--- title: "Measure an AI function without promising hours saved" description: "Measure an AI function through accepted outputs, review effort, exceptions, and cost per result, using a baseline and clear rules for counting each run." canonical: "https://scalewithsearch.com/articles/measure-ai-function-business-results" --- ## Site navigation - [Scale With Search](https://scalewithsearch.com/) - Services - Services - [AI systems](https://scalewithsearch.com/ai-systems) - [Site build](https://scalewithsearch.com/work#site-build) - [Whole website fix](https://scalewithsearch.com/work#whole-website-fix) - [Content library with SEO](https://scalewithsearch.com/work#content-library) - [Working session](https://scalewithsearch.com/audit) - Start here - [Send your brief](https://scalewithsearch.com/work#send-your-brief) - [Example scope](https://scalewithsearch.com/work#review-the-example-scope) - [Prices and terms](https://scalewithsearch.com/work#prices-and-terms) - Industries - Real estate - [Real estate agents and teams](https://scalewithsearch.com/for/real-estate) - [Platforms, ownership, and IDX](https://scalewithsearch.com/for/real-estate/real-estate-website-platforms-ownership-idx) - [Search, content, and AI](https://scalewithsearch.com/for/real-estate/real-estate-search-content-ai) - [Inquiries, campaigns, and handoffs](https://scalewithsearch.com/for/real-estate/real-estate-website-inquiries-campaigns-handoffs) - [Buyer, seller, and local pages](https://scalewithsearch.com/for/real-estate/real-estate-buyer-seller-local-resources) - [Rules, privacy, and trust](https://scalewithsearch.com/for/real-estate/real-estate-website-rules-privacy-trust) - Browse by business - [Trades and home services](https://scalewithsearch.com/for/#group-trades) - [Health, wellness, and animal care](https://scalewithsearch.com/for/#group-practices) - [Law, finance, and IT](https://scalewithsearch.com/for/#group-professional) - [Marketing and SEO agencies](https://scalewithsearch.com/for/#group-agencies) - [Shops and manufacturers](https://scalewithsearch.com/for/#group-manufacturing) - [Restaurants, shops, studios, and nonprofits](https://scalewithsearch.com/for/#group-local-services) - Chiropractic practices - [Chiropractic and wellness practices](https://scalewithsearch.com/for/chiropractic-and-wellness-practices) - [Chiropractic and physical therapy clinics](https://scalewithsearch.com/for/chiropractic-and-physical-therapy-clinics) - [All industries](https://scalewithsearch.com/for/) - Resources - Free tools - [AI readiness check](https://scalewithsearch.com/check) - [Free tools](https://scalewithsearch.com/tools/) - [Free Starter Kit](https://scalewithsearch.com/kit/business-memory-starter-kit.zip) - Guides - [AI-native library](https://scalewithsearch.com/ai-native) - [Industry intelligence](https://scalewithsearch.com/intelligence) - [Office job guides](https://scalewithsearch.com/guides/) - [Business memory](https://scalewithsearch.com/articles/business-memory-for-ai-agents-guide) - [Agent governance](https://scalewithsearch.com/articles/ai-agent-governance-guide-small-business) - [Migration and ownership](https://scalewithsearch.com/articles/leaving-vendor-ai-memory-guide) - Articles - [All articles](https://scalewithsearch.com/articles/) - [What changed](https://scalewithsearch.com/changes/) - About - Evidence - [Proof](https://scalewithsearch.com/proof) - Scale With Search - [About](https://scalewithsearch.com/about) - [How it works](https://scalewithsearch.com/how-it-works) # How to measure an AI function without guessing hours saved? Measure whether the function produces an accepted result from the right sources, within its approval and spending rules. Record review effort, failures, and corrections alongside output counts. Compare those observations with a defined starting process. A model run is not automatically a business result. A prepared draft may be rejected. A completed report may still wait for approval. A blocked run may have followed the correct rule by refusing to use missing information. For an owner, a useful scorecard should answer a decision: keep the function, change it, pause it, or expand its scope. Choose measures that help make that decision. Avoid using a large activity count as the only evidence of useful work. ## Define the unit before counting the work Choose a unit that matches the function's written result. It might be an accepted weekly exceptions report or a customer reply draft accepted for sending. State the conditions that make that unit complete. Keep the unit narrow. If the function prepares drafts, measure draft preparation. Track transmission separately when another capability performs the send. Do not count an approved draft as a confirmed customer message. Write the counting rule beside the scorecard. An illustrative rule could count a report only after its required fields, source checks, and reviewer acceptance are recorded. A revision of the same report should not become another business result unless your rule explains why. Also define eligible work. If a queue contains records outside the function's scope, exclude them using a stated rule and retain the excluded count. The denominator should reflect the work the function was authorized to handle. ## Record a starting process you can compare Observe the existing process on representative examples. Keep the