What an AI assistant can actually do inside a small business

An assistant earns its place in a small business when the job is to read something, draft something, and hand it to a person who decides. It loses money when it is given authority: a credit limit, a refund, a price, an answer to a customer that nobody reads before it is sent. The distinction is not about how capable the model is. It is about whether a wrong output can be caught before it costs anything.

Which jobs does it actually do well?

Four, and they share a shape: text in, a proposal out, and a person who already reviews that kind of work.

Sorting an inbox. Messages from customers, suppliers, couriers and job applicants arrive in one mailbox with equal weight. An assistant can classify each one and route it to the right queue. The value is not the writing; it is that nothing sits unread for three days because it looked like a delivery notice.

Drafting a reply that a human sends. A draft that is mostly right removes the blank page, which is where the time goes. The rule that keeps this safe is narrow: the assistant drafts, a named person sends. Not "a human is in the loop" as a principle — a name attached to the reply.

Extracting fields from a document. A supplier invoice, a delivery note, a photographed order form. The assistant reads it and proposes the fields: supplier, date, line items, amounts. A person confirms, and only then is the record written. This is the highest-value use we see, because it removes retyping without removing the check.

Searching the company's own documents. Contracts, procedures, earlier quotations, the warranty terms nobody can find. An assistant that answers with the passage it used is useful; one that answers without a source is a rumour with good grammar.

We built one such tool for teachers: the assistant drafts the exam, every question stays editable, and the teacher approves the paper before export.

Job	What the assistant does	Who signs off
Inbox triage	Sorts, tags, drafts	Whoever replies
Document fields	Reads a page into a record	The record's owner
Internal search	Finds the clause, cites it	The reader
Customer reply	Drafts, never sends	A named member of staff

Where must it never have authority?

Four decisions, and the reason is the same in each: the cost of being wrong is paid outside the building.

Credit. Whether a shop receives goods on account depends on its history, its exposure and a judgement about the people running it. An assistant can assemble the history in seconds; it should not approve the limit.

Payments and refunds. Money needs an approval trail with a person at the end of it. An assistant can prepare a refund request and check it against the order; the release is a human action with a name on it.

Prices and quotations. A quoted price is a commitment the customer will hold you to. An assistant may assemble a quote from a price list it is permitted to read; a person signs it.

Unsupervised answers to customers. A wrong delivery date or stock claim becomes a promise the business must honour. Customer-facing text without review is the fastest way to turn a helpful tool into a complaint.

How do you bound it?

Four bounds, all of them checkable rather than aspirational.

Sources. The assistant reads a named set: this folder, this table, this mailbox. Not "the internet", and not the whole drive because it was easier to index.

Permissions. It retrieves only what the person asking could already open. The filter runs in the search, not in the prompt.

Human in the loop. A specific step where a named person approves, and the approval is recorded. "Reviewed by the team" is not a step; it is a sentence that means nobody.

Logging. Every question, the sources retrieved, the answer given, and who approved it. Without that log you cannot answer the first question after an incident: what did it say, to whom, on what basis?

What does a task cost, rather than a seat?

Per-seat pricing hides the real cost. What decides whether the tool pays is cost per completed task: the model call, the retrieval, the review time, and the rework when the draft is wrong.

Take one job — extracting fields from supplier invoices. If a clerk spends four minutes checking a proposal that saves twelve minutes of typing, the task pays. If the invoices arrive as angled phone photographs and the fields take fifteen minutes to correct, the same tool is a loss, and no licence price rescues it. Measure on twenty real documents from last month, not the vendor's sample.

How do you start without a project?

1. Pick one job that already ends with a person checking the work.
2. Count how often it happens and how long it takes now, on real examples rather than estimates.
3. Name the person who approves the output, and write down what approval means.
4. Restrict the sources to the minimum that job needs.
5. Run it beside the current process for two weeks, with the log switched on.
6. Compare time saved, corrections needed, and anything the assistant got wrong that a person missed.
7. Only then connect it to a system that writes.

The order matters. A tool connected to a system before it has been measured writes its errors into your records, where they are expensive to find.

Where this goes wrong

The pilot on tidy data. The trial runs over the neatest folder in the company and works. Production starts on the folder where the scanned documents live, and the gap between the two is the whole project.

No reviewer in practice. The reviewer exists in the design, is on leave for a week, and the queue is cleared by whoever is nearest. A rule without a name is a hope.

Answers built on yesterday's price list. Retrieval reads an export that stopped updating silently. The answer is confident, the prices are last quarter's, and the customer has it in writing.

Staff route around it. If the assistant is slower than the old habit for a week, people stop using it and the project is a failure by month two. A first version has to be faster on the most common case, not the most interesting one.

The unlogged incident. A customer says the assistant promised a delivery date. Without a log you cannot establish whether it did, so you compensate or deny on instinct.

What this does not cover

This is not about choosing a provider or comparing models, and it is not a security architecture. It says nothing about regulated advice — medical, legal, tax — which needs its own review and often should not be automated at all. It also assumes someone in the business can define what a correct output looks like, which is the part that cannot be delegated to a supplier.

We build assistants that draft and propose inside a reviewed workflow, and leave the decision with the person who signs it.