Pick one AI workflow and stick with it for 30 days

By Jason Sibley

A new AI tool can make an ordinary Tuesday feel like a fresh start. By Friday, another launch has arrived. The first login is forgotten, a second trial begins and the team still writes the same follow-up emails by hand.

Small firms rarely suffer from a shortage of ideas for AI. They suffer from too many half-starts. Every tool asks for setup time and good examples. It also needs a little trust. Switching before that work has paid back makes AI look weaker than it is.

A 30-day focus is a useful correction. Pick one repeated workflow, use one agreed method and measure what happens. You are not choosing your forever system. You are giving one change enough time to show you something true.

What people get wrong

Tool choice gets more attention than work design. A team debates which model writes best, then gives the winner no source material and no clear definition of a good result. The software changes, but the job stays vague.

Another mistake is choosing a glamorous task that happens once a quarter. A pilot needs repetition. If the team only runs the workflow twice, it cannot spot where instructions fail or where review takes longer than expected.

Owners also mistake activity for adoption. Ten accounts and a busy group chat may feel like progress. If nobody can point to a repeated task that now takes less effort or produces fewer avoidable errors, the business has learned very little.

The last trap is allowing every person to invent their own route. Personal experiments are useful at the start. They do not become a company capability until the inputs, steps and checks are visible to someone else.

A practical frame for the 30 days

1. Choose a job with enough repetitions

Look for a task that happens at least several times a week and has a recognisable finish. Good candidates include turning call notes into a follow-up draft, preparing product descriptions from an approved fact sheet, or sorting incoming enquiries before a person responds.

Avoid work where a wrong answer would carry a high cost until you have stronger controls. Also avoid a task so loose that success becomes a matter of taste. “Make our marketing better” is not a workflow. “Draft the first version of the weekly customer email from these approved notes” is.

Write the start and finish in one sentence. If the sentence keeps growing, narrow the task.

2. Capture the old way before changing it

Run the task normally a few times. Note roughly how long the active work takes and where it stalls. Record one quality signal that matters, such as the number of corrections before approval or the share of drafts that need a complete rewrite.

This does not need a dashboard. A simple sheet will do. The point is to have a fair comparison with the new route. Memory tends to flatter both the old way and the new one, depending on the mood of the day.

Keep the measure close to the job. If the workflow drafts sales follow-ups, measure preparation time and whether the salesperson uses the draft. Revenue is important, but too many other factors sit between a draft and a signed deal for a short test to claim that connection.

3. Fix the input before polishing the prompt

Most weak outputs begin with thin source material. Give the tool the facts, examples and constraints a colleague would need. For a proposal draft, that might include agreed scope, customer language and the current offer. Do not ask the model to guess missing commercial decisions.

Store the approved input in one place. Remove old versions. If staff copy from whatever document happens to be open, you are testing inconsistency rather than AI.

Check what information is suitable for the chosen product and account. Vendor data terms differ by service and tier. OpenAI, for instance, separates its approach to individual services from products sold for business use in its data use policy. Read current terms for the product you actually use rather than assuming the brand name answers the data question.

4. Give the workflow an owner

One person should keep the working instructions, collect awkward cases and decide when a change enters the shared version. The owner does not need to be a technical specialist. They need to understand the task and care about its result.

Without an owner, each correction remains in someone's head. The next person repeats the same mistake, and the team concludes that the model is unreliable. With an owner, a correction becomes a better example or a clearer rule.

Protect the owner from constant tinkering. Change the shared method on a set day unless there is a data or safety problem. Otherwise, you cannot tell which version produced which result.

5. Use a human checkpoint that matches the risk

Decide what a person must check before the output is used. A low-risk internal summary may need a quick read against the source. A customer quote needs closer attention to prices, scope and promises.

Write the check beside the workflow. “Review it” is too vague. “Compare every price and delivery date with the approved sheet” gives the reviewer a real job.

The person reviewing should know that fluent wording is not evidence. AI can state a guess with confidence. Keep source links or source text close enough that checking is easier than trusting.

6. Keep a short exception log

For 30 days, record the cases that do not fit. Perhaps a call transcript is too poor to summarise. Perhaps the customer asks a technical question that must go to an expert. Perhaps the draft is fine, but moving it between systems eats the time saved.

Do not treat every exception as a reason to abandon the workflow. Group them. Some need a better input. Some need a clear route back to a person. A few show that the task was the wrong choice.

7. Decide on day 30

Compare the new route with the old one. Look at the measure you chose, then read the exception log. Ask the people doing the work if they would keep the method when nobody is reminding them.

Choose one outcome: keep it, change it for another short test, or stop. Stopping is a valid result if the workflow creates more review than value. The wasted move is to leave it running without a decision while the team starts three new trials.

The decision to make this week

Choose one repeated task and put the 30-day review in the diary now. Name the owner. Write one measure on the calendar invitation. Then tell the team which other AI experiments will wait until this test ends.

That pause matters. Focus is partly a decision about what not to start. A month of consistent use will teach you more than a folder full of trial accounts.

How we can help

Our AI sales & marketing desk works with one live commercial workflow at a time, from source material through review. If your team has been collecting tools without settling on a method, we can turn one repeated sales or marketing task into a working routine and help you judge it on evidence.