Where does AI genuinely help?

  • Sorting and routing enquiries: reading a message and sending it to the right person.
  • Drafting routine replies: a first version for a person to review and send.
  • Summarising: long documents, call notes or customer feedback.
  • Forecasting: estimating demand, workload or cash needs from past data.
  • Spotting oddities: flagging unusual invoices, orders or expenses for review.
  • Extracting data: pulling details from forms, bills or emails into a system.

Notice the pattern. Each job is narrow, repeated often and easy to check.

Where does it fall short?

  • Decisions with serious consequences and no human check.
  • Anything needing deep knowledge of your specific situation that the tool has not been given.
  • Tasks where one wrong answer is costly, such as legal, medical or financial advice.
  • Messy processes. AI cannot rescue a workflow nobody understands.

What do you need in place first?

  • A clear task: one job, described in a sentence.
  • Usable data: past examples, in a form the tool can read.
  • An owner: a person responsible for checking the results.
  • A way to measure: what better looks like, such as time saved or fewer errors.

How do you start small?

Pick one task that takes time and annoys your team. Run the AI version alongside the manual one for a few weeks. Compare speed and quality, and note where it gets things wrong. Expand only when you trust it on that task. This is slower than a big launch, and far safer.

Why does human review matter?

AI tools can sound sure of themselves and still be wrong. A person reviewing drafts, flags or forecasts catches mistakes before customers do. Build the review step into the workflow, and make it easy to correct the tool so it improves.

What risks should you plan for?

  • Accuracy: test on your own examples, not only on the vendor's demo.
  • Privacy: know what data leaves your systems and who can see it.
  • Lock-in: keep your data portable so you can change tools.
  • Over-trust: staff may stop checking if the tool is usually right.

How should you think about cost?

Compare the cost of the tool and its set-up with the hours or errors it removes. A tool that saves a few hours a week for a small team may pay back quickly. One that needs months of set-up for a minor gain does not.

Which three small projects are safe places to start?

  1. Enquiry sorting. Messages are read and tagged, then sent to the right person. A human still replies.
  2. Draft replies. The tool suggests a first answer to common questions, and staff edit before sending.
  3. Document extraction. Details from bills, forms or emails are pulled into a sheet or system, with unclear items flagged for a person.

All three keep a person in the loop and have low cost if the tool makes a mistake.

How do you judge a vendor's AI claims?

  • Ask for a demonstration using your own examples, not theirs.
  • Ask how often it is wrong, and how errors are caught.
  • Ask what data is stored, where, and who can see it.
  • Ask how you can override or correct the output.
  • Ask what it costs when volumes grow.

How do you keep your team on board?

Be direct about what the tool is for. If people hear that a tool will replace them, they will not help it succeed. Frame it as removing the dull parts of the job, involve staff in testing and let them flag what is wrong. Their corrections are the best training material you will get.

When is AI not the answer?

If a simple rule or a basic automation does the job, use that. Plain automation is cheaper, more predictable and easier to explain. Reach for AI when the task involves understanding language, spotting patterns or predicting from data.

Related: AI and workflow automation · SMEs and Enterprise · All articles