Small Business AI Coach: Turn a Chatbot Into a Weekly Advisor

To turn an AI chatbot into a weekly business coach, create a saved project holding a short brief about your business and goals, upload the same one-page numbers sheet each week, run a fixed review prompt, and finish by recording three commitments. Next week the AI opens by asking what happened to those commitments.

The commitments loop is what makes it coaching rather than conversation. Ask a chatbot for advice at random and you get plausible suggestions that fade by Thursday. Give it the same structure every week, the same numbers, and a record of what you promised, and it can notice that you have postponed the same task three weeks running, which is exactly what a good human coach does.

Follow me on Instagram@sagnikteaches

The set-up has five pieces: the brief, the numbers sheet, the weekly prompt, the accountability step, and a clear view of the limits. The illustrations use a six-person signage company, because a business with quotes, installs and materials has enough moving parts to show how the review works.

Connect on LinkedInSagnik Bhattacharya

What you are building and why it lasts

There are four parts, and each fixes a particular way these set-ups fail.

Subscribe on YouTube@sagnikteaches

1. A standing brief so the AI knows your business without being told each time.
2. A weekly numbers sheet so advice rests on figures, not impressions.
3. A fixed review prompt so the questions are consistent and the output is comparable week to week.
4. A commitments log so the AI can check what you did, not just what you intended.

Most owners start with part 3 and skip the others. The result is generic advice ("focus on customer retention") that ignores their situation. The brief and the numbers are what make advice specific, and the log is what makes it stick.

Where to run it: a project, not a loose chat

Use the feature your chatbot offers for saved context. ChatGPT and Claude both have Projects, which keep instructions and uploaded files together and let every new conversation in the project start from them. Claude's Free plan lets you create up to five projects with their instructions, and ChatGPT's Projects are available across plans, including Free (which allows five files per project), so you do not need a paid plan to start. OpenAI is retiring custom GPTs (standard retirement is scheduled for 11 December 2026), so do not build this as a custom GPT. If Gemini is your assistant, build it as a skill: Google's help centre says Gems on personal accounts become skills and turn read-only on 17 November 2026, with work accounts following in 2027.

Name the project plainly, for example "Weekly business review". Keep one project per business. Within it, start one new conversation per week, titled with the date, so that the history reads like a diary.

Do not rely on the chatbot's automatic memory for this, even if it is switched on. You want to control exactly what the AI treats as fact, and the project's instructions and files do that; memory decides for itself what to keep.

Step one: write the standing brief

The brief goes in the project's instructions. Keep it to one page. It should say who you are, what the business sells, what you want, and how you want to be coached. The last part matters more than owners expect: tell the AI to be direct, to challenge you, and not to flatter.

You are my weekly business coach. Be direct and specific. Challenge my assumptions. Do not praise me unless the numbers support it. Never invent figures; if you need data you do not have, ask.

BUSINESS
- A signage company, six staff (two designers, three installers, one office manager/owner's assistant). I am the owner.
- Services: shop-front and vehicle signage, interior wayfinding. About 60% of revenue is repeat commercial clients.
- Average job value roughly $2,800. Quote-to-win rate has been about 35%.
- Busy period: spring and autumn. Slow: January and August.

GOALS FOR THIS YEAR
1. Raise gross margin from 38% to 42%.
2. Cut average days from quote sent to decision from 11 to 6.
3. Take two weeks' holiday in August without the business stalling.

CONSTRAINTS
- I can give about 6 hours a week to improvement work, not more.
- I cannot hire anyone this year.
- Cash is tight in the weeks before quarter-end tax payments.

HOW TO COACH ME
- Each week, start by asking how last week's commitments went.
- Then review the numbers I upload, compare with prior weeks, and name the one or two things that matter most.
- Suggest no more than three commitments for the coming week, each small enough to finish.
- If I keep postponing the same item, say so plainly and ask why.
- Tell me when a question needs a human adviser (tax, law, employment, large spending).

Notice the constraints section. Without it, the AI will cheerfully suggest hiring a salesperson and rebuilding your website. Telling it you have six hours a week and cannot hire changes the quality of the advice more than any amount of clever prompting. If you are unsure how to describe your business so answers fit, giving AI your business context goes into more detail.

