Three starting points that fit a small team
- Drafting. Replies to the same ten customer questions, job descriptions, handover notes. A person still sends them.
- Documents. Invoices, delivery notes, application forms — extract the fields, type less.
- Your own knowledge. A search box over policies and past jobs, grounded with RAG, so staff stop asking the same person.
These are the same patterns as in AI for business, sized for a team that cannot spare a project manager for six months.
Two things to leave alone for now
Autonomous agents that can pay suppliers or write to customers without approval. And “train our own model”. You do not have the data, and you do not need it. Use a hosted model, ground it, measure it.
Five starting projects that fit a small team
- Draft replies to the ten questions customers already repeat. A person still hits send.
- Pull supplier, date and totals off invoice PDFs into the accounts draft.
- A search box over current policies, with citations, so people stop asking the same colleague.
- Summarise a long job or case file for the next person on shift.
- Classify inbound mail into “accounts / support / other” so it lands in the right queue.
Each of those can go live in weeks if the files are already digital. None of them needs a data-science hire. All of them need an owner who will check the first hundred results.
What it costs a small firm, roughly
One of the jobs above, bolted onto software you already run, is typically a few weeks of AI integration plus a small monthly model bill. A “company-wide AI platform” is how small firms buy a login nobody opens. Start with one job. Measure hours. Then buy the second.
A week in a ten-person firm, without a “platform”
Monday: the same three customers email “where is my order”. Someone copies from the spreadsheet into a reply. Tuesday: a supplier PDF sits in the inbox until accounts has an hour. Wednesday: a new joiner asks which leave form is current. Thursday: a job file is twenty emails and nobody wants to write the handover. Friday: inbound mail is a mix of invoices, spam and a real complaint, all in one inbox.
Each of those is a small AI job if the files are already digital. None of them needs a strategy offsite. The owner’s job is to pick one, name who will check the first hundred results, and refuse the vendor who wants to “roll out AI across the company” first.
What you already have — and what you should not buy
You already have email, PDFs, and probably Tally or a billing package. That is enough to start. You do not need a data lake, a vector database you will administer yourself, or a fine-tuned model on fifty invoices. Hosted model, tight prompt, write into the screen people already use. That is adding AI to an existing application, sized for a shop that cannot spare a project manager.
Do not buy: an “AI receptionist” that books appointments you do not have a calendar API for; a chatbot on the website that cannot see orders; a tool that emails customers without a person. Those demos look busy. They create cleanup.
A thirty-day plan that fits around the real job
Days 1–5: pick one of the five projects above. Write the definition of done in a sentence. Confirm the files are reachable. Days 6–20: build the button or the mailbox watcher, read-only. Days 21–30: a named person reviews every result. Count hours and corrections. Then either widen that job or stop. A month is enough to know. Six months of workshops is how small firms fund someone else’s slide deck.
What “small” changes about the advice — and what does not
You cannot spare a project manager. So the owner is you, or one supervisor, for an hour a week, not a programme office. You cannot spare a data scientist. So you do not hire one. You use a hosted model and a button on software you already run. You cannot spare a bad month of cash. So you do not buy a platform licence that needs twelve departments to pay back.
What does not change: permissions, a definition of done, and a number you measure. A ten-person firm that skips those gets the same silent failure as a two-hundred-person firm, just faster, because there is nobody else to notice. The sequence is still add AI to an existing application, not “roll out ChatGPT”.
What we would tell an owner who has two hours this month
Pick the invoice PDFs or the ten repeated emails — not both. Write one sentence for done. Give one person the first hundred results. Do not meet a vendor who wants to discuss a platform. If you have no application you can add a button to, use a mailbox watcher that creates a draft in the tool you already pay for, still with a human post. That is enough AI for a quarter. The rest is curiosity, and curiosity is cheaper as a ChatGPT licence than as a project.
If even that feels heavy, you may want automation of a form you already have, and no model. That is a good outcome. It is not a lesser one.
Cost for one of those jobs, bolted onto software you run, is weeks plus a small model bill — not a platform. If a vendor needs a six-month programme to draft the ten emails, they are building a different company than yours. Start with AI for business if you want the menu; start here if you want the first Tuesday.
If the work is a fixed sequence with clean inputs, you may want automation and no model at all. See AI versus automation.
Two people in accounts, one mailbox
They already file supplier PDFs. They already know which ten customer lines they type every day. Give them extract-to-draft on the PDFs, or drafts of those ten replies — one, not both. The person who posts bills owns the first hundred extracts. You, the owner, look at the hours at the end of the month. That is a small-business AI project. It fits in the same month as GST filing if you do not add a steering committee.
What will not fit: a chatbot on the public site that cannot see orders, a voice agent on the phone, or “train a model on our WhatsApp”. Those are demos you will babysit. Keep the ChatGPT licence for curiosity. Keep the project for the mailbox you already have.
After the month, either the hours moved or they did not. If they did, buy the second job — the ten emails if you started with PDFs, or the other way around. If they did not, stop. A second feature will not rescue a mailbox that was never the bottleneck. That honesty is the whole advantage of being small: you can see the number without a programme office.