A three-person startup should spend its first 90 days on three jobs only, chosen from the tasks that repeat every week and take at least two hours a week each: usually inbound email replies, meeting notes and follow-ups, and research or first drafts. Use one shared chat assistant, a written review rule, and a 20-minute weekly check-in.
The common failure is the opposite: each founder signs up to a different tool, tries it on whatever is nearest, and by week six nobody can say what it saved. This plan is about adopting AI inside a company that already exists, however young, not about building an AI product. It is built for three people who also have a product to ship, so every step is sized to fit around that.
Why three people need a different plan from a company of thirty
A larger firm can appoint an owner, run a pilot group and wait a quarter for a report. A startup of three has no spare person and no spare quarter. Everyone is already the sales team, the support desk and the finance department on some day of the week. That changes three things.
First, the job list has to come from what the founders actually do, not from a vendor's feature page. Second, the tool choice has to be boring: one assistant, one set of settings, one place where prompts live. Third, the review rule has to be written down, because with only three people one unchecked mistake reaches a customer or an investor quickly, and there is no second layer of editors to catch it.
If you want the wider map of where small businesses sit before starting, AI adoption stages helps you decide whether you are at "trying things" or "running a process". Most startups in week one are at the first stage, and this plan moves you to the second.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowDays 1 to 10: the job audit
Step one: list every repeating task, by person
Give each founder a 30-minute slot. They write down every task they did in the last two weeks that they have done at least three times. Pull from the calendar, the sent folder and the chat history, not from memory. Memory favours the interesting work and forgets the dull work, and the dull work is where AI pays off first.
For each task write three numbers: how often per week, minutes each time, and a score from 1 to 5 for how bad it would be if it went out wrong. A newsletter draft nobody reads scores 1. A reply to a paying customer scores 4. A message to an investor scores 5.
Step two: filter to the first three jobs
Multiply frequency by minutes to get weekly time. Strike anything scoring 4 or 5 on risk for now; those come later with tighter review. Then sort by weekly time and take the top three. If two of the three need the same input (say, both rely on a customer-conversation log), keep them: they will share context, which makes the setup cheaper.
| Task (illustrative) | Per week | Minutes each | Weekly hours | Risk (1-5) | Verdict |
|---|---|---|---|---|---|
| Reply to inbound product questions | 15 | 8 | 2.0 | 2 | First three |
| Write up sales and customer calls, then send follow-ups | 6 | 25 | 2.5 | 3 | First three |
| Research competitors and prospects before calls | 5 | 30 | 2.5 | 2 | First three |
| Draft the weekly investor update | 1 | 90 | 1.5 | 5 | Later, review hard |
| Chase unpaid invoices | 4 | 10 | 0.7 | 4 | Later |
| Write social posts | 3 | 20 | 1.0 | 1 | Backlog |
In this illustrative table the top three together come to 7 hours a week across the team. Saving even a third of that, about 2.3 hours, is the realistic early target. Do not plan around saving all of it; review time eats a share, as you will see below.
Days 5 to 15: choosing the tool without a comparison spreadsheet
For a team this size the choice is between a general chat assistant on a business or team plan and the AI that already sits inside the suite you use for email and documents. List prices change often, so confirm on the vendor page before you buy. As of November 2026, ChatGPT Business lists a standard seat at $25 a month on monthly billing or $20 billed annually, for teams from two users, and Claude's Team plan charges the same for a standard seat, for teams of 2 to 150. Google's consumer AI Pro tier is $19.99 a month, and Google Workspace's Business Starter, Standard and Plus plans include Gemini in Gmail.
Three rules settle it quickly:
- Pick the one that is already closest to the work. If the founders live in a Google or Microsoft suite, test that suite's built-in assistant first, because it sees the inbox and calendar without extra connections.
- Choose a business plan, not three personal ones. ChatGPT Business, Claude's Team plan and Google Workspace state that your content is not used to train models by default, which matters once customer details enter the picture. A personal plan such as ChatGPT Plus may use chats for training unless each person turns that off in their own data controls.
- One tool, one trial month. Put the plan on a monthly term first. Switch to annual only after day 90.
If one founder insists on a different tool, let them keep it, but only for tasks outside the three chosen jobs. The three jobs run on the shared tool so the weekly check-in compares like with like.
Before paying list price, look at the startup credits and discounts a small firm can actually claim, which shows what might offset the tool bill in the first year.
Days 10 to 20: write the one-page rules before anyone uses it for real
Write a single page and pin it where the team chats. It has four short parts.
What may be pasted
Say plainly: public information, your own marketing copy and anonymised examples are fine. Customer names with contact details, contracts, investor figures, credentials and unreleased source code are not, unless you are on a plan where you have confirmed the data handling terms. If a founder is unsure, they ask in the chat before pasting.
Who reviews what
Match the review to the risk score from the audit. Here is a workable ladder for three people:
- Risk 1 to 2 (internal notes, research, social drafts): the person who ran the task reads it before using it.
- Risk 3 (customer follow-ups, proposals): the person who ran it reads every line and checks every number against the source.
- Risk 4 to 5 (money, legal, investor, hiring): a second founder reads it before it leaves, and the AI draft is labelled as a draft in the file name.