Step two: design the weekly numbers sheet

This is a single page, the same layout every week, with no more than 12 to 15 figures. The consistency is the point: the AI can compare week 9 with week 8 only if both use the same rows and definitions.

Pick numbers that move when your decisions work. For the signage company they might be these. Figures are illustrative.

Two practical notes. First, include a small notes row for anything unusual: a big client lost, a machine breakdown, a holiday week. The AI cannot see these, and without them it will treat an anomaly as a trend. Second, define each measure once and stick to it. If "gross margin" is sometimes after labour and sometimes not, every comparison afterwards is wrong.

The sheet can live in a spreadsheet you update on Friday afternoon. If your accounting package already produces most of it, export those rows and paste the rest. For help deciding which measures to track, see AI KPIs for small businesses.

Rounding and summarising

You do not have to upload raw ledgers. Rounded weekly totals are enough for coaching, and they reveal far less. Leave out customer names, individual invoices, bank details and anything about staff pay. If you want the AI to discuss one client's pattern, call them "Client A".

Step three: the weekly review prompt

Save this prompt as a note, and paste it in at the start of each session, with the numbers sheet attached or pasted below it. Keep it identical from week to week and change it only on purpose.

Weekly review for the week ending [date].

1. Ask me first: how did each of last week's three commitments go? Wait for my answer before continuing.
2. Then read the numbers below. Compare with the previous week and the 4-week average. Tell me:
   a. The two figures that moved most, and whether the move is probably real or noise (say what you are basing that on).
   b. Anything that conflicts with my stated goals.
   c. One thing that looks good and one that looks worrying.
3. Ask me up to three questions you need answered to judge whether I am on track. Do not guess.
4. Suggest at most three commitments for next week. Each must be doable inside my 6 hours, name a concrete output, and say how I will know it is done.
5. List the assumptions behind your advice and the data that would change it.

NUMBERS: [paste or attach]
NOTES FOR THE WEEK: [anything unusual]

Step 1 is the accountability hook, and you should keep it even when it is uncomfortable. Step 2a asks the AI to say whether a move is real or noise, which is the single most valuable habit it can teach you. A rise from nine quotes to twelve in a week is not a trend; a margin that has slipped for four weeks running is.

What a session looks like

Here is an illustrative exchange, shortened. The business is the signage company, week ending 18 July.

AI: Before the numbers: last week you committed to (1) phoning the six quotes older than 8 days, (2) checking the vinyl supplier's price list against your job costings, and (3) writing a one-page install checklist. How did each go?

Me: Phoned five of the six, two said yes. Didn't get to the supplier prices. Checklist half done.

AI: Noted. Two of six old quotes now won is a good return on about 90 minutes, which is worth repeating. The supplier check has now slipped twice, and your margin goal depends on it: gross margin is 36% this week, down from 38%, against a 4-week average of 37% and a goal of 42%. Materials per completed job rose to about $1,120 ($5,600 over 5 jobs) from about $1,050 the week before ($6,300 over 6). What stops you doing the supplier check?

Me: I keep forgetting; it feels like a big job.

AI: Then make it small. Compare only your three most-used vinyl lines this week...

The coach noticed a repeated postponement, tied it to a stated goal, and shrank the task. That is the whole value. It did not know the answer to your margin problem; it kept you pointed at the task most likely to reveal it.

Step four: the commitments log

At the end of each session, ask the AI to write the three commitments in a fixed format. Copy them into a simple log, a document or a note, along with their status the following week.

Commitment 1: Compare the three most-used vinyl lines against current supplier price list.
Output: a one-row-per-line table showing cost per square metre in my job template versus supplier price.
Done when: table saved in the "Pricing" folder.
Time: 45 minutes.
Due: Wednesday.

Paste last week's log into the new conversation, or keep the log as a file in the project so the AI can read it. Each week it will start by asking about the previous three, which is the loop. After a month, ask it to summarise: which commitments were completed, which slipped, and what pattern shows up. You may find that you finish tasks that take under an hour and shelve everything that needs a half-day, which is a scheduling problem, not a motivation one.

