An AI tool writes a client email in seconds. That looks like a saving, until someone checks the facts, removes a promise you never agreed to and rewrites half of it.
Sometimes that still beats writing from scratch. Sometimes it takes longer. A small business needs a way to tell the difference before buying tools for everyone.
We suggest a 30-day test on one recurring task. It should answer two questions: does the team finish useful work faster, and does the benefit justify the full cost?
The quick version
- Choose work that repeats often and that someone on your team can check reliably.
- Measure the time to a usable result, including preparation, review and corrections.
- Count setup, training and tool spending before deciding what you saved.
- Keep cash savings separate from the value of time freed up for other work.
There is a reason to test this locally. In the OECD's 2025 SME survey report, 65% of businesses using generative AI reported improved employee performance. The survey did not measure the size of that improvement. It is evidence of potential, rather than a forecast for your business.
Pick one task with a clear finish line
“Adopt AI” is too broad to measure. “Turn client-call notes into a checked follow-up email” is specific enough to test.
Other starting points include drafting routine proposals from an approved template, organising meeting actions or summarising supplier documents. Choose something frequent enough to produce several examples during the month.
Write down what counts as good. For the follow-up email, that might mean correct facts, no invented commitments, a clear next action and your usual tone. Name the person responsible for checking it.
Start with material your business permits in the chosen tool. If client information cannot be shared there, use anonymised examples or choose another task.
Days 1–7: find out what the task costs today
Record several examples completed the usual way. A small table is enough:
| What to record | Why it matters |
|---|---|
| Type of task and difficulty | Stops a simple AI example being compared with a difficult manual one. |
| Total staff minutes | Includes everyone involved, even when a colleague does the checking. |
| Corrections and final quality | A faster result must still meet your standard. |
| Number completed | Shows whether the task happens often enough to matter. |
Use representative cases, including a messy one. Keep the same quality standard during the pilot. A polished-looking draft with missing facts is unfinished work.
Days 8–21: run a small pilot
Choose two or three people who do the task, set a modest spending limit and show them one example of a good result. Include anyone who will have to review their work.
Give the model the source material, constraints and desired format. A reusable instruction for the email task could be:
Draft a client follow-up email from these notes. Use only facts in the notes. Do not add prices, deadlines or commitments. Put missing information under “Questions to confirm”. Keep the email under 200 words. Match the tone of the approved example.
Try a lighter model first. If corrections are frequent, compare a stronger one on similar cases. Our guide to choosing AI models gives starting points.
Keep recording total time and quality. Include failed attempts, retries, copying between tools and a colleague's review. Track elapsed turnaround separately if faster delivery matters to clients.
Count the whole job, including the first month
Here is an illustrative example, using US dollars throughout. These are invented figures, not measured results or a Drephos price quote.
| One month's client follow-ups | Without AI | With AI |
|---|---|---|
| Emails completed | 40 | 40 |
| Preparation and drafting per email | 20 minutes | 5 minutes |
| Review and corrections per email | 10 minutes | 10 minutes |
| Total time for the emails | 20 hours | 10 hours |
| Additional pilot setup and training | None | 3 hours |
The pilot frees seven hours in its first month: 20 minus 10 minus 3. At an illustrative staff cost of $30 an hour, that is $210 of capacity. Subtract $30 of additional tool spending and the estimated net capacity value is $180.
That is not automatically $180 saved in cash. If salaries stay the same, the payroll bill stays the same. The benefit might be seven hours for client work, less overtime or a shorter backlog. Record what actually happens to that time.
Cash savings need a separate calculation: expenses actually avoided, less additional cash costs. Include cancelled subscriptions, reduced paid overtime or avoided contractor spend only where those reductions really occur. Do not count the same recovered hours twice.
Spend less on tools while you learn
Check what your current software already includes. An existing tool may handle the pilot without another purchase.
Then compare billing against observed use. A fixed seat can suit someone using AI heavily every day. Usage billing can suit a team where only a few people use it regularly. Include the features and administration you need, not just the price of a message.
List overlapping subscriptions before cancelling any. Make sure the replacement handles the work, and check renewal dates or commitments. The pay-as-you-go guide explains that billing option; the calculator helps estimate tool spending, but does not measure productivity savings.
Days 22–30: decide what earns a place
Review the examples with the people doing and checking the work. Did quality hold up? Was time saved consistently? Can the team use that time, and does the benefit cover the costs?
Keep the workflow if those answers are convincing. If it works only on simple cases, define that narrower use. If review takes longer than the original task, revise the instructions or stop the pilot.
Low-volume or seasonal tasks may need a longer test. Before expanding, include ongoing administration and any extra setup for new users in the next estimate.
Where Drephos fits
Drephos lets a team try several models in one web workspace, using a shared prepaid wallet without seat fees. Owners can set monthly spending caps per member and see their usage. Shared chats let a colleague review the work in the same conversation.
That gives you a way to contain tool spending during a pilot. The time and quality records still need to come from your team. Start with one useful workflow, keep its results and costs visible, and use those results to choose the next one.