Skip to main content

Automation and Agents

How to Measure AI Business Value Beyond a Vendor ROI Example

Build a business-value claim that distinguishes accepted work, released capacity and actual spending changes.

An open record book separates valued work time from expense and support records beneath the OpenAI publisher mark.
On this page
  1. 1. Choose the benefit before choosing the formula
  2. 2. Find a baseline that resembles the work you want to improve
  3. 3. Measure the whole route to accepted work
  4. 4. Explain what the released capacity made possible
  5. 5. Use the contract and invoice for the cost side
  6. 6. Work one small scenario without inflating the result
  7. 7. Show variation and uncertainty before annualizing
  8. 8. Make the next investment decision explicit

Saving an hour can be useful without reducing a bill by an hour's wages. It can give a team more room to serve customers, improve a deliverable or finish earlier. Before spending more on an AI workflow, name the benefit the team received and the cost of achieving it.

OpenAI's September 16, 2026 business-value guide includes a hypothetical 245% annual return based on estimated capacity value. It labels its figures as hypothetical and includes assumptions about productive use and first-year costs. Treat that example as a way to inspect assumptions, not as a measured return to copy into your budget.

The stronger business case follows a particular workflow from its starting point to an accepted result. Compare similar work, include review and failed attempts, distinguish available capacity from realized spending changes, and keep your billing basis clear. The arithmetic below is a proposed method with hypothetical inputs, not a customer trial or a standard prescribed by NIST.

Key takeaways

  • Declare whether you are measuring capacity, actual cost reduction or extra business output.
  • Compare accepted work and total effort, not just the speed of the first draft.
  • Count the same labor and benefits once, using a consistent period.
  • Keep assumptions visible so the next investment decision can change when the evidence changes.

1. Choose the benefit before choosing the formula

Start with a result the team wants to improve. “AI ROI” is too broad to tell you which records to collect or what a favorable result would mean. Choose the specific improvement first.

A team may want to prepare proposals with less effort, respond to a backlog, deliver more client work, reduce purchased services, or improve the quality of decisions. Those are worthwhile goals, but they do not all become the same kind of financial benefit. Decide which one the investment is meant to produce.

To assess capacity, ask how much work the team can handle and how much effort each task needs. For cash savings, the question is which expense actually fell or which planned expense no longer needs to be incurred. For extra output, the question is whether the team used the extra capacity to deliver something the business wanted and could use.

Do not treat an hourly salary valuation as proof that payroll changed. It can help describe the scale of released capacity. It cannot establish that the business spent less, hired fewer people, or won another customer. Those claims need their own records.

Valued capacity remains separate from spending and customer records. View image detail

Choose Actual size to read the graphic closely.

You can also choose a nonfinancial benefit explicitly. A clearer account brief or a less frustrating preparation task may matter even if there is no immediate revenue change. Record the result in its own terms. Forcing it into a dollar estimate can make the case less understandable rather than more rigorous.

Write a one-sentence claim before calculating anything: this workflow is intended to help this team produce this accepted result with less total effort. Then name the business consequence you hope follows. Keeping those sentences separate helps the reader see where observed improvement ends and a future expectation begins.

It also makes an unsuccessful trial useful. If effort falls but no additional work is needed, you may still have a capacity benefit. If quality improves at the same effort, that is a different result. The outcome should be described faithfully before someone decides whether its value justifies the investment.

2. Find a baseline that resembles the work you want to improve

Use a baseline that answers the same question as your trial. A fast new task compared with a difficult old task does not tell you how much AI helped. A comparison becomes more useful when its workload and completion criteria are reasonably alike.

NIST's December 2, 2025 measurement-science discussion raises questions about validity, baselines, uncertainty and whether evaluations generalize to other settings. It does not prescribe a corporate ROI formula or a mandatory trial length.

For a hypothetical proposal workflow, a baseline could use previous proposals with a similar scope and input quality. Record what had to be gathered, who prepared the document, what the reviewer checked and what counted as ready to send. If those records do not exist, label a reconstructed estimate as an estimate rather than presenting it as logged work.

