The office manager is chasing an engineer's job notes. A customer wants an update. The contracts manager has the answer somewhere in WhatsApp, while the invoice sits unfinished in Xero. Someone suggests buying an AI tool.
I'd start somewhere else. An AI readiness assessment should tell you whether a particular job is ready for AI. How modern your business sounds has nothing to do with it.
You do not need to transform the whole company before trying anything. But you do need a clear task, usable information, someone responsible for the result and a way to tell whether the change helped. Otherwise, I would treat the proposed tool as an experiment until it proves itself.
An AI readiness assessment starts with the work
My starting question would be: which piece of work do you want to get through the business more reliably?
Asking “where could we use AI?” mostly invites a demonstration. Ask where work waits, gets copied, comes back incomplete or needs someone to explain it again. Then follow an actual job from arrival to completion. Sit with the person doing the chasing as well as the person buying the software.
For an illustrative field service business, the problem might be turning engineers' notes into customer visit summaries. Write the task plainly: take the completed job notes, prepare a draft summary and pass it to the service coordinator for checking. That is something you can test. “Improve productivity with AI” is not.
The UK Government's AI Playbook explicitly recommends choosing the right tool and recognising when established technology is a better answer. It is written for government and doesn't bind your contractor business, but I would borrow that principle without hesitation.
Before considering a product, write down:
- What starts the task and what counts as finished.
- Who does it now and who receives the result.
- Where the information comes from.
- What usually makes the work stop or come back.
- What an acceptable result looks like.
If the team cannot agree on those points, I would resolve the disagreement before testing a tool.
Check whether the information is usable
You do not need a spotless company database to run a useful trial. I would look for a small, reliable set of inputs for the task you have chosen.
Take that illustrative visit-summary example. Can the coordinator find the engineer's notes, identify the correct site and distinguish completed work from follow-up work? Does the record say what happened, or does it simply say “all done”?
Here is the limitation to keep in view: generative AI, the kind that produces text, can confidently generate incorrect information. The US National Institute of Standards and Technology's generative AI profile describes this as confabulation, often called hallucination. A polished answer can still rest on wrong information.
My practical test would be to hand the same records to a competent colleague who was not on the job. Could they prepare a useful summary without ringing the engineer? If not, improve the notes before expecting a generated draft to be trustworthy.
Try a consistent job-note structure: work requested, work completed, outstanding issues and next action. Agree where the current version belongs. Keep invented practice records clearly labelled and use them for early experiments rather than casually copying live customer material into a new service.
The readiness question is not “have we digitised everything?” It is “can we supply the information this task actually needs?”
Separate process fixes from automation and AI
I would put each proposed improvement into a plain category before spending anything.
Process change: the work needs clearer ownership, a better form or a different handover. In our illustrative example, perhaps the coordinator cannot finish the summary because nobody has agreed who confirms outstanding work. Assign that responsibility first.
Rules-based automation: the next step follows a clear instruction. For example, when a job is marked ready for review, create a task for the coordinator. If you can write the instruction unambiguously, I would test that route before asking an AI model to interpret it.
AI assistance: the task involves drafting or interpreting varied material, with a person checking the result. Turning complete job notes into a readable first draft is a candidate worth testing, with no promise it will work.
The government's AI Playbook lists summarising information and improving readability among potential uses of generative AI. It also stresses understanding the technology's limitations and having meaningful human control. That distinction matters: producing a draft and accepting it as correct are different jobs.
You might decide to combine a better form, an automatic notification and an AI-assisted draft. Fine. But make each part earn its place. Do not buy an AI workflow to avoid agreeing how a completed job should be handed over.
Name the owner and the person checking
“Someone will check it” would not pass my readiness assessment.
Name the person who owns the task, the person who checks the output and the person who handles a failure. In a small firm, these may overlap. What matters is that everyone knows where the responsibility sits.
For our illustrative service coordinator, the review instructions could be:
- Compare the draft with the original job notes.
- Check that the site and description of completed work match.
- Remove any claim that the notes do not support.
- Return unclear material to the engineer rather than guessing.
- Approve the summary before it goes to the customer.
Give that person time to do the checking. I would not count a fast draft as a saving if checking and rewriting it takes longer than preparing the summary manually.
Also agree the fallback. If the tool is unavailable or the draft is unusable, can the coordinator finish the job using the existing records? Keep that route available during the trial.
Ask the coordinator what would make them reject the output. Their answer should shape the test, so ask before anyone buys the system.
Run a bounded trial against the existing method
I would start with a task where a rejected draft causes inconvenience rather than an irreversible action. Keep the output inside the team until the reviewer accepts it. Do not begin by allowing generated text to send customer messages or change job records on its own.
Before the trial, record how the existing method performs. Include preparing the input, doing the work, checking it and correcting mistakes. Also note why work gets returned. Use the same boundaries when measuring the proposed method.
Include awkward examples alongside the clean ones: incomplete notes, conflicting descriptions, an unresolved follow-up and a record that should be sent back rather than summarised. Decide in advance what should happen in each case.
I would keep a simple review log with the input, the output, the corrections and whether the result was accepted. Keep supplier promises separate from what your own test demonstrates.
Then make a decision: continue, change the approach or stop. A trial that shows a clearer job sheet is more useful than an AI subscription has done its job. It has stopped you buying the wrong thing.
Decide what you are ready for
I would avoid giving the whole business a readiness badge. You may be ready to trial visit-summary drafts while being nowhere near ready to automate quoting. Assess the workflow in front of you.
My decision would fall into a practical category:
- Fix the process first: the task or handover is unclear.
- Prepare the inputs: the task makes sense, but the records need work.
- Try a simpler solution: a form or rules-based workflow looks sufficient.
- Run a controlled AI trial: the task, inputs, reviewer and success test are clear.
This week, choose a troublesome handover and trace it with the people involved. Write down the missing information, the owner and what would make the next person's job easier. That is the kind of starting point our Assess stage is designed around: understand the operation before deciding what to change.
You do not need to be ready for AI everywhere. You need a worthwhile problem and enough discipline to reject a tool that does not help.
If you want a second pair of eyes on that starting point, talk to Mise.