It is easy to start an AI project with a tool and then look for something to use it on. A more reliable route is the reverse. Begin with a piece of work that is costing your people time, and ask whether automation, with or without AI, is the right way to remove it.

Begin with a task, not a tool

Pick one specific task that someone in your business does repeatedly and would happily stop doing. Keep a short list of candidates and choose the one with the clearest owner. "Improve our admin" is too broad. "Sort incoming certificates and file them against the right person" is a task.

Describe it in a few sentences: who does it, how often, what they start with, what they produce, and who checks it. Watching someone do the task for an hour is often more revealing than asking them about it. If you cannot describe it, you are not ready to automate it. If you can, you have the beginnings of a brief.

Note also that many tasks do not need AI at all. Plain rules, reminders and workflows will often do the job, more cheaply and more predictably. AI is worth reaching for when the input is untidy, such as a scanned form or a free-text email, and rules alone cannot handle it.

A test for a good first candidate

A good first task should pass most of these five checks.

  1. It is repetitive. The same steps, again and again.
  2. It is high volume. There is enough of it that saving a few minutes each time matters.
  3. It follows rules. You could explain how to do it to a new starter in a short document.
  4. It needs little judgement. The decisions are routine, not delicate.
  5. It is easy to check. A person can look at the result quickly and tell whether it is right.

If a task fails two or more of these, leave it for now and choose another.

Of the five, being easy to check matters most. If checking the output takes as long as doing the task, there is little to gain.

Examples that can suit it

These are the kinds of work where automation, and sometimes AI, can earn its place. In each, a person reviews the result.

  • Sorting and filing documents. Matching incoming certificates, delivery notes or job paperwork to the right record.
  • Extracting fields from forms or certificates. Reading names, dates and reference numbers so nobody retypes them.
  • Drafting summaries. Producing a first draft of a weekly summary across jobs, costs and compliance, for someone to correct and approve.
  • Reminders, which need no AI. Plain rules can prompt people about actions that are due, records that are missing and documents that are about to expire. They are listed because they are cheap to try.

Whichever example you choose, decide who will review the output and how quickly they can do it. A task that produces a pile of results nobody has time to check has simply moved the problem.

Where AI is a poor fit

Be cautious in these situations.

  • Decisions that call for judgement, such as whether to take on a job that stretches your team, or how to settle a dispute about what was agreed on site.
  • Safety decisions. A person must make and own them.
  • Anything you cannot check. If you have no practical way to tell whether the output is right, do not rely on it.
  • Occasional or one-off tasks. Automation takes effort to set up, and a task that seldom happens may never repay it.

A simple rule helps: people review AI output and approve the decisions that matter.

Questions to ask any supplier about data

Before you hand business information to any tool, ask the supplier to answer these in writing.

  1. What data leaves my business, and what exactly is sent?
  2. Where is it processed?
  3. What is stored, for how long, and who can see it?
  4. Is my data used to train or improve anyone else's product?
  5. Can I have my data deleted, and how do I confirm that it has been?

For anything involving personal data, check your obligations with your own adviser. We are not giving legal advice here.

Running costs, ownership and how to start

AI usage can add a running cost that continues after the build. Ask how it is measured, who sees the usage, and what happens if volume grows. Agree it before you start, not after the first bill.

Then start small. Choose one task, run a pilot with one team, and agree a single success measure in advance, for example the time the task takes before and after, measured the same way both times. If the pilot does not earn its place, stop. If it does, extend it one step at a time. Resist running several pilots at once, because you will learn less from each.

Say clearly who owns each automation in your business. Someone should be responsible for noticing when it stops working, when the input changes, or when the people reviewing it are no longer checking carefully.

This is how we approach it at SENRAB. Workflow automation is scoped per project. The things we typically build include approval workflows, document processing and data extraction, integrations with other systems, reporting and reminders. AI is used selectively, and people review its output and approve the decisions that matter.

If you are not sure which task to try first, a good place to look is the paperwork that holds up invoicing. Our article on the time between job completion and invoicing shows how to find it, and our piece on signs you have outgrown spreadsheets may help you spot other candidates.