Check the surrounding process. A new template, a simpler approval path or more complete client inputs can change results alongside AI assistance. You do not need to ignore those improvements. You do need to avoid assigning all of their benefit to the tool.

A task is supported by changes to its template, inputs, approval path and tool assistance. View image detail

Choose Actual size to read the graphic closely.

Choose a comparison that fits the decision and the work available. Comparable before-and-after tasks can support a modest operational decision while leaving uncertainty about cause. If the business needs stronger evidence that AI caused the change, it may need a more controlled comparison. Do not assert that dividing staff into two groups automatically removes seasonality, learning effects or workload differences.

Keep a note of material changes during the comparison. A quiet month, a changed customer mix or help from a senior reviewer can affect what the numbers mean. The note is useful context for interpreting the result, not an excuse to disregard every inconvenient finding.

If the starting records are incomplete, keep the claim modest. “These comparable proposals needed less recorded preparation effort during the trial” can be defensible. “AI increased departmental productivity by this percentage” reaches much further. Choose the statement your records can support and decide whether that is enough for the next bounded investment.

3. Measure the whole route to accepted work

The result that matters is an accepted deliverable, not the first generated version. Measure the effort needed to get there, including checks, corrections and attempts that did not lead to a usable result.

Define acceptance with the people responsible for the work. For a proposal, it may mean that the scope, price and client requirements are correct and the owner is ready to send it. For an engineering change, the team should choose its normal behavior and release checks. The evaluation should not invent a lighter standard simply because the initial draft was quick.

Separate hands-on effort from elapsed time. A document may take less writing time while waiting longer for approval. That could mean the team freed capacity without changing delivery speed. Conversely, a faster response might come from an approval change rather than less labor. Recording both gives you a clearer account of what improved.

Include the work done before generation. Gathering approved inputs, resolving missing information and setting up the task can be significant parts of the workflow. If the evaluation clock starts after those steps, compare it with a baseline that starts at the same point. Otherwise the apparent saving can be a difference in measurement boundaries.

Two workflow comparisons use matching start boundaries and retain the preparation before generation. View image detail

Choose Actual size to read the graphic closely.

Record review and correction once. If your AI-assisted total already includes them, subtracting their wage value again from the resulting labor saving would count the same cost twice. You can show a breakdown for clarity while still using one complete effort total in the calculation.

Count attempts that fail too. If several drafts are abandoned and completed manually, include that fallback work. An average calculated only from successful assisted tasks can hide the cost of learning which tasks do not fit. A useful report can show successful and unsuccessful attempts separately, then explain which future workload it expects.

Ask whether quality remained acceptable. Less effort accompanied by missing facts may not be an improvement. Better quality at the same effort can be worth continuing. Do not require every positive result to appear as a speed increase, and do not let an attractive speed result erase the standard the business needs.

4. Explain what the released capacity made possible

Time saved creates an opportunity. What happens next determines which business claim you can make. Name the next use rather than assuming every freed hour became revenue.

A team might use capacity to clear a backlog, strengthen customer follow-up, improve an existing deliverable, reduce overtime, or make a demanding week more manageable. Each consequence has different evidence. A workload record can show backlog movement. An expense record can show lower overtime. A sales record can show a later commercial outcome, though attributing that outcome to one workflow may remain difficult.

For a hypothetical agency, proposals prepared with less effort might let the same team serve more prospective clients. That only becomes extra business output if enough suitable requests exist and the team actually handles them. It does not follow merely from multiplying possible hours by a billing rate.

Check what still limits the team's output. If a senior partner still has to approve every proposal and already has a full queue, faster drafting may not create more completed proposals. If there is no additional demand, extra production capacity may remain unused. Those facts affect the expansion case without making the drafting improvement imaginary.

Drafting capacity remains separate from an approval queue and an unfilled demand slot. View image detail

Choose Actual size to read the graphic closely.

Avoid double counting the consequence. If you value released time at a labor rate and also count the margin from work done in that same time, explain which measure belongs in the financial total. Showing capacity and additional margin side by side can be informative. Adding both as independent benefits can overstate the case.

