When AI is not the answer
I build with AI every day, and a fair share of my advice is still 'do not use AI for this'. Here is how I tell when a better process, a spreadsheet, a standard feature, a simple rule or a conversation will beat it.
By Gary Kellett6 min read
For owners and directors under pressure to 'do something with AI' who want an honest view of when not to.
I build software with AI every day. It has changed how fast I can deliver, and I use it inside client systems where it earns its place. So it may seem odd that a good part of my advice to directors is 'do not use AI for this'. It is not a contradiction. It is the job.
AI is very good at a narrow set of things: reading messy text, drafting, summarising, classifying and spotting patterns in information that does not fit neatly into boxes. It is not good at being exactly right every time, explaining precisely why it did something, or working without the data it needs. Most business problems do not need what AI is good at. They need the boring things done properly.
When a better process wins
The most common case I see is a broken process wearing an AI costume. A team spends hours every week checking, chasing and correcting, and someone suggests AI could do the checking. Often the checking only exists because an earlier step is unreliable: an order form with free-text fields, a hand-off with no clear owner, two people keeping two versions of the same list.
Fix the earlier step and the check disappears. No licence, no build, no running cost. Automating a bad process just makes the mess happen faster, and now it is harder to see.
When a spreadsheet or a standard feature is enough
Plenty of jobs that get pitched as AI projects are really lookups, totals and reminders. A well-built spreadsheet with a few formulas and some data validation will do them for years. That is not a failure of ambition. It is the right size of tool.
Before that, check what your existing systems already do. Accounts packages chase overdue invoices. CRMs send follow-up reminders and score leads by simple rules. ERPs match purchase orders to invoices. These features are often paid for and switched off because nobody had time to set them up. Configuring what you own is almost always cheaper than adding something new.
When a simple rule beats a clever model
If you can write the logic down as 'if this, then that', you probably want a rules-based automation, not AI. Rules give the same answer every time, they are easy to test, and when something goes wrong you can see exactly which rule fired. AI gives a likely answer, which is the right trade-off for reading a scanned delivery note and the wrong one for deciding which bank account a payment goes to.
- Moving data between two systems when the fields line up: a rules-based integration.
- Sending a reminder when a date passes: a scheduled rule.
- Routing an enquiry by product or postcode: a lookup table.
- Reading a supplier's PDF invoice that arrives in twenty different layouts: here AI genuinely helps, with a person checking the exceptions.
Worked example
A wholesaler wants AI to decide which customers get a credit hold. When we write the policy down, it turns out to be three rules: over the limit, more than 60 days overdue, or flagged by the credit controller. That is a rule in the accounts system, configured in an afternoon. AI would add cost, add doubt about why a customer was held, and make the decision harder to defend.
When the data is not there
AI works with the information it can reach. If your job records are on paper, your pricing logic lives in one person's head, or your customer history is split across three inboxes, no model will make up the gap. It will produce something that looks confident and is built on guesswork.
In these cases the first project is getting the information into a system in a consistent shape. That work is less exciting and far more valuable. It also makes AI a realistic option later, if you still want it.
When a person must decide
Some decisions should stay with a person because of what they mean for someone else: who gets hired, who gets credit, what an individual customer is charged, whether a complaint is upheld. UK data protection law has specific rules about decisions made solely by automated means that have legal or similarly significant effects on people, and the governance cost of getting this wrong is high.
AI can still help here, for example by summarising a file so the person deciding has it in front of them. But the decision, and the accountability for it, stays with a named person.
When the work is regulated
In pharma, healthcare, financial services and other regulated settings, the questions are 'can you prove what happened?' and 'will it do the same thing next time?'. A validated workflow system with a clear audit trail usually answers those better than a model whose output can vary. AI can still sit inside that system in narrow, checked places, but it should not be the system.
When a conversation is the answer
Sometimes the problem is not technical at all. Two teams disagree about who owns a step. A customer keeps sending orders in a format nobody can use. A supplier has never been asked to change their invoice layout. One phone call, or one meeting with the right people in it, can remove a problem that would otherwise justify a software project.
A quick test to tell
Before anyone spends money, I run through five questions. Answer them honestly for the job in front of you.
Step 1: Can you write the rules down?
If yes, start with a rules-based automation or a feature in a system you already own.
Step 2: Is the process working as intended?
If no, fix the process first. Map it, remove the rework, then look again.
Step 3: Is the data in a system, in a consistent shape?
If no, the first project is the data, not the AI.
Step 4: Would a wrong answer harm a person or break a rule?
If yes, keep a person making the decision, and use AI only to prepare information for them.
Step 5: Is the input messy text, documents or images that vary every time?
If yes, and the other answers are clear, AI may well be the right tool. Prove it on a small, measured job before you widen it.
This is not an argument against AI
When the job suits it, AI can take hours out of a working week, and I build those systems for clients. The point is to reach for it on purpose, not by default. The best AI projects I have seen sit on top of a sound process, clean data and a clear owner.
If you are weighing up a specific problem and are not sure whether it needs a process fix, some configuration, a connection between systems, a product or a build, the build or buy check walks through those options in order and tells you where to start.
Questions directors ask
When should a small business not use AI?
When the steps can be written down as fixed rules, when the process itself is broken, when the data is not in a system, when a decision has a legal or significant effect on a person, or in regulated work where an exact audit trail matters more than speed. In those cases a process fix, a standard feature or a rules-based automation is usually cheaper and more reliable.
What is the difference between AI and a rules-based automation?
A rules-based automation follows logic you write down, such as 'if the invoice is over 60 days old, send a reminder', and gives the same answer every time. AI produces a likely answer from patterns in data, which suits messy inputs like varied documents but is less predictable and harder to explain.
Can AI still help with decisions that a person must make?
Yes, by preparing information, for example summarising a file or highlighting points to check. The decision, and the accountability for it, should stay with a named person.