How to measure AI business value across the whole workflow

Measuring AI business value requires following work through to an accepted result. Time spent reviewing and correcting output belongs in the calculation, alongside quality and total cost. A faster drafting step can coexist with a growing queue of unfinished work.

McKinsey's Technology Trends Outlook 2026 notes that AI can generate output faster than organizations can validate and put it into use. That observation raises a practical question for a workflow already using AI: where does the time saved actually go? The example and measurement approach below are a proposal for answering that question, rather than findings from the report.

Follow a proposal through to approval

Consider an entirely illustrative sales proposal. Preparing the first draft takes 40 minutes, followed by 20 minutes of checking and revision. Total hands-on work is 60 minutes.

Now suppose AI produces the draft in 10 minutes. Checking prices, adding missing conditions, and correcting the text takes another 45 minutes. Total work is 55 minutes. The drafting stage saved 30 minutes, but the overall saving is five. Model usage charges and system maintenance have yet to enter the calculation.

Hands-on time also differs from elapsed time. If the proposal waits until tomorrow for approval, faster drafting may have little effect on when the customer receives it.

None of this establishes that the AI system should be abandoned. It identifies where further investigation could improve the result. The next change might concern the information the system uses, the division of work, or the approval process.

Define what counts as completed work

For this example, a useful unit is a checked proposal that the responsible person has accepted as ready to send. That definition includes the quality requirements the business needs.

Compare tasks of similar difficulty. Giving AI the straightforward requests while leaving complex cases in the old process makes the comparison misleading. Failed attempts and work handed back to a person should remain in the record.

A small log can capture when a request arrives, when the first draft is available, time spent reviewing it, reasons for revision, and final approval. Add usage cost per accepted proposal. Keep setup and maintenance effort visible as well, while distinguishing initial investment from recurring expenses.

The aim is to connect a claimed saving to work the organization can use. A high volume of generated drafts says little about that on its own.

Use the evidence to choose the next change

Suppose reviewers spend most of their time checking whether prices are current. Investigate the source of pricing information. Who owns it? Which version is valid? Can the system identify when the information was last updated? Requiring a person to resolve the same ambiguity for every proposal creates recurring work.

If corrections concentrate on commitments made to customers, define which terms the system may use and when it must request approval. Permission to prepare a proposal does not automatically include permission to change prices or send it.

These decisions make trust more concrete. A reviewer who can trace a claim to its source has a clearer basis for accepting or correcting it. Speed obtained by removing necessary access controls would fail the definition of an acceptable result.

Bring one workflow to the next review

Choose a workflow in which AI is already being used. Name its owner and the person responsible for accepting the output. Set a bounded observation period covering comparable tasks. Two weeks might be a practical starting point, but it is an example of a review window, not evidence that every workflow can be evaluated that quickly.

Put completed work, elapsed time, review effort, and cost beside the usage figures. Explain how quality was assessed. If volume is low or outcomes vary widely, collect more observations before deciding to expand.

The next investment decision can then address a specific question: which change would help this workflow produce an acceptable result with less total effort?

Source

McKinsey, *Technology Trends Outlook 2026*, September 15, 2026. The report provides the observation about production speed and organizational validation capacity. All timing figures are illustrative; the measurement approach is a practical proposal.