Often yes, but only for the tasks where you can show it. A small firm can tell whether AI is helping by tracking five numbers on one job before and after: hours spent, errors found, response time, revenue per enquiry and cost per task. If none of them moves within six weeks, it isn't helping.
Most owners judge AI by feel, and feel is unreliable in both directions. A tool that writes a clean first draft feels quick even when checking it takes longer than writing from scratch. A tool nobody enjoys using can still be saving an hour a day. Five plain numbers, noted once a week, settle the argument in about five minutes.
Why general reports on AI benefits tell you nothing about your firm
Search for whether AI is helping businesses and you will find large surveys, analyst forecasts and vendor case studies. They describe averages across thousands of companies. Your firm has perhaps four to forty people, a handful of recurring jobs and a payroll that makes a saved hour worth a specific amount. The only evidence that matters is what happens to your own jobs.
That is a narrower claim than it sounds. You do not need to prove that AI helps "the business". You need to prove it for one job at a time: enquiry replies, quote drafting, invoice chasing, job-sheet write-ups. Each job is small enough to time with a stopwatch and a tally sheet, and the result is a yes or a no you can act on, such as keeping the plan, changing the instructions or cancelling.
The method below is deliberately dull. It works because it is dull: a team of three can run it on a Friday afternoon without a dashboard, a consultant or a data analyst.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThe five measures, defined so two people would record the same number
A measure is only useful if different people record it the same way. Each of the five has a definition, a way to collect it and a trap.
| Measure | Definition | How to collect it | The trap |
|---|---|---|---|
| Hours | Staff minutes spent per completed task, from start to sent or filed | Tally sheet: task, minutes, date; five tasks a day is plenty | Counting only the typing and forgetting the checking |
| Errors | Tasks that needed a correction after they left the desk, per 20 tasks | A tick whenever a customer, colleague or you spots a mistake | Missing quiet errors that nobody reports |
| Response time | Working hours between an enquiry arriving and the first useful reply | Timestamps in the inbox or CRM; sample ten enquiries a week | Counting an auto-acknowledgement as a reply |
| Revenue per enquiry | Value of work won divided by the number of enquiries received | Monthly: total accepted, divided by enquiries | Reading too much into a single week of small numbers |
| Cost per task | Tool fees plus staff time at an hourly cost, divided by tasks done | Monthly sum from the other measures and the subscription bill | Leaving out the time spent setting up or fixing the tool |
Three of these (hours, errors, response time) you can read weekly. Revenue per enquiry and cost per task move slowly and are best calculated monthly, but the inputs are the same tally sheet. If you want a single page, keep one sheet with a row per week and a column per measure.
A before-and-after, worked through for an insurance broker
This is an illustration with made-up figures. Take a five-person insurance brokerage that gets about 60 renewal and new-business enquiries a month by email and web form. Each needs a first reply and, for most, a quote summary the client can read. The owner suspects AI could help with the first reply and the summary, and decides to test one tool for six weeks. Before anything changes, she runs a two-week baseline.
The baseline fortnight
Staff time each task. The owner samples ten enquiries a week for response time. Mistakes are logged when they surface. At the end of two weeks the sheet reads like this:
| Measure | Before (two-week average) | Note |
|---|---|---|
| Minutes per enquiry (reply plus summary) | 22 | Mostly retyping details from the form |
| Errors per 20 enquiries | 3 | Wrong renewal date, missing excess figure |
| Response time (working hours) | 5.5 | Replies batched twice a day |
| Revenue per enquiry | $310 of accepted commission | Last full quarter average |
| Cost per task | $11.00 | 22 minutes at a $30 hourly cost |
Notice what is already useful: the owner learns that the errors are two specific kinds, not "mistakes in general". That will shape the instructions she gives the tool.