input, required result, review notes, and measured effort. Note unusual cases and missing sources rather than smoothing them into a clean average. Use the same acceptance criteria when testing the AI function. If the new process checks more fields or prepares a different artifact, record that scope change. A shorter task and a longer task should not appear as identical units. Label estimates clearly. If nobody measured the old review effort, say that the baseline is incomplete. Begin collecting evidence instead of filling the gap with a remembered number. Keep a record of what each observation covers. A trial on ordinary inquiries gives evidence about those inquiries. It does not establish performance for a different language, a new service, or an unusual customer dispute. ## Separate accepted outputs from attempts Record eligible inputs, attempted work, accepted outputs, rejected outputs, blocked runs, and failed runs. State whether each count describes a business item or a system run. One business item may require several attempts. Keep that relationship visible through a stable item identifier. A retry should add to the attempt count without silently adding to the number of customers served or reports completed. For a draft-preparation function, a proposed acceptance measure is accepted drafts divided by eligible drafts submitted for review. List unresolved reviews separately. Do not hide pending work by treating it as either accepted or rejected. Use the existing [AI agent reliability metrics guide](/articles/ai-agent-reliability-metrics-small-business) for operational measurement detail. The owner's scorecard should preserve those distinctions while presenting the decision the business needs to make. ## Measure the check at the end Record how long reviewers spend checking an output when that measurement is practical. Separate source verification, editing, exception handling, and approval administration if the distinction helps explain a problem. Use a consistent start and stop rule. For example, measure active review from opening the draft to recording its disposition. Keep interruptions or incomplete observations labeled so they do not appear as precise review measurements. A shorter review alone should not decide acceptance. Check whether the reviewer verified the required facts and account details. If the review skipped required checks, record that as incomplete evidence. Use 10/80/10 as a responsibility model. A person directs the job, the system prepares the middle, and a person checks the output. The percentages do not establish that review will take a fixed share of the old task's time. ## Count correct stops and recurring problems separately A missing source should trigger the function's stopping rule. Record the block, its cause, and the person responsible for resolving it. Do not call it a completed output merely because stopping was appropriate. Use the [no-work receipt guide](/articles/ai-agent-no-work-receipt) to define an empty queue separately from an unreadable source. At the same time, do not treat every block as the same type of failure. A correctly blocked request and an error that prevented a valid job from running need different follow-up. Keep both visible on the scorecard. Review recurring causes. Repeatedly missing exports may call for a source-process change. Repeated factual edits may call for a clearer rule or better acceptance examples. Assign each correction to the part of the process that produced the problem. NIST's AI RMF Core recommends documenting test methods, measuring performance in conditions similar to deployment, and monitoring behavior in production. Keep the scorecard tied to those tested conditions. [NIST AI RMF Core, Measure 2.1 through 2.4](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/), checked 2026.10.26. ## Relate spending to the accepted unit Record the cost of attempts as well as completed outputs. Include failed and repeated attempts in the cost boundary you define. Identify any shared service costs you cannot assign directly. A proposed measure is period operating cost divided by accepted results in that period. State which costs it includes. If there are no accepted results, report the cost and zero outputs separately. Do not display a zero cost per result or perform a division by zero. Keep setup costs and ongoing costs distinguishable. If you allocate setup costs across a period, explain the method and retain the original cost. Label missing or estimated charges instead of presenting them as final. FinOps Foundation distinguishes technical usage measures from business unit measures. Its guidance supports connecting technology spending to the work it serves. [FinOps Foundation: Unit Economics](https://www.finops.org/framework/capabilities/unit-economics/), checked 2026.10.26. ## Make the scorecard explain a change Put the reporting period, function version, model identity, source state, and acceptance rule beside each measurement. If one changes, mark the change before comparing periods. A model replacement can be tested against the accepted examples with the same sources and checks. Keep the resulting outputs and review notes. Approve the replacement for that function only after its acceptance requirements pass. Use a similar process for a changed prompt, source rule, or approval path. Change one bounded part when practical. When