Practical thinking about AI, automation and data.

Most small businesses do not need an AI strategy. They need one repetitive, text-heavy task taken off someone’s plate — and a clear way to check the result.

The best first project usually has three traits. It happens often, so the savings add up. It involves reading or writing — emails, forms, invoices, documents — which is exactly where today’s language models are strong. And a person can verify the output quickly, so mistakes are caught before they matter.

Three good places to start

• Document processing. Invoices, purchase orders, application forms and contracts are read by people and re-typed into another system. An AI step can extract the fields, flag anything unusual and hand a clean record to the system you already use.
• Inbox triage. Sort inbound requests by intent and urgency, draft a first reply and route it to the right person. The human still presses send.
• Answers from your own files. A knowledge assistant that responds from your handbooks, proposals and procedures — with a link to the source — so new team members stop asking the same questions.

What to avoid at first

A generic chatbot bolted onto the website, or an open-ended writing assistant with no process around it. Both are hard to measure and easy to distrust. Start where the work is structured, the volume is steady and the result can be checked.

How to judge it

Count the hours the task takes today, run the AI step alongside the existing process for a few weeks, and compare accuracy and time. If it saves time without adding rework, extend it. If not, you have learned something cheaply.

Automating a broken process only makes mistakes faster. Before you connect any tools, look at five places where work usually leaks — and ask one question about each.

1. Customer intake

Enquiries arrive by form, email and chat, and are copied by hand into a spreadsheet. Ask: where does the first record of a customer actually live?

2. Invoice and expense approvals

Requests sit in inboxes waiting for a decision. Ask: who needs to approve what, and what happens if they do not respond?

3. Project handoffs

Work moves from sales to delivery and details get lost on the way. Ask: what information must travel with the handoff?

4. Status reporting

Someone compiles the weekly update by hand from three different tools. Ask: which numbers are always requested, and where do they come from?

5. CRM-to-accounting sync

The same customer is typed twice, with slightly different spellings. Ask: which system is the source of truth for each field?

Then automate

Simplify the rules first. After that, automate with clear triggers, automatic retries and an alert when something fails — so nobody finds out about a problem from a customer.

A beautiful dashboard on contradictory spreadsheets produces confident confusion. The chart is only as trustworthy as the numbers beneath it.

Four things to settle first

• One definition per metric. “Revenue” should mean the same thing in sales, finance and the board pack. Write it down.
• One owner per metric. Someone is responsible when the number looks wrong.
• One source of truth. Decide which system wins when two disagree, and feed the dashboard from that.
• Basic hygiene. Remove duplicates, standardise names and dates, and automate the clean-up so it stays clean.

Start small

Pick the three numbers leadership asks for every week. Build those well, show where each one comes from, and add more only when the first set is trusted.

Add alerts, not just charts

A dashboard that waits to be opened is passive. Set thresholds — a project reaching most of its budget, a lead waiting too long — and let the system tell you.

Every growing business reaches this question. The answer is rarely “build everything” or “buy everything”.

Ask these questions in order

• Is it a differentiator? If the capability is part of what makes your business different, building may be worth it. If every competitor does the same thing, it probably is not.
• Does something existing cover most of it? If a product fits most of your needs, buy it and accept the small gaps.
• Can the gaps be closed by connecting tools? An integration or automation often covers what a product lacks, at a fraction of the cost of a custom build.
• What is the total cost? Include per-seat licences, setup, training, maintenance and the cost of leaving later.
• Who owns it? Prefer approaches that keep your data portable and your options open.

A rule of thumb

Buy what is common, integrate what connects, build only what makes you different.

A short, independent review before you commit is usually cheaper than a wrong turn.

AI is useful precisely because it is fast and tireless. It is risky for the same reason: it can be confidently wrong at scale. The answer is not to avoid it, but to decide where a person stays in the loop.

Where a human should decide

• Anything involving money — payments, refunds, pricing, contracts.
• Anything sent outside the company — replies to customers, proposals, public posts.
• Anything involving personal or sensitive data.
• Anything that is hard to undo.

Where AI can act on its own

Sorting, tagging, summarising, extracting fields and drafting — work that is reversible and easy to check.

Make the review step easy

Show the AI’s answer next to its source, surface low-confidence items first and make approval a single click. If review is painful, people skip it.

Keep a record

Log what the AI proposed, what the person decided and why. An audit trail builds trust and shows where the system can improve.