The six weeks with a tool
The team uses a business chat plan on the task: paste the form fields, ask for a reply draft and a plain-English summary, then a person checks every figure against the policy documents. The plan costs, say, $25 a seat for two seats, so $50 a month, spread across roughly 60 enquiries, or about $0.83 an enquiry. Readings from weeks three to six:
| Measure | Before | Weeks 3 to 6 | What happened |
|---|---|---|---|
| Minutes per enquiry | 22 | 13 | Drafting fast, checking still manual |
| Errors per 20 | 3 | 2 | Fewer typos, but one new kind: an invented policy clause |
| Response time | 5.5 hrs | 2.0 hrs | Replies sent as drafts are ready, no batching |
| Revenue per enquiry | $310 | $335 | Too early to trust; flagged as a watch item |
| Cost per task | $11.00 | $7.33 | 13 minutes at $30 an hour is $6.50, plus $0.83 of tool fee |
The sums are simple. Thirteen minutes at $30 an hour comes to $6.50; add the $0.83 share of the monthly fee and each enquiry costs $7.33 against $11.00 before. Whichever method you choose for spreading the fee, write it on the sheet so next month's figure is calculated the same way.
The verdict is mixed in an honest way. Minutes and response time clearly improved. Errors fell slightly but a new type appeared, so the owner adds a rule: the tool may only reword figures supplied in the prompt, never supply them. Revenue per enquiry is parked until there are three months of data. That is what a useful check looks like: it ends in a decision and an adjustment, not a headline.
Setting the baseline when you have only a week
You do not need a perfect baseline. You need an honest one, taken before the new way of working begins. A few practical choices keep it cheap.
- Pick one job. If you measure everything, you will measure nothing. Choose the job with the highest weekly volume, since that is where a small saving multiplies.
- Time five tasks a day for five days. That gives 25 readings, enough to see the typical figure and a couple of outliers.
- Write down the definition of "done". Sent? Filed? Approved by the owner? Pick one and keep it for the whole test.
- Record the cost of an hour. Salary plus the extras the firm pays, divided by working hours, rounded to a figure you will actually use. $30 or $45 will do; precision is not the point.
- Keep the same person timing. Where possible, one person records for both the before and after period, so differences in how people count do not look like differences in performance.
If the tool is already live and nobody recorded the old way, fall back on a reconstruction. Ask two people who did the job before to time it the old way once or twice, using their spare moments in a quiet week. It is a weaker baseline, so label it "estimated" on the sheet, and expect to revisit it.
Four small firms, four different measure choices
The five measures do not carry equal weight in every business. These sketches are illustrations, not results, and they show how the emphasis moves.
An electrical contractor and the quote follow-up
The contractor sends about 25 quotes a month, and half never get a reply. The job is the follow-up message at day three and day ten. The lead measure is revenue per enquiry, since the only reason to automate follow-ups is to win more of the quotes already sent. Hours matter less: each message takes two minutes. Before the test, 25 quotes produce 9 accepted jobs averaging $1,400. After six weeks of drafted follow-ups, if 25 quotes produce 11, the extra $2,800 a month dwarfs the tool fee. If they produce 9 again, the follow-ups are not working, whatever the tool's marketing says.
A plumbing firm and the job-sheet write-up
Engineers finish a call-out, then spend ten to fifteen minutes typing up what they did and what they used. The lead measure is hours, and the second is errors, because an unclear job sheet causes invoice disputes. The firm lets engineers dictate notes into a phone and has an AI tool turn them into a structured sheet. If write-up time falls from 12 minutes to 5, across 120 jobs a month the saving is 14 hours. Errors are counted as sheets queried by the office; a rise in queries would show that speed was being bought with clarity.
A roofing company and the first response
Homeowners with a leak often contact several firms and book whichever one replies first with a firm survey date. The lead measure is response time, with revenue per enquiry as proof it matters. The company measures the working hours between a web form arriving and a human-checked reply that names a survey date. An AI-drafted holding reply is not counted, because it does not offer a date. If median response time drops from 4 hours to 40 minutes and surveys booked per ten enquiries rise from 4 to 5, the link is plausible, though three months of data are needed to rule out a spell of bad weather.
An insurance broker choosing between two tools
Months later the brokerage owner is deciding between two tools at different prices. Minutes and errors are identical in a short trial, so she compares cost per task directly: staff time plus fee per task. This is the situation the fifth measure exists for. When two options look equally good, the one with the lower cost per task wins, unless its errors are higher, in which case the saving is an illusion.
