It says nothing about value.

  2. Usage vs. productive utilization

    Usage counts attempts. Productive utilization counts the attempts that produced accepted work.

    How it’s calculated

    Acceptance rate = AI-assisted work accepted ÷ AI-assisted work attempted

    Worked example, fictional figures

    Of 1,000 AI-assisted changes started in a month, 620 were accepted after review: 62%. The other 380 consumed time and review without producing accepted work.

    What it doesn’t tell you

    Acceptance must be defined before you measure — and an accepted change can still be the wrong change.

  3. Productivity vs. operational output

    Productivity is output per unit of effort on one task. Operational output is what the whole system delivers.

    How it’s calculated

    Modeled output ≤ the lowest stage capacity (a simplified serial-workflow model)

    Worked example, fictional figures

    If AI raises code-candidate capacity from 47 to 62 a month but review can clear 40, deployments stay at 40. Task productivity rose about 32%; delivered output rose 0%.

    What it doesn’t tell you

    Actual output also depends on demand, variability, work mix and quality. Relieving one constraint exposes the next, so re-measure the whole flow after every change.

  4. Released capacity vs. cost avoidance

    Hours AI saves are released capacity. They avoid a cost only when they replace spending you had planned.

    How it’s calculated

    Cost avoidance = planned cost not incurred because of released capacity

    Worked example, fictional figures

    Saving 20 engineers 4 hours a week releases 80 hours — about 2 full-time equivalents. If that lets you cancel a planned $300,000 contractor engagement, that is $300,000 of cost avoidance. If the hours go to backlog work, it is capacity: valuable, but not a saving.

    What it doesn’t tell you

    Cost avoidance keeps a future budget flat. It does not reduce today’s spend.

  5. Cost avoidance vs. realized savings

    A realized saving means a like-for-like expense actually falls, net of what the AI and the change cost. Cash-flow savings are reported separately, from payment timing and cash records.

    How it’s calculated

    Net expense reduction = baseline expense − actual expense − incremental AI and implementation cost, same scope and period

    Worked example, fictional figures

    A team spent $1.2M last year on external content production. This year it spends $1.0M on the same scope, plus $72,000 on AI tools. Net expense reduction: $1.2M − $1.0M − $72,000 = $128,000. Whether that is also $128,000 of cash depends on when the payments fall.

    What it doesn’t tell you

    Scope and period must match, and the reduction must be attributable to the change. If volume or quality changed, adjust before calling it a saving.

  6. Estimated vs. verified financial benefit

    An estimate is a model. A verified benefit has been measured against a baseline and reconciled with financial records.

    How it’s calculated

    Modeled → Measured → Verified

    Worked example, fictional figures

    “Review automation could save $400,000” is modeled. “Review wait fell from 4.6 to 3.0 days in the pilot group” is measured. “Contractor spend fell $300,000 against budget, confirmed by the finance owner” is verified.

    What it doesn’t tell you

    We show all three, and never let a modeled number pass for a verified one.

  7. Cost per successful outcome

    The included cost of getting one outcome accepted — not the cost of the tool alone. Every report names the cost categories it includes and omits.

    How it’s calculated

    Cost per successful outcome = (AI cost + people time + rework) ÷ outcomes that passed acceptance

    Worked example, fictional figures

    A month with $6,000 of AI tools, $54,000 of engineering time and $12,000 of rework that produces 40 accepted deployments costs ($6,000 + $54,000 + $12,000) ÷ 40 = $1,800 per successful deployment.

    What it doesn’t tell you

    Define “successful” before measuring, count engineering time and rework once each, and compare against the same measure without AI.

  8. Technical acceptance vs. business acceptance

    Technical acceptance means the work passed its tests and review. Business acceptance means the person who owns the outcome confirms it does what the business needed.

    How it’s calculated

    Delivered value counts only after business acceptance

    Worked example, fictional figures

    A claims-routing change can pass every test and still route the wrong cases. It counts as delivered value only once the claims owner accepts it on real cases.

    What it doesn’t tell you

    Business acceptance takes longer to collect. We report both, and say which one a number rests on.

What can invalidate a result

  • Attribution: other changes landed in the same period.
  • Baseline drift and seasonality: the comparison period was not like-for-like.
  • Adoption ramp: early results understate or overstate the steady state.
  • Quality: more output at lower quality is not an improvement.
  • Implementation and change costs left out of the comparison.
  • Double counting: the same hours or savings claimed twice.

Opportunity cost matters to the decision, but we never report it as cash. A matched comparison group or a phased rollout strengthens attribution; neither proves causality on its own.

What we won’t do

  • Report productivity as savings.
  • Present a modeled benefit as a verified one.
  • Hide the assumptions that could overturn a recommendation.