Use three layers of evidence
AI measurement becomes confused when generated output is treated as completed work, or completed work is treated as incremental revenue. Track all three separately. A draft is output. An approved quote delivered to the correct buyer is completed work. An additional profitable order is a business outcome.
Each layer answers a different question. Output metrics help understand activity. Completion metrics show whether the process works. Business metrics show whether the work matters. A useful dashboard connects them without pretending that one proves the next.
Count the tools, staff time, review and repairs needed to run the process.
Count accepted completed results separately from drafts and unresolved work.
Divide the agreed total cost by accepted results, keeping severe failures visible.
Read the visual explanation
A practical sequence for this decision. Apply your actual records, responsibilities and acceptance criteria.
Define the eligible population
Write the entry rule before the pilot. An enquiry might be eligible only when it concerns a supported product, arrives through an approved channel and has enough information to identify the buyer. Count ineligible and unresolved records separately.
Changing eligibility during the pilot can make completion rates look better without improving delivery. Keep versions of the rule and report results under the relevant version. Track duplicate events so retries do not count as additional successful jobs.
Calculate cost per accepted result
Divide operating cost by jobs meeting the acceptance criteria. Include model usage, infrastructure, software, human resolution, supervision and recovery work. Keep setup investment separate or amortize it over an explicitly stated period.
For illustration, suppose monthly operating cost is $3,000 and 600 jobs meet the acceptance criteria. Operating cost is $5 per accepted job. If 150 of those jobs later require reopening and do not meet the definition until corrected, include correction cost and avoid counting the same job twice. These are hypothetical numbers, not Looqus pricing or measured client results.
Treat time savings as capacity first
Saved minutes do not automatically reduce payroll or create revenue. Establish whether the released capacity was used for additional productive work, lower overtime, reduced external spend or faster service. Document the actual change rather than multiplying all saved minutes by an hourly rate and calling it cash.
Use a baseline measured with comparable work. Seasonality, demand mix and staffing changes can distort before-and-after comparisons. Sample the difficult cases as well as the easy ones; an average that excludes escalations understates the operating burden.
Separate contribution from booked value
An order value is not collected cash, and revenue is not contribution. Where relevant, account for discounts, returns, cost of goods, shipping subsidies and incremental acquisition cost. State which costs are included and which remain unknown.
For a controlled experiment, estimate incremental impact using the agreed comparison. For an observational rollout, report the limitations plainly. A before-and-after improvement can be useful evidence, but it does not isolate the workflow from simultaneous price, stock or marketing changes.
Agree the measurement contract
Define the baseline period, owner, data sources, exclusions, acceptance criteria and review cadence. If commercial terms depend on upside, agree how the counterfactual and disputed records are handled before launch. Avoid leaving attribution decisions until money is due.
Looqus positions the work around measurable outcomes and shared upside within an agreed engagement. Measurement should remain inspectable by the customer. A strong report includes accepted jobs, unresolved work, correction effort, observed business movement and uncertainty. That is more useful than a large count of agent actions.
Measurement contract worksheet
Use this starting worksheet to agree the process with the people who will operate it. Replace the examples with your own rules and owners.
| Measure | Definition to agree | Avoid counting |
|---|---|---|
| Eligible jobs | Records that satisfy a versioned entry rule. | Repeated events as new jobs |
| Accepted jobs | Unique jobs passing all acceptance criteria. | Drafts or unverified API responses |
| Resolution effort | Time spent correcting and carrying exceptions. | Only the easy cases |
| Operating cost | Software, models, infrastructure and human work. | Token spend as the whole cost |
| Business movement | Observed orders, contribution or completed jobs. | Booked value as collected cash |
| Incrementality | Impact relative to an appropriate comparison. | All post-launch revenue as AI-generated |
| Uncertainty | Known data gaps and alternative explanations. | Missing cost data as zero |
Sources and scope
This is Looqus’s practical design guidance. Worked scenarios are illustrative; they are not client results or guaranteed outcomes. The sources below provide the referenced platform or engineering context.
AI does the work.We own the outcome.
Measure completed work against the full cost of delivery, review and repair, then connect the result to a business decision worth making.
- Agree the result
- Connect the work
- Own the follow-through