Did that automation actually give you time back?

01Monday: agree what would make this worth keeping

Start with a job you already understand: preparing a draft quote, sorting incoming enquiries, or turning approved notes into a weekly update. Keep the trial small enough to check every result. Write down what a finished, acceptable job looks like. A polished draft that still needs a complete rewrite is unfinished work.

The US Small Business Administration recommends starting small and testing whether AI adds value. Its guidance also recommends human assessment of AI-generated messages and outreach. [1] Here is our practical version: write your pass conditions before opening the new tool.

• Time: how many minutes must this return over a normal week, after checking and fixes?
• Quality: which errors need correction, and which single mistake would stop the trial?
• Money: what is the maximum trial spend, and how much ongoing cost is acceptable?
• Control: who approves the output, and how do you return to the old process?
• Evidence: how many representative cases will you review before deciding? If this week is too quiet, extend the test instead of guessing.

02Copy this scorecard into your working sheet

Swipe or scroll to compare

03Tuesday to Thursday: count the awkward cases

Run ordinary work through the trial and keep the old route available. Include a messy enquiry, missing information, or an unusual request when those occur naturally. Record the case before you know whether the automation handles it well. Cherry-picking easy jobs makes the sheet less useful.

For the week's time balance, subtract all trial effort from the usual effort for comparable completed work. Show setup separately, but include it in the first week's total. For recurring effort, exclude that one-off setup and label the result clearly. Count checking or corrections once, under a consistent category. If a job falls back to the old process, include both the failed attempt and the manual work.

NIST's AI Risk Management Framework calls for documented tests, metrics, and evaluation under conditions resembling actual use. It also asks teams to document what cannot be measured. [2] For this trial, a simple sheet with honest gaps is more useful than a dashboard filled with guessed numbers.

04Keep time, cash, and customers in separate columns

Recovered hours are capacity. Decide where they will go: serving existing customers, clearing a backlog, following up suitable enquiries, or finishing on time. Multiplying those hours by an hourly rate gives a planning estimate, not money in the bank. Record cash savings only when an actual expense is avoided or reduced, after added costs.

Faster replies do not establish more sales. Define a qualified enquiry before judging the trial—for example, a person with a relevant need and a realistic next step. Track enquiries through the same process, and leave later orders or margin unknown until you have the records. One week may reveal operational problems well before it reveals a commercial result.

05Friday: make one decision

Keep the trial when it clears your agreed effort, quality, and spending limits. Adjust it when a specific, fixable step is causing the trouble, then retest that change. Stop when a serious error crosses your boundary, costs exceed the cap, or the old process remains better. Keep approval on customer-facing commitments, prices, and sensitive information during the trial.

The UK Government's AI Playbook recommends establishing baseline measures and considering ongoing maintenance costs and alternatives. [3] You can apply that discipline without a large project: record this week's decision and book the next review. This is a proposed owner worksheet, not a tested savings claim. A week can justify the next small step; it cannot prove lasting revenue growth.

Sources are linked above so the argument can be checked independently. Editorial revisions retain the same public address.