Five ways to point the same set-up at different questions

A margin check

Once a month, add a job-costing extract to the weekly numbers: ten recent jobs with revenue, materials and hours. Ask the AI to find the three least profitable and say what they have in common. For a signage firm it might find that jobs needing a second site visit lose most of their margin, which suggests a surcharge or a better survey step. Verify any arithmetic yourself; checking margins product by product with AI shows how.

A quote-follow-up review

Paste a list of open quotes with their age. Ask which to chase first and what to say. This turns a vague worry ("I should follow up more") into a 20-minute task with named clients.

A decision memo

When you face a one-off choice, such as whether to buy a second wide-format printer at $14,000 or keep outsourcing, ask the AI to write both sides, list what numbers it would need, and say what would change its view. Use it to organise your thinking, then decide with an accountant.

A monthly zoom-out

Once a month, ask for a longer review of the last four weeks against the year's goals, with a verdict on whether the six-hour improvement time is being spent on the right things. This is where a pattern like "three weeks on the website, zero on margin" shows up.

A "what am I avoiding" question

Every so often ask: based on the commitments log, which important task have I postponed the most, and what is the smallest first step? It sounds soft, but it works because the answer is drawn from your own record.

A bad week: using the coach when the numbers fall

The first poor week tests whether the set-up is worth keeping. Say the signage firm sends only five quotes, wins one, and cash drops by $6,000 because an installer's van needed a major repair. A weak prompt would lead the AI to sympathise and suggest "refocusing on sales". The weekly structure gives it better material.

In the notes row you write: "Van repair $5,400, unplanned. One installer off sick for three days." The AI should separate what happened from what it means. A sensible reply would say that the quote count fell mainly because of the notes, not because of a change in demand, that the cash dip is a one-off rather than a trend, and that the single figure worth watching is the unpaid-invoices line, because cash has dropped below your $15,000 floor and overdue invoices are the quickest way back above it. It might then propose three small commitments: chase the two largest overdue invoices, check the van's warranty and insurance position with the right person, and block two hours on Monday to send the quotes that slipped.

If the AI instead treats the dip as a crisis, push back. Ask: "Which of these changes would still matter if you ignore the notes row?" Making it reason with and without the one-off events is a good habit, and it stops you overreacting to noise.

What the log looks like after a month

A filled-in log makes patterns visible. Here is an illustration, four weeks in.

Look at what the table shows. The supplier comparison was postponed twice, then shrunk to two product lines and completed. Every task with a small, clearly defined output got done; the 45-minute comparison slipped because "compare against the price list" felt open-ended until it was cut to two named lines. In a monthly review the AI can say this directly, and suggest that you shrink any task that has been postponed twice before you schedule it a third time.

Where the advice goes wrong

AI coaching has real limits, and it is worth stating them plainly.

• It does not know your market. It can tell you that your quote win rate fell, not that a competitor opened down the road. Add context in the notes row.
• It is eager to please. Even with a "be direct" instruction, models drift towards agreement. Ask it periodically to argue the opposite case, or to say what a sceptical adviser would object to.
• It can misread numbers. A pasted table with unclear headings may be misread, and long calculations may be wrong. Ask it to show sums, and check any figure you are about to act on.
• It is not a professional. Tax, employment law, contracts, insurance and borrowing decisions belong with a qualified adviser. Treat the AI's view as a way to prepare better questions for them.
• It forgets, or remembers wrongly. Keep the log and the brief as your own files, so the facts it works from are the ones you chose.

The sharpest risk is acting on a confident but unfounded suggestion. A good guard is the final step of the prompt: assumptions, and the data that would change the advice. If the AI cannot name them, the advice is not ready to act on. Using ChatGPT as a business adviser without being misled goes deeper on checking its reasoning.

Keeping your figures private

Your weekly sheet describes cash, margins and customers, so decide on privacy before the first upload.