What a good week looks like on the sheet
Keep the weekly check to five minutes. Here is a filled-in line from a hypothetical sheet, then the rule for reading it.
Week 4 | Job: enquiry reply + summary
Tasks done: 16
Avg minutes per task: 12.5 (baseline 22)
Errors found: 2 of 16 (baseline 3 of 20)
Median response: 2.1 hrs (baseline 5.5)
Notes: one invented clause caught at check; tool re-instructed
Decision: continue, review again week 6
The "Decision" line is the one people skip, and it is the one that makes the sheet worth keeping. Each week choose one of four: continue, change the instructions, change the process around the tool, or stop. Write the reason in a sentence. A sheet that never contains the word "stop" probably isn't being read honestly.
Reading the numbers without fooling yourself
Small firms have small samples, and small samples jump about. Three habits protect you.
- Look at four weeks, not one. A single week with a holiday or a large job distorts everything. Compare four-week averages.
- Change one thing at a time. If you introduce a new CRM, a new price list and an AI tool in the same month, you will not know which moved revenue.
- Check the unflattering measure first. Before celebrating an hours saving, read the errors column. Speed bought with extra mistakes is a loan, and customers repay it.
Where the numbers mislead: three ways firms get a false yes
A false yes is worse than an honest no, because it keeps paying for something that is not working. Watch for these.
The checking time nobody logged
A roofing company's office manager trials a tool to draft supplier emails. Typing time falls from eight minutes to two, and everyone is pleased. Later she notices that she re-reads and re-edits every draft because early ones got a delivery day wrong. The true time is five minutes, not two. The fix is to record time from opening the task to sending it, so the sheet captures drafting and checking together. The tutorial on setting up human review for AI work without slowing down covers where checking is worth its minutes.
Cherry-picked tasks
If staff use the tool only on the easy enquiries, the average looks wonderful and says nothing about the hard ones. State the scope at the start: all enquiries of this type, or a random one in three. Without that, you are measuring staff's taste for easy work.
A rise in volume mistaken for a gain
If enquiries double in a good month, hours worked can rise even though each task is quicker. Use per-task figures (minutes per task, errors per 20 tasks), not totals, so busy and quiet months compare fairly.
Turning the readings into a decision about money
At the six-week mark you have a before and an after on one job. Now put a price on it, still with illustrative figures, using the brokerage's numbers.
Before the tool, 60 enquiries cost 60 x $11.00 = $660 a month in staff time. After it, they cost 60 x $6.50 = $390 in staff time plus the $50 fee, $440 in all. The net gain is $220 a month, or $2,640 a year, before any change in revenue. That would be a clear keep for a task that takes minutes a week to monitor.
Two caveats stop you over-reading it. First, a saved minute is only worth money if it is used. If the saved nine minutes an enquiry is spent scrolling, the saving is theoretical. Decide in advance where saved time goes: more enquiries answered, quotes sent sooner, or fewer overtime hours. Second, include the setup time. If the owner spent six hours writing instructions and testing, count that once against the first quarter's gain.
For a wider view of what tools cost across a whole business, see how much AI costs a small business, and use an audit of your AI subscriptions when your measures say some of them are not earning their place.
When the tool helps but the firm cannot show it
Some benefits are real and still awkward to count: a manager who now finishes at five, a team that dreads a report less, a more consistent tone across replies. Do not discard them, and do not pretend they have been proven. Record them as notes beside the numbers. If the measured gain is flat but the notes are consistently positive, the tool may be worth keeping on a monthly plan while you look for a measurable use.
Conversely, be wary of the opposite case: strong numbers on a job that does not matter. Saving 10 minutes a week on an internal newsletter is a result, but it is not a reason to expand AI across the firm. Put the effort where an hour is worth the most, which is usually customer-facing or cash-related work.
Questions to ask the team at the six-week review
- Which step now takes longest, and was it the longest before?
- What did the tool get wrong that a person would not have?
- Did anyone stop using it, and why?
- What would we have to see in the next four weeks to expand it to a second job?
