How to tell if AI spend is producing business results

Name the result

Choose one workflow and one result worth paying for. For quotes, measure accepted, accurate quotes per staff hour and the time customers wait. For support, measure resolved requests and reopened cases, not replies generated. Record the starting value, the data source, the quality standard and the review date before extending the spend.

Name the owner

Name the person who can verify the result and change the work. They need the operating records as well as the subscription invoice. Write down who checks output quality, who receives the released capacity and who can approve the next period of spending. Owning the login is not the same responsibility.

Name the workflow change

Describe the old and new workflow from input to accepted output. Include preparation, prompting, checking, corrections and handoff. Compare similar tasks and workloads; where practical, keep a comparable group using the old process during the same period. Record staffing, demand or process changes that might explain the difference. Call the result an observation when you cannot separate those effects.

Name the stop rule

Before the next renewal, agree the minimum useful improvement, required quality and maximum additional spend for this workflow. These are business-specific choices, not a universal AI ROI threshold. Record three possible decisions: continue when the result is useful and costs and quality are acceptable; change a named part of the workflow and retest when the evidence identifies a fix; pause expansion or stop the use when it fails the agreed conditions. Missing evidence is a reason to limit the next test, not to report success.

Compare the same work before and after.

Choose a task and a period you can compare fairly. Record completed items, total working time and errors or rework. Include checking, corrections and handoff time in both versions. Keep the required quality standard the same. If the tasks became easier, that change cannot be credited to AI.

Illustrative calculation: 50 comparable quotes used to take 20 minutes each, including checking. With AI, drafting takes eight minutes and checking takes six: 14 minutes total. The difference is six minutes per quote, or 300 minutes across 50 quotes. That is five hours of capacity released if the quotes still meet the same standard.

If those hours are valued internally at $30 each, the capacity value is $150. An $80 incremental monthly tool cost and $60 of setup time leave $10 against that first-month cost comparison. All figures are hypothetical. Salaried time released is not automatically cash saved, additional sales or profit; say what useful work actually filled it.

Before extending the test, inspect rejected quotes, corrections and missed commitments. Stop or change the workflow if quality falls below the agreed standard, review work erases the time saving, or the released time has no useful application. Record the decision date and repeat the comparison on a suitable later batch. A before-and-after observation alone cannot isolate every other change in the business.

Count the costs that the next decision adds.

For the same period and workload, list additional licenses or API usage, setup and integration, training, review and rework, maintenance, and any extra service or support. Keep one-time setup separate from recurring costs. Allocate a shared subscription consistently to the workflows using it rather than charging the whole bill to each one. Include staff time once: if checking is already in total task time, do not subtract it a second time as a separate labor cost.

The quote example above leaves $10 in its hypothetical first-month capacity-value comparison. Without repeating the $60 setup charge, the same assumptions would leave $70 in a later month before any new maintenance, training or other costs. Neither number is cash profit. For a renewal, compare the benefits and costs still ahead; keep past setup spending visible in the overall account without assuming another payment will recover it.

Turn “five hours saved” into a checkable use: did a backlog shrink, did quotes reach customers sooner, or did overtime actually fall? Record the result separately. If the same time is used to generate extra sales, do not count both the full hourly capacity value and the resulting gain as independent benefits of those same hours.

What if nobody measured the work before AI?

Start with comparable dated records: completed jobs, elapsed handling time, corrections and invoices. Mark gaps instead of replacing them with a confident estimate. If records cannot support a fair comparison, run a limited test with a comparable old-process batch where practical and appropriate. Keep the same acceptance standard. Until that evidence exists, report usage, cost and observed quality separately; the business return is still unknown.

At the review, the owner should be able to say: “We compared this work over this period, using these records. This result changed, these costs were added, these other conditions changed, and this is the next decision.” A renewal decision needs that explanation more than a single impressive percentage.

What the sources support.

Microsoft’s measurement guidance separates usage from business value and supports baselines, comparison groups where possible, quality measures and tracking where released time goes. Its account of its own AI investments explains scenario-specific measures and the limits of claiming ROI without a defensible cost model. These are measurement references, not evidence of a return for your company. The quote arithmetic and decision examples on this page are illustrative business applications.