Business
Use AI where it actually makes sense.
AI is very good at a narrow set of tasks and unreliable at everything else. Knowing the difference is most of the value of hiring someone for it.
At a glance
- Applied to specific, bounded tasks
- A human reviews anything that matters
- Clear rules on what data is sent where
- Measured against the manual process it replaces
The useful question is narrow, not visionary.
Most AI proposals aimed at small businesses describe a transformation rather than a task. The projects that actually pay for themselves are unglamorous: reading a supplier invoice and pulling out six fields, sorting an enquiry into the right category, drafting a first reply someone then edits, or finding the answer to a question buried in three hundred pages of internal procedures.
Those work because they are bounded, verifiable and sit inside a workflow where a person still has the final say. Anything where a confident wrong answer would cause real damage either keeps a human in the loop or does not get built.
Good candidates
- Data typed in from PDFs, forms or emails
- Enquiries sorted by hand into categories
- The same first reply written repeatedly
- Staff searching documents for a known answer
- Long threads that need summarising
- Notes transcribed after every job or call
What it covers
Where AI genuinely earns its place
- Document and form extraction
- Pulling structured fields out of invoices, applications, timesheets and PDFs, then writing them into the system that needs them.
- Enquiry classification
- Sorting incoming messages by service, urgency and location so routing and response times improve without a person triaging first.
- Drafted responses
- A first draft prepared from your own previous replies for staff to review and send — faster, and consistent in tone.
- Search across your own material
- Asking a question of your procedures, manuals and past quotes, and getting an answer with a reference to the source document.
- Summarisation
- Long email threads, call notes and meeting transcripts condensed into the decisions and actions that came out of them.
- Guardrails
- Confidence thresholds, human review steps and fallbacks, so an uncertain result escalates to a person instead of proceeding quietly.
Where your data goes
Before anything is built, it is agreed which data leaves your systems, which provider processes it, whether it can be used for training, and what stays entirely internal. For sensitive material, options that keep processing private are assessed rather than assumed away.
Measured against the status quo
An AI step has to beat the manual process on accuracy, time or cost to justify existing. If it does not, it gets removed. That comparison is made deliberately rather than taken on faith.
What will not be recommended
AI making decisions with legal, financial or safety consequences without review. AI-generated content published at volume for search. Chatbots that answer customers with confident invented detail.
Is our business big enough for AI to be worth it?
It depends on repetition, not size. If someone spends several hours a week doing the same language or document task, there is likely a case. If the task happens twice a month, there almost certainly is not.
What about accuracy?
Accuracy is treated as a design constraint. Extraction is validated against rules you define, uncertain results are escalated to a person, and anything consequential keeps a review step. Nothing is built on the assumption that the model is always right.
Will our data be used to train someone's model?
That depends on the provider and the plan, and it is settled before anything is built rather than discovered later. Where data sensitivity requires it, configurations that keep material out of training are used.
Do we need AI, or just automation?
Very often just automation. If the rules can be written down, ordinary automation is cheaper, faster and more predictable. AI earns its place specifically where the input is messy, unstructured language or documents.
Related
Often needed alongside this
Wondering whether AI applies to your business?
Describe the task you have in mind. You will get a straight answer, including if the answer is that plain automation would do it better.