What the tool must never decide
Prices, discounts, refunds, anything that commits the company to a date, and any statement about what the product does that has not been checked against the product. The AI drafts; a founder decides.
Where prompts live
One shared document, or a shared Project in the assistant if your plan has it. Every prompt that works gets copied there with a one-line note on what it is for. Without this, each founder reinvents the same instructions and the quality varies by who wrote the prompt that day. Giving AI your business context shows how to write the standing context block that sits at the top of every prompt.
Days 15 to 45: run the first three jobs, one at a time
Resist starting all three on the same day. Stagger them by a week, so each has its own bedding-in period and its own measured result.
Job one: inbound product questions
Create a context file with your product description, plan names, known limitations and the five answers you give most often. Then paste each incoming question with the instruction below.
You are replying on behalf of a three-person software company.
Use only the facts in the context block. If the answer is not in it,
say "I need to check this" and list what is missing.
Write a reply of under 120 words, plain tone, no exclamation marks.
End with one question that moves the conversation forward.
CONTEXT: [paste product facts]
CUSTOMER MESSAGE: [paste message]A realistic first output (illustrative) for a question about exporting data reads: "Yes, you can export all your records as a CSV from the Settings page. I need to check whether attachments are included, so I'll confirm that tomorrow. Are you moving to another tool, or just keeping a backup?" What you fix: the founder verifies the CSV claim against the product, because the AI only knows what the context block told it. If the context block was silent about attachments, the "I need to check" line is the instruction working correctly, and that is the behaviour to reward.
Record two numbers for the first week: minutes per reply before and after, and the number of replies you had to rewrite substantially. If more than one in four needs a heavy rewrite, the context file is thin; add the missing facts and rerun.
Job two: call write-ups and follow-ups
Use the transcript or notes from the call, strip personal details you are not allowed to paste, and ask for a structured output. A usable template:
From these call notes produce:
1. A five-line summary.
2. Decisions made, as bullets.
3. Actions, each with an owner and a date (write "no date given" if none).
4. A follow-up email under 150 words that states the actions back.
Do not add anything that is not in the notes.
NOTES: [paste]The risk is invented dates and invented owners. The "write no date given" instruction exists because assistants will happily guess a Friday. In an illustrative case, the notes say "we'll send the quote soon"; a careless draft turns that into "by Thursday". Your review rule for risk 3 catches this: the person who was on the call checks every date against what was actually said.
Job three: research before calls
Ask for a one-page brief on a prospect from their public website and public material, and require sources. Tell the assistant to separate "stated on their site" from "my inference". Check any figure, funding amount or headcount against the original page before it enters a message to the prospect; a wrong detail in an opening line does more damage than no detail. Where the assistant cannot browse, paste the page text in yourself.
For comparison with a different small business, the rollout plan for a ten-person agency shows how the same staging works with more people and a named owner.
The weekly 20-minute check-in
This meeting is what makes the plan a plan. Hold it at a fixed time, with an agenda that never changes, so it stays under 20 minutes:
- Numbers (5 minutes). For each live job: runs this week, minutes saved versus the audit estimate, rewrites needed, errors caught in review, errors that reached a customer.
- The worst output (5 minutes). Each founder brings the one result they liked least. Decide whether the fix is a better prompt, more context, or removing the job from AI.
- Prompt library (5 minutes). Add one improved prompt, retire one weak one.
- One decision (5 minutes). Continue, change, pause, or add the next job from the audit list.
Keep the tally in a sheet with one row per week. A filled-in illustrative set of rows looks like this:
| Week | Job | Runs | Minutes saved | Heavy rewrites | Errors caught | Errors sent |
|---|---|---|---|---|---|---|
| 4 | Product replies | 14 | 70 | 2 | 1 | 0 |
| 4 | Call follow-ups | 5 | 55 | 1 | 2 | 0 |
| 4 | Prospect research | 4 | 45 | 0 | 1 | 0 |
Week 4 here totals 170 minutes saved, which is 2.8 hours, in line with the earlier target. The "errors sent" column is the one that decides whether a job continues: any non-zero figure triggers a conversation about the review rule, not about blame.
Days 45 to 75: the realistic mistakes and how they show up
The confident wrong answer about your own product
An assistant told a prospect, in an illustrative example, that a feature "integrates with" a popular accounting tool. It did not. The assistant had filled a gap in the context file with a plausible guess. The fix was twofold: add a "what we do not do" list to the context, and add the rule that every capability claim in a customer message is checked against the product by the sender.
Prompt drift
By week six, three versions of the reply prompt exist, one per founder, each with small edits. Replies start to sound like three different companies. Fix it in the check-in: one master prompt, with changes agreed and dated.
Time saved that quietly disappears
A founder saves 20 minutes on a draft and spends 20 minutes tinkering with the prompt. If your numbers show savings but your week still feels full, count the setup time honestly for the first month, then expect it to fall. If it does not fall by week eight, the job is a poor fit; drop it.
Sensitive material pasted in a hurry
The one-page rules exist for this. If it happens, do not hide it: note what was pasted, check the plan's retention and training settings, and tighten the rule. A near-miss reported openly makes the next one less likely.
