AI is useful in business systems far more often than it’s used well. The gap isn’t the technology — it’s the discipline around it. Added to the right task with the right guardrails, AI removes genuine drudgery. Added because it’s fashionable, it introduces a component that’s confident, occasionally wrong, and hard to debug. The difference comes down to where you put it and how you contain it.
Where AI tends to help
AI earns its place on tasks that involve messy, language-shaped information and tolerate a human check:
- Reading and sorting documents. Pulling key fields from invoices, forms or emails that arrive in a hundred slightly different formats.
- Retrieving knowledge. Answering “where’s the policy on this?” from a pile of internal documents, with links back to the source so a person can verify.
- Drafting, not deciding. First-draft replies, summaries or categorisations that a person reviews before anything is sent or committed.
- Triage. Routing or prioritising incoming work so humans spend attention where it matters.
Notice the pattern: in each case AI reduces effort, and a person stays accountable for the result.
Where it adds risk
AI is a poor fit where correctness must be guaranteed and can’t be checked, where the cost of a confident mistake is high, or where a clear deterministic rule would do the job better. Calculating a payment, enforcing a legal threshold, making an irreversible decision unsupervised — these want logic you can read and test, not a probabilistic guess. If you can write the rule, write the rule.
Use AI for the judgement-shaped, language-heavy parts of a task. Use ordinary code for the parts that must be exactly right every time.
The guardrails that make it dependable
Where AI does belong, a few practices keep it trustworthy:
- Keep a human in the loop for anything consequential, with the AI proposing and the person approving.
- Show the source. An answer you can trace back to a document is useful; an unverifiable claim is a liability.
- Design for being wrong. Assume it will occasionally err and make sure that’s caught cheaply, not discovered in front of a customer.
- Be honest about data. Know what information the model sees, where it goes, and whether that’s acceptable for your obligations.
- Watch the running cost. AI features have ongoing per-use cost and latency; both belong in the decision, not just the demo.
A grounded way to decide
Before adding AI to a system, it’s worth answering: what specific task is it doing, what happens when it’s wrong, who checks it, and would a plain rule do the job more reliably? If those answers are solid, AI can quietly remove real work. If they’re not, the most valuable thing we can do is say so — and that’s the same standard we hold our own recommendations to.