Skip to content
Sudbury Consulting Group
← All posts
2 min read

Where AI actually earns its keep in a mid-sized business

AI is useful for some business work and a costly distraction for the rest. Here is a practical test for telling the difference before you spend anything.

Every software vendor now sells AI. Some of it is genuinely useful. A lot of it is ordinary automation with a new label, or a solution looking for a problem. For a mid-sized business, the question isn't whether to "do AI". It's which specific pieces of work it would do better, cheaper or faster than the alternative.

A practical test

AI tends to earn its place when the work is:

  • High-volume. Enough of it that small gains per item add up.
  • Judgment-heavy but repetitive. Work that needs reading and interpreting, not just copying, but where the same kinds of judgment come up again and again.
  • Measurable. You can say what a good result looks like, check it, and know the error rate.
  • Tolerant of review. A person can check the output before it matters, at least at first.

If a task fails most of those tests, ordinary automation, or simply removing the task, is usually the better answer. Plain rules are cheaper, more predictable and easier to maintain than a model.

Where it tends to work

Reading and sorting documents

Invoices, purchase orders, inspection forms, applications and emails arrive in endless formats. Pulling the key details out of them and routing them to the right place is tedious for people and well suited to AI, as long as uncertain cases are flagged for a person.

Summarising and drafting

First drafts of routine correspondence, summaries of long documents, notes from structured information. The person still decides what goes out; the time saved is in not starting from a blank page.

Answering questions from your own documents

Internal assistants that can find answers in your procedures, policies and past work can save time for teams who spend their days asking "where's the document for…". They're only as good as the documents behind them, which is often where the real work is.

Where it tends not to

  • Anything that needs to be exactly right every time with no review, such as final pricing or anything with safety or legal consequences.
  • Rare tasks. If it happens twice a year, the setup will never pay back.
  • Broken processes. AI on top of a messy process produces messy results faster. Fix the process first.

How to do it responsibly

Before building anything, measure the current cost, volume and error rate of the work. Without that baseline, nobody can say whether the AI version is better.

Then build in review and monitoring from the start rather than bolting them on later: track the error rate, keep a person in the loop where mistakes are costly, and know the cost per task. That's how AI becomes a dependable part of the operation instead of an experiment someone has to babysit.

We take a deliberately cautious approach in our AI implementation work: where the case can't be made, we say so and build something simpler. If you're weighing up where AI might fit in your business, get in touch. The first conversation is free.