Those four answers, with the sheet, are your whole case for or against. If you are unsure whether to bring in help to set this up, this tutorial on whether a 1:1 AI consultation is worth it lays out when it pays off. And if the question is really whether the work needs a person at all, AI or a new hire uses the same hours-and-cost reasoning.
A one-page template you can copy
Use one sheet per job. Keep the baseline in the top block and the weekly readings below. The example is filled in for an electrical contractor's quote follow-ups, with all figures illustrative.
JOB: quote follow-up at day 3 and day 10
OWNER OF THE SHEET: office manager
HOURLY COST USED: $30
SCOPE: every quote over $500
BASELINE (weeks -2 and -1)
Minutes per follow-up: 2.5
Errors per 20: 1
Reply time to customer replies: 6 hrs
Quotes sent: 25 a month | Accepted: 9 | Avg job: $1,400
Revenue per enquiry: $504 (9 x 1,400 / 25)
Cost per task: $1.25
WEEKLY LINE
Week | tasks | min/task | errors /20 | reply time | note | decision
1 | 12 | 3.4 | 2 | 5 hrs | learning | continue
2 | 14 | 1.8 | 1 | 4 hrs | tone odd | change instructions
3 | 13 | 1.2 | 1 | 4 hrs | | continue
MONTHLY
Quotes sent | accepted | revenue per enquiry | tool fee | cost per task
Notice that the sheet records scope, hourly cost and definition at the top. When someone reads it six months later, they can recompute every number.
Making the weekly check stick
A measurement habit lasts when it is short, has an owner and sits inside something that already happens. Attach the five-minute check to the Friday wrap-up, give one named person the sheet, and put the next review date at the top. When the review date arrives, make the decision even if the data is imperfect, since a firm that never decides drifts back to feel.
Finally, keep what you learn about errors. The log of mistakes is the most valuable output of the exercise, because it tells you what to put in the instructions, what to keep a person on and what to add to your simple AI risk register. Numbers tell you whether AI is helping, and the mistake log tells you how to make it help more.
Questions owners ask when the numbers look odd
How long before AI results show up in the numbers?
For a single task such as quote drafting or enquiry replies, four to six weeks of weekly readings is usually enough to see a trend. The first two weeks often look worse because staff are learning the tool and checking everything twice. Judge the trend from week three onward, not the first reading.
What if the numbers improve but the team says it feels worse?
Believe both. Time may be saved on the task while checking or fixing output is spreading to other people, or the work has become more tedious. Ask the team which step they now dislike, and add a sixth, informal reading: one sentence a week on how the task feels.
Can I measure AI if I never recorded anything before?
Yes, but you need a baseline now. Time the task by hand for a week, using a stopwatch or a simple tally sheet, before you change anything. If the tool is already in use, ask staff to estimate the old way, then treat that as a rough starting point and label it as an estimate.
Which of the five measures matters most?
It depends on the job. For admin tasks it is usually hours and errors. For anything customer-facing it is response time and revenue per enquiry. Cost per task is the tie-breaker when two options look equally good, because it puts tool fees and staff time in one figure.
When is it right to stop using an AI tool?
Stop when four consecutive weekly readings show no gain on the measure you chose for that job, or when errors rise and stay up after you have tightened the instructions. Cancel the plan, keep the baseline sheet and move to the next job on your list.
Further reads
- Who in Your Team Actually Needs a Paid AI Licence? — Match paid seats to the people whose numbers improved.
- Learn AI for Business Owners: A Four-Week Plan Built on Real Jobs — Learn AI on four real jobs, then decide what to keep.
- Questions to Ask an AI Vendor Before You Sign Anything — Ask the right things before a trial begins.
- Do I Need an AI Consultant? 8 Signs It's Time to Get Help — Eight signs that outside help would shorten the work.
- AI Security Risks for Small Businesses and How to Close Them — Check what a new tool exposes before staff rely on it.
- Is AI Worth It for a Small Subscription Box? Five Jobs First — Support, churn, curation, emails and stock forecasting for a small box business, ranked by hours saved against what each tool costs.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: the author's working method for small-firm baselines. All figures in the examples are illustrative.