Pick one recurring job and compare the finished work. Did it meet the same quality bar? What happened to the full effort, including review and correction, and to cash? What did the business do with any time released?
Those questions tell an owner more than "we use AI every day" or "the first draft took five minutes". A quick draft can still need a long review. Time released gives people room for other work, and payroll stays where it was.
A small decision can rest on a handful of comparable jobs and records you already keep. A wider rollout, or a return-on-investment figure you plan to state in public, needs stronger evidence. Start with the decision you have to make, then collect enough evidence to settle it.
Where value can show up
AI may help people finish more work in the same time, respond sooner, cut revisions or take on work the business could not support before. Lower spending is one possible benefit among several.
| Possible benefit | Evidence worth looking at | What could cancel the gain |
|---|---|---|
| Capacity | Total effort to complete comparable, acceptable work | More review, rework or support elsewhere |
| Faster turnaround | Time from the request to an accepted result | A quicker draft that waits longer for approval |
| Better quality | Fewer material omissions or corrections against the same requirements | Confident mistakes or an easier mix of cases |
| Better client experience | Specific client feedback or fewer clarification requests | Polished but inaccurate work sent sooner |
| Lower spending | An actual drop in expense with the required result kept | New software, setup or supplier costs |
| A new option | Work that used to go unfinished now gets done | More output that nobody needs or uses |
State the result you can actually see. A proposal that goes out a day sooner may matter to a seller; whether it won is a separate fact. Fewer clarification emails may matter to a client lead; satisfaction needs the client's own response.
Keep each benefit in its own unit. Drafts generated show use. Net time shows capacity. Cash shows spending. Accepted work shows output. Pricing released minutes at a wage rate can help planning, and the wage bill stays the same. Revenue needs extra work sold, delivered and paid for.
Test the job you actually do. In a field experiment with 758 BCG consultants using GPT-4 (Harvard Business School Working Paper 24-013, September 2023), AI help improved results on tasks inside the AI's range. On a task outside it, consultants using AI were 19 percentage points less likely to produce a correct solution than those working without it. Averages from other people's work can point either way for yours.
Ten client updates, with all the effort counted
Made-up example. The business, records and figures are invented to show the method. They are not a client result or output from a real AI run.
A team prepares ten weekly client updates the usual way, then ten of similar length and difficulty with AI preparing a draft from approved project notes. Every update must match the notes, show open decisions and contain no client promise nobody approved. A person checks each one before it goes out.
| Effort for ten comparable updates | Before | With AI |
|---|---|---|
| Preparation | 10 × 40 min = 400 min | 10 × 15 min = 150 min |
| Fact and promise review | 10 × 10 min = 100 min | 10 × 18 min = 180 min |
| Corrections and rework | 10 × 5 min = 50 min | 10 × 9 min = 90 min |
| Extra support and setup in this period | 0 min | 30 min |
| Total counted effort | 550 min (9 h 10 min) | 450 min (7 h 30 min) |
Preparation alone looks like it releases 250 minutes. Longer review takes back 80, extra rework takes 40 and support uses 30. The net difference is 100 minutes, or 1 hour 40 minutes, across ten updates.
If all ten updates meet the same quality bar, that is a provisional gain of ten minutes per update. Provisional matters: ten cases can support a local decision and may not represent the next ten.
The wrong promise the review caught
In one update, the draft said "the client approved additional analysis" when the notes only recorded a request. The reviewer changed it to "additional analysis requested; approval and scope still open" before it was sent, and that correction time sits in the rework row.
Did the team save money?
Say the team also paid CAD 60 in added software cost for this period. Payroll and other spending stayed the same.
| Measure | Result in this example |
|---|---|
| Accepted output | Ten updates, if all pass the same quality check |
| Net capacity released | 100 minutes |
| Added cash paid | CAD 60 |
| Cash released | CAD 0 |
| Revenue or profit increase | Not shown |
| Use of released capacity | Not yet observed |
For this comparison, the business spent CAD 60 more. The team may still judge the capacity worth it. The next step is to say what the time is for and check that the work gets done. Who decides where the time goes covers that choice.
Suppose the owner gives the 100 minutes to an overdue client handover. That is a plan. A finished, accepted handover is an observed output. A better client response would be later evidence of benefit.
If a licence serves several jobs, or setup helps later months, split its cost on a stated basis and keep the original totals visible.
A measurement card for the owner
Use one card for one job and one decision.
| Question | What to enter | Filled line from the example |
|---|---|---|
| Which work and cases? | Job, period, number and comparable difficulty | Ten weekly client updates before; ten with AI |
| What counts as acceptable? | Required facts, decisions, promises and exceptions | Accurate project facts; open decisions visible; no unapproved promise |
| What was the full effort? | Preparation + review + rework + support | 550 minutes before; 450 with AI |
| What changed in cash? | Added costs and spending actually reduced | CAD 60 added; CAD 0 released |
| What happened to the time? | How it was used and what got finished | 100 minutes available; use not yet observed |
| What happens next? | Keep, change, expand or stop, with a reason | Keep the limited use and check whether the time goes to needed work; wider rollout undecided |
Made-up example from Sidekick Orchestration. Replace the assumptions with your own.
Project notes, expense records and a few ordinary examples are enough. The person responsible for the work judges whether each output is acceptable. You need no new dashboard or time diary for every employee. Individual feedback can stay private; a team decision needs the combined result and the exceptions.
A solo owner can use the same card for recurring quotes: accepted scope and price, full preparation and correction effort, cash cost and what the released time made possible.
Decide what the evidence supports
Keep the use when the work is acceptable and the benefit justifies the full cost. Change it when sources, instructions or the review step cause avoidable corrections. Stop it when it adds effort or produces work the business cannot rely on. Expand only when the method has held up on the work and people the expansion will involve.
Compare like with like. Busy seasons, simpler cases, a more experienced employee or a different client can each change the answer. A convenient before-and-after comparison can mix those effects in with AI's. Larger investment decisions need a better comparison and finance input that fits the size of the claim. For the full cost side, read What an AI setup costs.
If you need to decide whether to keep, change or expand an AI setup, book a free 30-minute call. Bring the job and the decision; we'll help you work out what evidence would settle it.