1. Check the plan's data settings. Consumer ChatGPT and Claude plans let you switch off use of your chats for training in privacy settings. Business plans (ChatGPT Business, Claude Team) do not train on business content by default. Both require two seats at minimum, so a one-person business may prefer an individual plan with training switched off. How to stop AI tools training on your business data walks through it.
2. Summarise, do not dump. Weekly totals rather than ledgers; letters instead of customer names.
3. Leave out the regulated and the personal. No staff pay, no customers' personal details, no bank login information, nothing covered by a confidentiality agreement.
4. Protect the account. If the account sits behind a Google or email sign-in, the safety of that login is the safety of your history.
5. Delete when you wish. Know how to delete a conversation and the project, and do it for anything you regret uploading.

A simple test before each upload: if the sheet were forwarded to a competitor by mistake, would anything on it identify a customer, an employee or your bank? If so, take it out or replace it with a letter.

A four-week start plan

Week 1. Create the project, write the brief, build the numbers sheet from the last four weeks of history. Run a first review and choose three small commitments. Time: about two hours.

Week 2. Run the normal 30-minute review. Expect the AI's questions to feel obvious; answering them is the exercise. Keep commitments small enough to finish.

Week 3. Ask the AI which numbers on the sheet it never used. Remove them. Add one that you wish you had.

Week 4. Run the monthly zoom-out. Ask what has changed in your goals and whether the brief still matches the business. Update the brief if it does not.

After a month, judge it by what it changed, not by how helpful it felt. Count commitments completed, any figure that moved because of them, and any decision you made better. If none of those has happened, change the inputs before you give up, since thin numbers and vague commitments are far more common causes than the chatbot. How to review an AI tool after 90 days offers a way to decide whether to keep, fix or drop it.

If the review works, the same rhythm suits other recurring jobs, such as weekly quote chasing or monthly reporting. And if you would rather map this out with someone, that is what my AI implementation consultation covers; you can reach me through the contact page.

Questions about using an AI chatbot as a business coach

Can AI replace a human business coach?

Not entirely. AI is good at asking structured questions, spotting patterns in your figures and keeping you to a routine, and it is available at any hour. It does not know your market first-hand, cannot see your premises, and has no stake in your result. Many owners use it for the weekly rhythm and a human for big decisions.

Which chatbot should I use for weekly coaching?

Any of the main assistants can do it. What matters is a feature that keeps your standing instructions and files between sessions, such as Projects in ChatGPT or Claude. Pick the one you already use, and move only if the saved context or the privacy settings fall short for you.

Is it safe to upload my sales and cash figures?

Safer on a business plan that does not train on your content by default, and safer still if you upload rounded or summarised figures without customer names. On a consumer plan, check the privacy settings for training and chat history before uploading anything you would not want stored.

How long should the weekly session take?

Aim for 25 to 30 minutes: five to update the numbers sheet, ten for the AI review, ten to choose your commitments for next week. If it creeps past 45 minutes, trim the inputs, because a long session is one you will eventually skip.

What if the AI gives me bad advice?

Expect some, and build in a check. Ask it to state its assumptions and say what data would change its view, treat every recommendation as a hypothesis to test for a few weeks, and take anything involving tax, law, employment or large spending to a qualified adviser.

Further reads

• How Marketing Agencies Use AI to Automate Client Reporting — See how a recurring report is automated end to end.
• Is It Worth Automating a Task You Only Do Once a Week? — Judge which weekly jobs deserve automation.
• AI Session Notes and Follow-Ups for Coaches: A Weekly Workflow — A weekly notes-and-follow-up workflow from the coach's side.
• Why AI Projects Fail in Small Businesses (It's Rarely the Tech) — Spot the habits that make AI routines fade.
• How to Classify Business Data Before Using AI Tools — Sort which figures can go into an AI tool.
• Does ChatGPT Remember Conversations? Memory Settings for Business — What ChatGPT's memory carries between chats, and what it drops.
• Is It Safe to Put Customer Data Into ChatGPT? — What to keep out of a chatbot entirely.
• How to Scale AI From One Workflow to the Whole Business — Extend the weekly rhythm to other recurring jobs.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT and Claude help pages on Projects, project files and data controls; the ChatGPT Business overview and the Claude pricing page; OpenAI's help page on custom GPT retirement; Google's Gemini Apps Help page on the transition from Gems to skills.