Start with one repeating job, not a tool. Map how the work actually happens, decide what a person must still approve, and only then choose any AI to help with the preparation. That is how small businesses get a first result they can keep, rather than a licence nobody uses.
A lot of automation advice begins with a platform tour. That is the wrong end of the job. If you cannot name the work you want to change, you will not know whether the output is any better, and you will not know when to stop. The more useful starting point is a job your team already recognises, that happens often enough to be worth improving, and that a person can still check.
Pick one job, not a programme
Good first jobs are visible, frequent and reviewable. Preparing a weekly or monthly update from figures that already exist. Drafting a first reply to a common enquiry. Turning meeting notes into a list of actions for someone to approve. Finding the right policy or procedure without hunting through folders. Sorting a shared inbox so exceptions are flagged for a person.
Weaker first jobs are broad and hard to inspect: “automate operations”, “make the business AI-ready”, or “let the team use ChatGPT for whatever comes up”. Those may be later ambitions. They are not a first brief. If two possible jobs are competing, choose the one where you can describe what “better” looks like in a sentence, and where a mistake can be caught before it reaches a customer, a supplier or a service user.
Write the job down in plain language. Who does it now, how often, what they start from, what they produce, and where it gets stuck. That note is the brief. It is also the baseline you will come back to when you decide whether to continue.
Map the work as it really happens
Watch the current process, don’t rely on the official version. Where do the numbers actually come from? Which spreadsheet is the real one? Who is asked, informally, before anything is sent? Which steps exist only because a system doesn’t talk to another system?
List the information the job needs, and mark anything that should not go into an unmanaged tool. Customer records, staff data and anything you would not be happy pasting into a public chat tab need an agreed route, or they need to stay out of the first project. A first automation that uses a safe extract, or a template with the sensitive fields removed, is often enough to learn whether the idea works.
Then mark the decision points. Some steps are preparation: gathering, drafting, summarising, sorting. Some steps are judgements: a price, a promise, a rejection, a safeguarding concern, a change to someone’s record. AI can help with preparation. A person should still own the judgements. If you cannot tell those two kinds of step apart, you are not ready to automate the job.
Give the first version a narrow brief
Once the job is clear, build the smallest thing that helps with it. That might be better instructions and approved examples around a tool you already pay for. It might be a staff assistant that prepares a first draft in your voice, from templates and sources the team is allowed to use. It might be a short workflow that assembles a report and leaves it for review. It is rarely a new platform chosen before the job is named.
Write down the sources the system may use, the format of the output, and the person who signs it off. Test it on real examples, including awkward ones and cases where the right answer is “this needs a person”. If the output is only impressive on the easy cases, it is not ready.
This is the same discipline as the 12 questions to ask before your first AI implementation: one recurring job, evidence of what better looks like, a review route, and a named owner. Our AI process improvement work follows that sequence, starting with a free 40-minute call and a written next-step summary.
Measure one thing, then decide
You do not need a dashboard of AI metrics. You need a small before-and-after picture for this job. How long did preparation take? How often was the first draft usable? Were follow-ups missed? Did staff still have more checking to do than before? Speed alone is not a win if the review burden has simply moved.
Set a review date. Continue, change the brief, or stop. Stopping is an allowed answer. A tool that never shows up in the work is a cost, not a project. That is the point of AI Rescue when licences are already sitting unused, and of the practical advice in what to do when Copilot licences aren’t being used.
Common traps
Starting with the software. A convincing demonstration is not a design. If the first conversation is about features, bring it back to the job.
Letting anyone paste anything into a public chat. That is not process automation. It is an unmanaged habit, and it is a poor fit for customer or staff data. See can charities use ChatGPT with sensitive data if your organisation holds records of that kind, and can your AI be trusted for a wider checklist.
Automating a messy process as it stands. If the work depends on one person’s memory, or on a spreadsheet nobody quite trusts, tidy the steps enough that a system can follow them. You need a process you can describe, not a perfect one.
Skipping the handover. A first version that only the person who built it can run will fade the first week they are away. Write down how to use it, who owns it, and what to do when it fails.
Where this usually leads
Once one job is running, the next one is easier to choose: you already have a brief, approved sources, a reviewer, and a keep-or-stop point. You will not know whether the next step is a second role, a system connection, or simply better use of the original tool until the first job has been tried in real work, not in a slide deck.
Questions
Do we need new software to start?
Not necessarily. Many first jobs can be improved with the tools you already have, once the job, the sources and the review route are written down. New software is for when that is no longer enough.
How small can the first job be?
Small enough that you can test it on real examples and review it in a few weeks. A monthly report, a class of enquiry replies, or a meeting-note follow-up is plenty.
What if we already tried AI and it didn’t stick?
Go back to the original job, find whether the break was the brief, the sources, the review route or the tool, and decide whether to keep, fix, rebuild or stop. That is better than adding another licence on top.
Bring the job that keeps coming back. A free 40-minute call by phone or Teams. We’ll help identify one sensible first step and send a written summary of next steps.
Book a free call