What AI Can Actually Do for Your Business
Last updated:
The distinction that predicts success
There is one question that predicts whether an AI feature will work in production: does a wrong answer get caught before it does damage?
AI is probabilistic. It will be wrong sometimes, and no amount of prompt engineering removes that entirely. Features where a human reviews the output, or where a wrong answer is merely a slower path to the right one, work well. Features where a wrong answer goes straight to a customer or a ledger need a great deal more engineering than the demo suggests.
Five things it does reliably
- Reading unstructured text. Pulling structured fields out of emails, invoices, CVs, contracts. This is the highest-value, lowest-risk category and where most of our AI work sits.
- Classifying and routing. Which department, which priority, which category. Consistently better than keyword rules and dramatically cheaper to maintain.
- Drafting for a human to finish. Replies, descriptions, summaries. The human edit is the quality gate, and the time saving is real.
- Summarising long material. Call transcripts, threads, documents. Reliable when the source is provided rather than recalled.
- Searching by meaning. Finding the relevant policy or ticket when nobody knows the exact words used.
Three that usually disappoint
- Anything needing a guaranteed answer. Prices, legal positions, medical guidance, compliance determinations. Right 96% of the time is excellent for a draft and unacceptable for an invoice.
- Arithmetic and aggregation. Language models are poor calculators. Compute in code and let the model explain the result.
- Facts it was not given. Your stock level, your policy, today's price. If the model was not handed the data it will produce something plausible, which is worse than an error.
The failure mode that costs businesses money is not the model refusing. It is the model answering confidently from nothing.
What this looks like as a real project
A distributor receives 200 purchase orders a day as PDFs and emails, in dozens of formats. Someone types them into the order system. Extraction gets the fields, confidence scores route the uncertain ones to a human, and everything above the threshold posts automatically.
That project halves a person's day and cannot invent an order, because it only ever reads a document that exists. It is unglamorous and it is where the money is.
The rule we apply before quoting
We will build an AI feature where at least one of these is true: a human reviews the output before it matters, a wrong answer is cheap and obvious, or the output is checked against a deterministic system before use.
If none of the three holds, we say so. There is an enormous difference between a demo that impresses a boardroom and a system that runs unattended on a Tuesday afternoon, and the gap between them is almost entirely error handling.
Where to start
Find the most boring text-handling task in the business — the one where somebody reads something and types what it says into a system. It is high volume, tolerant of a review step and easy to measure.
- Invoices and purchase orders arriving as email or PDF
- Enquiries needing categorising and routing
- CVs, application forms, claim forms
- Meeting or call notes needing structuring
- Support tickets needing a first-draft reply
Frequently asked questions
Do we need our own data to start?
Is our data used to train someone else's model?
How accurate is accurate enough?
Will this replace staff?
Wondering whether AI would actually help here?
Describe the task and we will tell you plainly whether it is a good fit, including when the honest answer is that ordinary software would serve you better.