Days 60 to 90: adding job four and five, then deciding what to keep
By day 60 you have a measured baseline. Now draw on the audit list. Good next candidates are the ones you struck earlier for risk: the investor update, invoice chasers, and hiring messages. They need the second-founder review step, so add one at a time.
Take the investor update (risk 5) as a worked case. The founder keeps a running log of facts during the week: revenue figure from the finance sheet, new customers, shipped features, hires, blockers. On Friday they paste the log, with the real figures, and ask for a 250-word update in the format of last month's. The assistant organises and phrases; it never calculates or recalls a number on its own. The second founder compares every figure in the draft with the finance sheet before sending. The time saved is smaller, perhaps 40 minutes of a 90-minute job, but the consistency of the format is a gain in itself.
For invoice chasers, a tiered set of three messages (friendly, firm, final) works well, written once with placeholders for names and amounts. The AI is not even needed every time; once the templates are good, a founder fills in the blanks. Say so honestly at the check-in: sometimes the right outcome of an AI trial is a good template.
If you are an online seller rather than a software firm, the same method applies with different jobs: how a small online shop uses AI to compete shows which tasks come first there. Illustratively, a garden centre's first three might be supplier stock queries, product descriptions and answering opening-time enquiries; a sports club's might be fixture emails, membership renewals and sponsor outreach.
The day-90 review
Set aside one hour with all three founders and answer five questions from the sheet:
- How many hours per week did the live jobs save, averaged over the last four weeks?
- What did the tools cost in that period, including every seat?
- How many errors reached customers or investors, and what was changed afterwards?
- Which job should stop, and which should extend to a second use?
- Is the one-page rule still right, and does it need a section on a new tool or data type?
A simple sum frames the decision. If three seats cost $75 a month and the jobs saved 10 hours a month between them, the tool costs $7.50 per hour saved. Whether that is a good deal depends on what a founder hour is worth to you, and every founder hour matters more than most line items. If the sum does not clear that bar, either the jobs were poorly chosen or the review time is higher than assumed; fix that before adding tools. The method for this arithmetic is laid out in finding AI savings in your profit and loss.
Decide, at day 90, between three paths: keep the same tool and add jobs, move to annual billing if the numbers support it, or pause AI for a quarter if the founders agree the tasks were wrong. All three are legitimate outcomes. What you should not do is drift on without a decision.
What to keep manual, even after 90 days
Some work stays human on purpose. Anything that sets price or terms for a customer. Hiring decisions, as opposed to scheduling messages. Product direction, and the first conversation with a customer who is upset. AI can prepare notes for these; it should not hold the pen. Keeping this list short and explicit makes it easier for new hires to follow later, and it is easier to defend to customers.
If you would rather map the audit and the review rules with someone who has seen the pattern across businesses, that is what my AI implementation consultation covers, and you can reach it through the contact page. Otherwise, the plan above is complete enough to run yourselves: audit in the first ten days, three jobs staged by the end of day 45, a weekly check-in from day 15, and a hard review at day 90.
Questions founders ask before week one
How much should a three-person startup spend on AI tools?
Start with one shared chat assistant plan, which is roughly $20 to $25 per seat per month on the main business plans, and nothing else for the first month. Add a second paid tool only when a logged job proves it saves more hours than it costs. Three seats on monthly billing come to about $75 a month.
Should every founder use the same AI tool?
For the first 90 days, yes. One tool means shared prompts, shared Projects and one set of privacy settings to check. Different tools per person makes the weekly check-in impossible, because nobody can compare results or reuse each other's instructions.
What if the founders disagree about using AI at all?
Run the 30-minute job audit anyway and let the numbers decide. A sceptic usually engages once they see their own hours listed next to a task the tool handled. Agree one small trial job and a date to review it, rather than a company-wide policy.
When does a startup need a written AI policy?
Before anyone pastes customer data, investor material or code into a tool. A single page covering what may be pasted, which accounts are allowed and who reviews outputs is enough. Write it in week one and revisit it at day 90.
Further reads
- Barriers to AI Adoption in Small Businesses and How to Clear Them — Common blockers and how to clear them early.
- Small Business AI Statistics 2026: Adoption, Spend, and Results — Context on how other small firms are adopting AI.
- Can AI Manage Your Calendar and Book Meetings for You? — A natural fourth job once the first three run well.
- AI Email Triage for Professional Firms: Sort, Summarise, Draft — Inbox sorting rules that suit a busy small team.
- Why Most Landscapers Haven't Adopted AI Yet, and Where to Start — How a trade business picks its first AI job.
- Startup AI Tools for Growth: 8 That Earn Their Subscription — Eight growth-focused AI tools judged by the job they do, ranked by payback, with a 14-day trial rule and a cancel list.
- AI Marketing for Startups: A 30-Day Plan Without a Marketing Hire — A four-week founder schedule with copyable prompts, a worked budget and a decision at day 30 based on real conversations.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatGPT Business pricing page and OpenAI help article on how data is used, Claude pricing page, Google One plans page and Google Workspace pricing page (checked 5 Nov 2026).