Your First AI Workflow for a Five-Person Team: Pick One, Run It for 30 Days

A five-person company decides to "use AI". In one month the owner tries it for marketing copy, the admin for customer replies, the bookkeeper for invoices and the salesperson for proposals. Each has a few good moments and a few bad ones. Nobody measures anything. In month two, usage fades and the conclusion is that "AI is not ready for us". The technology was fine. The approach was four experiments with no owner and no finish line. A first AI workflow should be one thing, chosen deliberately, with a date on which you decide whether it stays.

What a workflow is

A workflow is a repeated piece of work with a start, steps and an end. "Customer service" is a function. "Reply to enquiries about price and availability within ten minutes" is a workflow.

AI fits workflows. A vague instruction to "help with sales" gives vague results. A defined job, with examples and limits, gives something you can judge.

Four tests for a first workflow

Score each candidate from 1 to 5 on four points.

1. Frequency. Does it happen many times a week? Savings on a rare task never add up, and you get too few examples to learn from.

2. Similarity. Does each instance look much like the last? Repetitive work is where AI is strongest.

3. Low cost of error. If the output is imperfect, is it caught before it matters? Prefer work where a person reviews before anything is sent or recorded.

4. Ready information. Is what the AI needs to know already written down, or easy to write: prices, policies, standard answers?

Add the scores. The highest total is your first workflow. Resist the temptation to choose the most exciting one.

Five candidates compared

The scores are illustrative. Yours will differ. Notice which ones lose: the quotation, which owners often want to automate first because it takes the most thought, scores lowest for exactly that reason.

For most small service and trading businesses, the first two are the ones to start with.

Before day one: measure the baseline

For one week, before changing anything, record for the chosen workflow:

• How many times it happened.
• How long each took, roughly.
• How long customers waited.
• How many mistakes or complaints.

Ten minutes a day with a tally sheet is enough. Without it, day 30 becomes a debate about impressions.

The 30-day plan

Days 1 to 3: prepare

• Name an owner. One person, with two hours a week for this.
• Write the reference material. One page of facts: products, prices, hours, delivery, payment, returns. One page of tone: how the business speaks, with three example replies you like.
• Write the limits. What the AI must not do: promise delivery dates, offer discounts, discuss complaints. What it should do instead: hand over.
• Choose draft mode. The AI prepares, a person sends.

Days 4 to 10: first use

• The team uses it on every instance of the workflow.
• Each time a draft is edited, note why in one line: wrong fact, wrong tone, missing information, too long.
• The owner reads the notes daily and adjusts the reference material or instructions.

Expect many edits this week. That is the system learning what your business is.

Days 11 to 24: steady use

• Keep using it on every instance.
• Track weekly: drafts sent unchanged, sent with edits, discarded.
• Review twenty real examples each week.
• Fix the top reason for edits each week. One fix at a time.

Days 25 to 30: decide

Measure the same things as the baseline week and compare.

• Time per instance: lower?
• Customer waiting time: lower?
• Mistakes and complaints: the same or fewer?
• Share of drafts sent unchanged: rising?
• The team's view: would they object if it were removed?

Set the pass mark before you start. For example: time per reply down by at least a third, no increase in complaints, and at least seven in ten drafts sent unchanged by week four.

Then one of three decisions:

• Keep and extend. Consider auto-send for the simplest category, or start a second workflow.
• Keep and fix. It helps, with a specific weakness to address in the next 30 days.
• Stop. It did not earn its place. Write down why, and try the next candidate on the list.

Stopping is a legitimate outcome. A trial that ends with a clear no has done its job.

What usually goes wrong

• No owner. Everyone's project is nobody's.
• No baseline.
• Thin reference material. The AI cannot know your prices or policies unless told. Most bad drafts trace back to something missing here.
• Skipping draft mode.
• Judging on day three. The first week is the worst week.
• Several workflows at once.
• Not telling the team why. People who think the tool is there to replace them will not help it succeed. Say what it is for: fewer repetitive replies, more time for customers who need a person.

Costs to count

• The subscription.
• The owner's two hours a week.
• The team's time in the first week, when edits are frequent.

Compare these with the time saved by day 30. If a team sends 60 routine replies a day and each takes two minutes less, that is two hours a day. Your own figures will tell you whether it is worthwhile.

After the first one

A second workflow goes faster, because the reference material exists and the team knows the routine. Choose it from the same scoring table. Three or four well-run workflows change how a five-person team works more than twenty half-used experiments.

In Polynusa

Polynusa's AI agents can draft replies in the sales inbox from your business information, in draft mode or sending directly as the workspace decides, and can summarise conversations for notes and handover. Both of the top-scoring workflows above can be tried inside the trial period. See the pricing page.

Summary

Choose one workflow by scoring candidates on frequency, similarity, cost of error and ready information. Measure a baseline week. Appoint an owner, write the facts, tone and limits, and start in draft mode. Correct daily in week one, track drafts sent unchanged in weeks two to four, and decide on day 30 against a pass mark set in advance: keep, fix or stop.

FAQ

What is a good first AI workflow for a small business?

One that happens often, follows a similar pattern each time, is low risk if imperfect, and has a person checking the result. Drafting replies to common customer questions is the usual winner, followed by summarising long conversations for handover and drafting follow-up messages.

How long before we know whether an AI workflow is working?

About 30 days. The first week is setup and correction, the second and third show real use, and by the fourth you can compare time spent and quality against the baseline you measured before starting. Without a baseline, you will only have opinions.

Do we need technical staff to start?

Not for a first workflow built on a ready-made tool. You need one person who owns it, writes down the business's standard answers and policies, reviews the output each week, and adjusts the instructions. That is an operational job more than a technical one.