Where should a small business actually start with AI?

Why the usual starting points fail

Most people begin one of three ways, and all three tend to stall.

Starting with a tool. Someone buys licences, sends a link round, and waits. A few people try it, most do not, and three months later the subscription is a line item nobody defends. The industry data on this is unkind: across paid rollouts, something in the region of a third of employees end up actively using the tool they were given.

Starting with the most exciting idea. Usually the thing that would be most impressive in a demo, which is rarely the thing that is costing you money. Impressive and valuable are different axes.

Starting with a consultant. I say this as one: if you pay me before you know where your time goes, you are paying me to find out something you could have found out for free, and I will have a commercial opinion about the answer.

The common failure in all three is starting from the solution. The useful move is to start from the evidence, and the evidence is already in your week.

The one-week retyping audit

Here is the whole method. It is not clever, which is why it works.

What you are looking for

Every moment where a person moves information from one place into another by hand. Copying a figure from an email into a spreadsheet. Retyping a quote into an invoice. Rebuilding a report from an export. Summarising the same meeting into three different systems. Finding something from last year so it can be rewritten this year.

That moment is the signature of a gap between two systems that should be talking, and it is where almost all of the recoverable time in a small business lives. It is also, bluntly, the work that most reliably makes good employees feel like administrators.

How to run it

1. Tell people what you are doing, honestly. You are looking for friction in the process, not clocking anyone's speed. If people think it is a productivity review they will hide their workarounds, and the workarounds are precisely what you need to see.
2. Five working days. Long enough to catch the weekly rhythm, short enough that people keep doing it.
3. Record five things each time it happens. Keep it to a line.
4. Everyone, including you. Owners are frequently the worst offenders and the least aware of it.

How to read the results

At the end of the week, multiply the minutes by how many times a month each one happens. Then sort.

What usually surprises people is that the top of the list is rarely the thing they complain about. The dramatic monthly ordeal that everyone grumbles about comes twelfth. The quiet eight-minute task nobody mentions, which happens four times a day, is first by a distance. Eight minutes, four times daily, is over thirty hours a month.

Frequency beats intensity, and frequency is invisible without writing it down. That is the entire reason for the exercise.

What your list will probably say

Having done this with a number of businesses, the same handful of patterns come up almost every time:

• The same information entered two or three times because the systems that hold it do not talk to each other.
• Hunting for a past example before producing a new one. Proposals, quotes, reports. The search takes longer than the writing.
• Rebuilding a recurring document from scratch each period because nobody saved the method, only the output.
• Translating between formats. A spreadsheet into an email. Notes into a document. A document into a different document for a different audience.
• One person being the only route to a piece of knowledge, so everything queues behind their availability.

Every one of those is a good AI candidate, which is convenient. But they are not equally good first candidates, and that distinction matters more than people expect.

How to pick the first one

Three filters, applied in this order.

Does the information already exist somewhere digital? If the inputs sit in a system, a folder or an inbox, you can start now. If they live on paper, in photographs of whiteboards, or in one person's memory, then your first project is not an AI project, it is getting that information written down. Worth doing, but be honest that it is a different job with a different timeline.

How often does it happen? Daily beats weekly beats monthly, by a wide margin. Something occurring twelve times a year will almost never repay the effort of systematising it.

What happens if it is wrong? Not disqualifying, but it decides the design. Anything touching pricing, contracts or client commitments needs a person approving the output. That is fine. It just means you are building a drafting aid, not an autopilot, and you should say so up front so nobody expects otherwise.

Then pick the most frequent item with clean inputs. Not the most painful one. You are not trying to solve your biggest problem first. You are trying to prove to the people who will have to change their habits that this is real, and nothing does that like a tedious daily task quietly disappearing.

The part everyone skips

If your list shows the same information being entered into three systems, there is a decision to make before any AI gets involved: which one of those is the real record?

This is a management question, not a technical one, and it is free to answer. If three people describe your quoting process differently, automating it will encode the disagreement rather than resolve it, and you will have paid for the privilege. Sort the disagreement out first. It is often the single highest-value hour in the whole exercise, and it has nothing to do with software.

Consolidate before you automate. Every time.

What not to start with

• Anything customer-facing and unsupervised. Not because it cannot work, but because the cost of it going wrong lands on your reputation rather than your calendar. Earn confidence internally first.
• A company-wide rollout. One team, one task, proven, then spread. Rollouts that begin everywhere tend to land nowhere.
• The thing you saw in a demo. Demos are built to be demonstrated. Your list is built from your actual week.
• Custom software, at this stage. Almost everything on a first list can be done with business-tier accounts and some structure. Building comes later, if at all.

What the first month actually looks like

Assuming you have your ranked list and have picked one:

1. Gather the raw material. Three or four real examples of the task done well. This is the step that decides whether the output is any good, and the step people skip.
2. Do the task alongside the AI, once, with a real case. Not a test case. A real one, with a real deadline, where you will notice if the output is wrong.
3. Write down what worked. The instructions, the examples, the corrections. This is your playbook, and it is worth more than the output of any single run.
4. Hand it to the person who actually does the task, and watch them use it without helping. Everything that confuses them is a flaw in your playbook, not in them.
5. Check in a month. Are they still using it? If not, find out why. That answer is more valuable than the time you were trying to save.

That last step is where most of these efforts quietly die, and it is also the cheapest one to get right.

If you do nothing else

Run the week. It costs you five minutes a day per person and no money at all, and you will end it knowing something concrete about your own business that you did not know on Monday.

You might conclude that AI is not the answer to the top item on your list. That is a legitimate and useful outcome, and it is a lot cheaper to reach this way than after a pilot.