We run an AI consulting team, so this might sound like an odd thing for me to write. But one of the most useful things we do in a pilot is tell a company where AI is the wrong tool.
If you sit on a board or run a business, you have probably heard a lot of pitches that start with the model and work backwards to the problem. I'd rather start with the problem. Sometimes the answer is AI. Quite often, for part of the task, it isn't.
When a rule will do
If the right answer can be written down as a clear rule, write the rule.
Checking that an invoice total matches the purchase order, routing a request by which country it came from, flagging a payment above a set amount for approval: these are rules. A rule gives the same answer every time, costs almost nothing to run, and anyone on your team can read it and see why it made a decision.
An AI model can do these things too, but it adds cost and a small chance of a wrong answer to something that never needed either. When we test a task, we often find one or two steps where a plain rule beats every model we try. We keep the rule and use AI only for the steps that actually need judgement, like reading a messy scan or understanding what an email is asking for.
When a better form will do
A lot of AI projects are really about cleaning up information that came in badly in the first place. Free-text emails instead of structured requests. Scanned paper instead of data. A form with ten optional fields that nobody fills in.
Before we build something to read all of that, it's worth asking a simpler question: can we change how the information arrives? A required field, a dropdown instead of a text box, or a supplier portal instead of an inbox can remove the problem at the source. That is cheaper to run and easier to trust than a model reading around it every day.
This isn't always possible. You can't redesign how your customers write to you, and some suppliers will keep sending PDFs whatever you ask. But where you control the input, fix the input first.
When the process is the problem
Sometimes a task is slow for reasons that have nothing to do with reading or writing. An approval sits with one person who is travelling. Two teams each think the other one owns a step. Work waits for a weekly meeting before anyone acts on it.
AI won't fix any of that. If we automate the slow part around a broken process, the work just piles up faster at the step that was actually stuck. In those cases the honest advice is to change the process: move the approval, name an owner, remove the step. It's less exciting than an AI project, and it usually works better.
When the volume is too small
If a task happens a few times a month, the cost of building, testing and looking after an AI system can be more than the time it saves. That's why the first thing we do in a pilot is measure how often the task happens and what each one costs today. If the money isn't there, we say so on day one.
When a mistake costs too much to risk
Some decisions carry consequences that a person needs to own: a legal position, a large payment, telling a customer something that can't be taken back. AI can still help here by gathering the facts, drafting a first version or pointing out what looks unusual. But the decision itself should stay with a person, and the system should be built so it can't act on its own.
How we say so in a pilot
During a pilot, we test each step of the task against real cases from your business. That makes these calls a matter of evidence, not opinion. For each step, the scorecard shows whether a rule, a model or a person does it best, and what each option costs.
So the readout can say things like: use a rule for this check, fix this form, move this approval, and use AI for these two steps where it clearly helps. Sometimes it says the task isn't worth automating at all, and you keep the numbers and the test set to make that decision with confidence.
We don't sell a model or a licence, so we have no reason to push AI into places it doesn't belong. The aim is a system your team relies on, and that usually means a mix of plain rules, better inputs and AI where it earns its place.
If there's a task you suspect is a good fit, or one you suspect isn't, start a chat with us on the site or book a call through the Applied Labs page. Bring a few real examples and we'll tell you honestly which parts AI should touch.
