The right question
A process might be fully automatable, yet cost more to build and maintain than the time it saves. Or it might be sensitive enough that removing the human creates more risk than benefit. So the right question isn’t can we automate this, but should we automate this.
Before choosing a tool, model or agent, put the current process on paper: what goes in, who receives it, what decision is made, what comes out, how often it repeats, and what an error would cost. Sometimes this alone shows the problem needs no AI at all - a better form or a simple integration is enough.
Four situations to pause before automating
1. When the process is rare
If something happens once or twice a month, a dedicated system may not be justified; designing, building, testing and maintaining it can take longer than just doing it by hand. Ask: how many times does this really happen in a year?
2. When the decision depends on human judgement
Experience, negotiation, reading a situation, or professional responsibility play a role in some work. AI might summarise a customer's information, but the final call in a sensitive negotiation should not be handed to a model. There, AI is better as an assistant than a decision-maker.
3. When errors are expensive
The higher the cost of a mistake, the more carefully the level of automation must be chosen. A misfiled email may not matter; a wrong financial or security decision can be costly. It must be clear what AI does, what it only suggests, where human approval is required, and whether an error is reversible.
4. When the process isn’t yet well-defined
If every person does a task differently, you’re probably not ready to automate it. Adding AI to an undefined process doesn’t turn it into a good system; first define the right way to do the work, then decide what part to hand to a system.
Run the economics before building
Convert current time into monthly hours, estimate its value to the business, then set the build cost, maintenance cost and the cost of current errors beside it to find the break-even point. Without this calculation, even a technically successful project can still be a bad business decision.
AI isn’t always the answer
Sometimes the best solution is a simple automation, a small script, an integration between two tools, or an automatic report. Only in some cases does a language model or agent genuinely add value. The goal isn’t using AI; it’s improving the process.
What usually makes a good candidate
- High repetition: work done many times daily or weekly
- Defined rules: steps that can be standardised
- High information volume: reviewing many documents, messages or records
- Response time matters: delay here costs an opportunity
- Measurable output: before and after can be compared with a clear metric
A simple decision framework
- How often does this happen?
- How much time or money does it cost today?
- If automated, what exactly gets better?
- How much does building and maintaining it cost?
- Which part should still stay under human control?
Sometimes the best decision is: don’t build it
If a project saves little, runs rarely, carries high risk, the process is still undefined, or a simpler fix solves it, building a dedicated system is probably the wrong call. A good consultant doesn’t just tell you what to build; they can tell you what not to.