The same care applies to reduced spending. If faster preparation lets the business stop buying a service, use the documented expense change. Do not also claim the full wage value of the same displaced work unless the benefits are genuinely separate and the finance owner agrees with the accounting treatment.

Some benefits need time to become observable. An improved brief may support a better customer conversation long before a sale closes. Describe the intermediate result now and leave the later outcome open. A decision maker can fund a further test based on a credible promising signal without pretending the final commercial result has already arrived.

5. Use the contract and invoice for the cost side

Consumption measures and financial charges answer different questions. Use the business's actual agreement and billing records when estimating investment. A dashboard trend alone cannot tell you what the invoice did.

OpenAI's current ChatGPT Work usage-and-cost guide distinguishes shared committed credits from additional invoiced charges. Applicable agreements, rates, remaining allocation and overage settings matter. It also separates seat fees from estimates of credit consumption.

That means a marginal workflow decision can differ from a total program evaluation. If you are deciding whether to use already committed capacity for one more suitable task, consider the available allocation and competing demands. If you are deciding whether to renew or expand the broader program, consider its committed costs as well. State which decision the cost estimate supports.

Use a consistent period. Monthly workflow benefits should be compared with costs for that month or an explicitly explained allocation. One-time setup should not vanish from the first-year case, and a setup expense should not be charged again in every month unless that is how the chosen model accounts for it.

Monthly benefits and costs share one period, while setup has a separate allocation reference. View image detail

Choose Actual size to read the graphic closely.

Include actual support work. Someone may maintain shared skills, improve source material, train users or investigate failed results. Count that effort when it is part of the investment, but do not invent a universal support burden for every deployment. The team can distinguish setup from recurring maintenance in its own records.

Separate total cost from a marginal cost estimate. A share of credits can help compare workflows even when no new invoice charge follows. Keep that estimate labeled. Do not present a calculated share of committed consumption as a new cash payment.

For mixed tools or services, describe the boundary of the evaluation. If a task depends on another system or a paid data source, say whether its cost is included. You do not need to assign the entire company's technology budget to one proposal, but you should avoid making a workflow look cheaper by quietly excluding a resource it requires.

6. Work one small scenario without inflating the result

A small calculation is easier to inspect than an impressive annual forecast. Begin with a period and a set of comparable deliverables, then show exactly how assumptions enter the result.

Here is a hypothetical monthly proposal workflow. These figures are invented teaching inputs, not observed results or current OpenAI prices. Assume the team completes 40 comparable proposals. The baseline requires 90 minutes of hands-on effort per accepted proposal. The assisted workflow requires 60 minutes, including preparation, review and corrections. Assume quality meets the same agreed criteria.

The difference is 30 minutes per proposal, or 20 hours across the month. If the business chooses an assumed loaded labor value of $50 per hour, that represents $1,000 of estimated capacity value. It does not establish a $1,000 reduction in payroll.

Now assume $300 of relevant tool cost and $200 of additional maintenance and enablement effort for the same month. Under this simplified capacity-value model, the estimated benefit less those costs is $500. Expressed relative to the $500 investment, the modeled net return is 100%. The interpretation is more important than the headline number: this is a scenario about capacity valued at an assumed rate.

A hypothetical capacity-value model keeps the assumed benefit, tool cost and support cost distinct. View image detail

Choose Actual size to read the graphic closely.

Review labor has already been included in the 60-minute assisted total. Subtracting another review-cost line from the $1,000 difference would count it twice. The $200 support line should contain work outside that per-proposal total. If the team included the same maintenance time in both places, the scenario would need correction.

Change one assumption and the decision changes. If comparable accepted proposals instead take 80 minutes with assistance, the same 40 proposals release about 6.7 hours. At the assumed $50 rate, estimated capacity value is about $333, below the $500 cost. That does not prove the product is unhelpful. It means this narrow capacity-value scenario does not yet support the same expansion claim.

