AI
Where AI actually helps a small business, and where it does not
A practical filter for AI ideas: the jobs language models are reliably good at, the ones they are not, and how to start small.
Codexan
6 min read
Most businesses we speak to have been told they need an AI strategy. Very few need one. What they need is a short list of tedious jobs that a language model can take off somebody's desk. Here is how to find that list.
What language models are good at
Modern models are dependable at work that involves reading and rewriting text:
- Extracting. Pulling names, dates, amounts and reference numbers out of emails, PDFs and forms.
- Classifying. Deciding which team an enquiry belongs to, or whether a message is urgent.
- Summarising. Turning a long thread or document into the five lines someone needs.
- Drafting. Producing a first version of a reply, a description or a report for a person to check.
- Answering from documents. Finding the relevant passage in your own manuals, policies or knowledge base and answering in plain language.
Notice what these share: the input is messy text, and a person can quickly check the output.
What they are not good at
- Arithmetic and exact records. Totals, stock levels and balances belong in a database. Let the model call a calculation, never perform it.
- Decisions with serious consequences. Approving a refund, rejecting an application or changing a medical or legal record needs a person, with the model assisting.
- Anything it has not been shown. A model does not know your prices, your policies or last week's change unless you give it that information each time.
A filter for ideas
Take any proposed use and ask four questions.
- Is the task done often? Automating something that happens twice a month rarely pays back.
- Is a wrong answer cheap to catch? If a person reviews the output, or a mistake is easily reversed, the risk is low.
- Is the information available? The model needs the relevant documents or data at the moment it answers.
- Can you measure it? Time saved per item, or the share handled without correction.
Ideas that pass all four are worth a trial. Ideas that fail the second are worth redesigning so that a person stays in the loop.
Start with a narrow trial
Pick one job and a few weeks of real examples. Build the smallest thing that handles it, and compare its output with what your team actually did. You will learn more from fifty real cases than from any demonstration.
Keep these in from the first day:
- A log of what went in and what came out.
- A way to say "I am not sure" and pass the item to a person.
- A spending limit, because usage-based pricing rewards attention.
Privacy is a design decision
Before any customer data goes to a model, know where it is processed, whether the provider keeps it, and whether it is used for training. Business offerings from the major providers generally do not train on your data, but confirm it in the terms you are signing, and send only the fields the task needs.
The honest summary
AI is at its best as a fast, tireless assistant doing the first pass of text-heavy work, with a person checking the result. Used that way it saves real hours. Asked to replace judgement, it creates new problems faster than it solves old ones.
- ai
- llm
- automation
- strategy