several parts must change together, record the combined scope and avoid attributing the result to one component. The [run receipt guide](/articles/receipt-is-part-of-the-output) explains the evidence behind a scorecard. An owner should be able to open a measurement and find the runs it counts, including rejected attempts and unresolved exceptions. ## Agree on decisions before expansion Write the conditions for keeping, revising, or pausing the function. Use checks suited to its risk and purpose. A source-verification failure may require immediate pause even if other measures look acceptable. For expansion, name the added task or input type and the new acceptance examples. Do not assume a function that prepares one report has permission to send messages or change accounts. Those actions need their own scope and approval rules. The [AI systems method](/ai-systems) defines the result and handoff checks before building a function. Bring a measurable output, its current process, and the limits you need enforced. Keep any savings claim separate until you have evidence that supports it. [Send your brief](/work#send-your-brief) with the recurring job and the result your team can accept or reject from a written checklist. ## Related: Agent Governance - [Five AI Agent Reliability Metrics a Small Team Can Verify](/articles/ai-agent-reliability-metrics-small-business) - [No Work or Broken Run? Make the AI Agent Prove Which One](/articles/ai-agent-no-work-receipt) - [The Receipt Is Part of the Output](/articles/receipt-is-part-of-the-output) ---- ```text .|########||. .|########||. .|########||. |##||. .||##|. |##||. .||##|. |##||. .||##|. |#|. .|#|. |#|. .|#|. |#|. .|#|. |#| |#| |#| |#| |#| |#| .#| |#. .#| |#. .#| |#. .#. .#| .#. .#| .#. .#| |#. .#| |#. .#| |#. .#| |# ...... #| |# ...... #| |# ...... #| .# ||#########| #| .# ||#########| #| .# ||#########| #| .#. |######||######|. .#. .#. |######||######|. .#. .#. |######||######|. .#. #. .##|###|##|#|######| .# #. .##|###|##|#|######| .# #. .##|###|##|#|######| .# || |##|#||||||||||||#||#| || || |##|#||||||||||||#||#| || || |##|#||||||||||||#||#| || # |#||||||||||||||||||||#| #. # |#||||||||||||||||||||#| #. # |#||||||||||||||||||||#| #. || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || # .#||||||||||||||||||||||||#| # # .#||||||||||||||||||||||||#| # # .#||||||||||||||||||||||||#| # || ....|||#||||||##|#|||#|#||##||||....|. || || ....|||#||||||##|#|||#|#||##||||....|. || || ....|||#||||||##|#|||#|#||##||||....|. || #. . ....|# ....#||| ||| .#||| ....#. .# #. . ....|# ....#||| ||| .#||| ....#. .# #. . ....|# ....#||| ||| .#||| ....#. .# # . ||||#| .#####|| . .#| .#||| ||||# #. # . ||||#| .#####|| . .#| .#||| ||||# #. # . ||||#| .#####|| . .#| .#||| ||||# #. .| ||||| #. |#|||||. || #. |#||. .||||| || .| ||||| #. |#|||||. || #. |#||. .||||| || .| ||||| #. |#|||||. || #. |#||. .||||| || || |.... #. ....|#. |. ...|. ....|| || || |.... #. ....|#. |. ...|. ....|| || || |.... #. ....|#. |. ...|. ....|| || #. .||||||##||||||##||####|||||||#||||||#| .# #. .||||||##||||||##||####|||||||#||||||#| .# #. .||||||##||||||##||####|||||||#||||||#| .# # .||####|###########################|. # # .||####|###########################|. # # .||####|###########################|. # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#||||||.#.|| # |. ..# #| #|||||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||||| .. |# ## |#| ...#|#|||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#|||#|# .# .#| #| ##|.#..#|||||#| # # .#||||||############|######|||||#| # # .#||||||############|######|||||#| # # .#||||||############|######|||||#| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # # |...||#||||||#||||#|#||||||#|||||#||| |.| # #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# #. .. ||||# .|||##|. |#| .|| || .|||#. #| # .# || | ...|| ...##| | #| .|. |. ##### .. .| || || | ...|| ...##| | #| .|. |. ##### .. .| || || | ...|| ...##| | #| .|. |. ##### .. .| || || ||||| || ||||#| . |. |. |#. ||||| |#| || || || ||||| || ||||#| . |. |. |#. ||||| |#| || || || ||||| || ||||#| . |. |. |#. ||||| |#| || || .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. .# |.....#|....|#.||||.|||##.|#|....||.#||.#. #. #.|||||||#############################| ||| .# #.|||||||#############################| ||| .# #.|||||||#############################| ||| .# || ##||#||#||||||#||#||#||#||##. || || ##||#||#||||||#||#||#||#||##. || || ##||#||#||||||#||#||#||#||##. || # .#|||||||||#||#|||||||||||#| # # .#|||||||||#||#|||||||||||#| # # .#|||||||||#||#|||||||||||#| # || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || || |#||||||||||||||||||||||#| || .# |#||||||||||||||||||||#| #. .# |#||||||||||||||||||||#| #. .# |#||||||||||||||||||||#| #. || |#||||||||||||||||||#| || || |#||||||||||||||||||#| || || |#||||||||||||||||||#| || #. .######|#|#########| .# #. .######|#|#########| .# #. .######|#|#########| .# .#. |#####||||#####| .#. .#. |#####||||#####| .#. .#. |#####||||#####| .#. |# ||########|| #. |# ||########|| #. |# ||########|| #. |# ...... #| |# ...... #| |# ...... #| |#. .#| |#. .#| |#. .#| |#. .#. |#. .#. |#. .#. .#| |#. .#| |#. .#| |#. |#| |#| |#| |#| |#| |#| .|#|. .|#|. .|#|. .|#|. .|#|. .|#|. |##||. .||##| |##||. .||##| |##||. .||##| .||########||. .||########||. .||########||. Scale With Search 2026 [scalewithsearch.com](https://scalewithsearch.com) ```