Part of our guide to What is an AI execution layer? A plain-English guide for operators
On this page · 6 sections
Consider a sales team preparing a weekly account brief. Is one job a draft, an approved brief or an email actually sent? Pick the point where the organisation receives the value it is paying for. If the result only becomes useful once the manager checks the figures, count an approved brief. Otherwise a system that produces five unusable drafts can look more productive than one that produces a single good one.
Define what counts as one finished job
Write a short acceptance rule. For example: the account owner can locate the original request, the brief identifies the latest commercial term, the next action has a person responsible and any uncertain date is marked for confirmation. Save one example that passes and one that fails. Give them to reviewers before changing the tools. The examples help distinguish a real difference in quality from two managers having different tastes.
Define the period you are measuring. A monthly subscription divided by one afternoon of work can make a useful tool look absurdly expensive. A busy launch week projected over an entire year can make a marginal purchase look irresistible. Pick a normal period, record the number of eligible tasks in it and repeat the calculation when the workload changes materially.
Make the arithmetic visible to the person paying the invoice
Use four figures: allocated fixed cost, variable usage, human review and rework. Divide their sum by the number of accepted results. Treat setup and training separately at first so the buyer can see both steady operation and the first month's cash requirement. Record whether tax is included consistently when using supplier quotes. Keep prices from different billing periods in separate rows until converted to the same period.
| Illustrative monthly input | Example amount in AUD | Why it belongs |
|---|---|---|
| Allocated subscription share | 120 | The portion assigned to this workflow under the team's stated allocation rule |
| Usage for all attempts | 30 | Failed and repeated attempts consumed resources too |
| Review, two hours at 60 per hour | 120 | Somebody checked the outputs before they entered the sales process |
| Rework, one hour at 60 per hour | 60 | Repairs happened outside the original review |
In this invented example the operating total is AUD 330. With 22 accepted briefs, the cost is AUD 15 per approved brief. If only 11 pass, the same spending yields AUD 30 per approved brief. These are arithmetic examples, not reported customer results or Fulcrum prices. Replace every figure with your own. For current Cortex plan details, consult pricing and confirm which usage the quote includes.
Avoid hiding the expensive second half of the job
Track the document after the assistant says it is ready. A reviewer may spend time checking the source of a claim, resolving a duplicate, correcting a figure and asking the author what the proposed next step means. Those minutes count even if they happen in another application. For a short pilot, a paper tally beside the reviewer's desk can be easier than changing every system to log events.
Distinguish correction from a change in the brief. If the account manager missed a required source that was available, book the extra work as rework. If the customer changed the delivery date after the brief was accepted, that is new work. A pilot that charges every subsequent change to the assistant can exaggerate its costs. A pilot that calls every failed draft a new request can make its results look better than they are.
Record the time spent waiting separately from the time a person spends working. A ten-minute delay may have little labour cost if the reviewer can work on something else. The same delay can be expensive if it causes an estimator to miss a submission window. Bring that operational consequence into the decision as a separate observation. Do not turn every elapsed minute into labour cost automatically.
Compare the purchase to the existing method
Give the normal process the same acceptance rule and the same mix of cases. Let experienced staff use the templates and shortcuts they already rely on. Include their existing review time. If the current method is already paid for, separate the cash you could actually avoid from time that might become available for other work. A saved half-hour does not lower payroll unless your staffing or spending changes.
Look at the difficult cases separately. Suppose routine briefs pass quickly while cases with conflicting sources take a supervisor an hour. If the pilot contains mostly routine cases, the average may hide the very category that drives overtime. Record how often exceptions occur in ordinary work and what a safe handover looks like when the tool cannot complete them.
Change one assumption at a time. Halve the expected volume, double review time or remove a subscription you genuinely expect to cancel. Does the choice change? A case that depends on every optimistic assumption deserves a smaller trial. A case that remains attractive under a range of ordinary volumes gives the buyer a clearer reason to proceed.
Decide what you will measure after renewal
Give each metric an owner. Finance can confirm invoices and overlapping subscriptions. The operations lead can count eligible jobs and exceptions. The reviewer can report how many outputs were accepted without reconstruction. Agree on one location for the monthly numbers. An account manager should be able to reproduce the calculation without access to the consultant who set up the pilot.
Set a point at which you would change the workflow. For example, if one category repeatedly consumes most of the review time, move it back to a person and keep the assistant on the routine subset. The new denominator must reflect that narrower scope. Carry the excluded jobs into a separate line so the improvement in cost per accepted result does not hide shrinking coverage.
A buying decision should capture who benefits from the released time. If sales staff use it to contact prospects sooner, measure the follow-up interval as a separate outcome. If the business needs faster reporting during renewal season, measure whether the deadline was met. These outcomes make the cost figure meaningful without guessing how many dollars of revenue a saved minute produced.
Related guides: What is an AI execution layer? A plain-English guide for operators · How to measure AI ROI in hours saved (without inventing numbers) · Can one AI subscription replace 20 SaaS tools? The honest maths.
Frequently asked questions
Should we divide by the number of users or tasks?
Use seats to understand the invoice and accepted tasks to understand the workflow. An infrequent user can still approve an essential job. Keep both figures so the purchasing team can see why a seat exists.
What about a tool used by several departments?
Agree on an allocation rule before comparing candidates. You might use measured activity or an agreed share of capacity. Also show the full invoice so that reallocating costs between teams cannot be mistaken for a saving.
When does it make sense to stop a pilot early?
Stop if the team cannot define an acceptable result or cannot obtain the evidence needed to review one. An extra month of usage will add little insight until that gap is resolved.



