Most AI projects should have been automation projects
The line that matters, and the mistakes that follow.
Most businesses asking “should we use AI?” should be asking “should we use automation?” — and the answer is usually different from what the vendor selling them AI would say.
This page is the operator’s version of that conversation. No vendor angle. No magic-wand framing. If you have a real operational problem and want to know which tool to reach for, this is for you.
The line that matters
Automation is deterministic. Same input, same output, every time. If/then logic, even when the logic is complicated. The intelligence is in the rules — you wrote them, you understand them, you can change them, and an auditor can read them.
AI is probabilistic. Same input, different output is possible. The intelligence is in a model trained on data. You can shape its behaviour with prompts, evaluation, and architecture, but you cannot guarantee a specific output. Run it a million times and you get a distribution, not a number.
That is the line. Everything else is implementation detail.
When the answer is automation
If you can write down the rules — even complicated rules — automation is the right answer. Not because it is cheaper (it usually is). Because it is the right tool for the job.
Examples where automation is the correct call:
- Generating an invoice from line items and a tax rate
- Pulling a CSV from a supplier and updating stock levels
- Routing a customer email by subject line, sender, or order ID
- Calculating the optimal cuts from a 6.5-metre profile to fulfil a 4-metre and a 1-metre order
- Sending an alert when a metric crosses a threshold
These have a right answer. They have an audit trail by default. Run them a million times and you get the same answer a million times — which is exactly what you want for anything financial, contractual, or operational.
If you use AI to do these, you have added a probabilistic step to a deterministic problem. The fancy demo will hide that for a while. Production will not.
When the answer is AI
If you cannot write down the rules — or the rules are too many, too fuzzy, or change too fast — AI starts to earn its place.
- Extracting structured data from unstructured documents (PDFs, emails, handwritten notes)
- Classifying or summarising free-form text the way a person would
- Generating human-quality language from a brief — drafts, explanations, replies
- Pattern-matching across complex inputs (images, audio, mixed media)
- Decisions where the right answer is a judgement, not a calculation
The test: would a human doing this work need experience and judgement, or would they just need a procedure manual? Procedure manual → automation. Experience and judgement → AI.
When the answer is both
Most real systems are hybrid. The pattern that works:
- AI handles the fuzzy part. Usually input parsing or judgement-heavy classification at the front of the workflow.
- Automation runs the spine. Rules, calculations, integrations, the actual operational flow.
- AI returns for the human-readable bits. Draft confirmations, summaries, explanations — where the output needs to read like a person wrote it.
Example: a supplier sends a PDF order. AI extracts the line items, quantities, and delivery date. Automation checks stock, calculates margin, routes the order to production, books it into the schedule. AI drafts the confirmation email. The deterministic spine is automation; AI is used where the inputs and outputs are messy.
The mistake one way is letting AI run the spine. The mistake the other way is refusing AI at the edges where it would save weeks of manual handling.
How to tell which you need
Reach for automation when:
- You can write the rules down in plain English
- The inputs are always structured (forms, databases, fixed formats)
- The same input has to produce the same output every time
Reach for AI when:
- The inputs are unstructured (free text, images, handwriting)
- The work requires human-like judgement
- The output needs to read like a human wrote it
Add a human checkpoint when:
- The cost of being wrong is high — whichever tool you’ve chosen
Where the prompt shops go wrong
They use AI for everything because AI is the only tool they have. A pure-AI implementation of, say, an invoicing workflow will:
- Produce slightly different invoices for the same inputs, because foundation models are non-deterministic by default
- Cost an order of magnitude more per transaction than the automation equivalent
- Need an evaluation harness, drift monitoring, and prompt versioning to prove it is doing the right thing
- Still fail on the first input that does not match the training distribution
This is the demo-that-falls-apart-in-production pattern. The most common failure mode in the field today.
Where the AI vendors go wrong
They sell you an “AI platform” and treat the surrounding automation, integration, and data work as your problem. You end up with a clever component sitting inside a fragmented system that nobody can audit and that has no LLM-agnostic upgrade path. When the model changes — and it will — you rebuild.
The platform did its job. Your operational reality did not.
The honest test
Before deciding which tool to use, ask:
- Could a junior employee with a procedure manual do this? If yes, automation will do it cheaper, faster, and more reliably.
- Does the work require a person to read context and make a call? If yes, AI can probably help — but a human likely needs to stay in the loop for any consequential decision.
- Could a wrong answer cost real money, real time, or real reputation? If yes, the system needs an audit trail and a review path — whichever tool you choose.
If you cannot answer all three with confidence, that is the conversation we should have.