Neither result is a cash-savings calculation. To make that claim, replace the assumed labor valuation with an actual spending change and check that it is attributable to the workflow. To make an additional-output claim, show what the team delivered with the released capacity and how the business valued it. Do not count all three views as independent gains from the same hours.

7. Show variation and uncertainty before annualizing

An average can hide a useful workflow for one group and a poor fit for another. Examine differences that affect the investment decision before extending a small result across the business or the year.

The 2025 Generative AI at Work paper studied a specific customer-support deployment and reported varied effects across experience and skill groups. Its findings do not establish the return of ChatGPT Work or your local rollout. They illustrate why the relevant population and task matter.

For a local evaluation, record whether difficult assignments, new staff or incomplete inputs produced different results. Do not turn every difference into a permanent worker category. Ask whether the workflow can be improved, whether training needs differ, or whether some work should remain outside the proposed expansion.

Show the amount of evidence behind the summary. A handful of accepted documents can justify another bounded test. It cannot establish the full range of exceptions a larger team will encounter. Keep the number of comparable tasks, missing records and excluded cases visible to the decision maker.

A small evidence set remains beside missing-record slots and excluded cases. View image detail

Choose Actual size to read the graphic closely.

Show how the answer changes when an assumption changes. Change the effort saving, demand, support burden or period of productive use and show what happens to the case. Those are planning choices, not error bars from a statistical analysis that was never performed. Label them accordingly.

Be careful when annualizing. A recurring task may not occur at the same volume throughout the year. Initial setup may be concentrated at the start. A benefit that depends on available customer work should not be projected as if demand were certain. Explain what would have to remain true for the annual estimate to hold.

A credible report can state that a workflow looks promising but has an uncertain effect on finances. That gives leadership a decision it can act on: continue gathering relevant evidence, narrow the scope, or fund a measured expansion. Correct arithmetic cannot rescue hidden assumptions. Show the inputs that would change the decision.

8. Make the next investment decision explicit

The purpose of measurement is to decide what to do next. End with the action the evidence supports, the work it applies to and the condition that would change the decision.

An expansion decision might apply only to similar proposals prepared by a team with complete inputs and an available reviewer. An improvement decision might focus on reducing corrections. A pause might be sensible when effort stays high or the outputs repeatedly miss the agreed criteria. None of those decisions requires a claim about every AI task in the business.

Give the business owner a short evidence statement. For the hypothetical proposal workflow, it would identify the 40 comparable proposals, the total effort per accepted proposal and the $500 cost assumption. It would call the result estimated capacity value, then say whether the team used that capacity. State whether the result is observed, estimated or hypothetical. Include the most consequential limit next to the conclusion, where it can affect the choice.

Name what the next review should resolve. If the workflow released capacity, ask whether the team used it for the intended work. If cost was estimated from consumption, check the actual billing effect. If quality improved, decide whether that improvement matters enough to justify the ongoing support. This keeps the next measurement tied to a decision rather than collecting numbers indefinitely.

The next review connects capacity to intended work, a cost estimate to billing, and quality to ongoing support. View image detail

Choose Actual size to read the graphic closely.

Adoption diagnosis is a separate but complementary job. Our guide to using OpenAI Admin Analytics to diagnose team adoption addresses how to choose support for a workflow. A team can need better support before a financial comparison is meaningful, and a high-use workflow can still have a weak investment case.

Do not impose a universal deadline for success or automatic removal. A suitable review period depends on the cadence of the work, the investment and the evidence required. The responsible owners should make that decision within their actual agreements and operating rules.

Choose one workflow and write the benefit claim in plain words before opening a spreadsheet. If you can explain how useful work became a measurable result and which costs belong beside it, the calculation can support a real choice. If that connection is missing, another percentage will not supply it.

Checked for this article

Sources

  1. How to connect AI usage to business valueOpenAI
  2. ChatGPT Work usage and costOpenAI
  3. Accelerating AI Innovation Through Measurement ScienceNIST
  4. Generative AI at WorkQuarterly Journal of Economics

Keep reading

All articles