When AI is the wrong answer

Failure mode 1 — the process is not documented

Every workable automation starts with a written procedure. Not a diagram, not a slide — a procedure. Five steps, clear input at the top, clear output at the bottom, and a decision rule for the edge cases. If you cannot produce that document in an afternoon, the process is not ready for AI. It is ready for a whiteboard.

A Sydney mortgage broker came to us wanting to automate his client review meetings. On paper it made sense — he was doing thirty a month and each one took ninety minutes to prepare. We asked him to write the five steps he took to prepare a review. He could not. Every client got a slightly different treatment depending on their situation, the market that week, and what he had read that morning. It was expertise, and it was valuable — but it was not a process. There was nothing to automate.

We sent him away for six weeks to document what he actually did, then built the automation in two. The build was fast because the thinking was done. That is the pattern. AI needs defined input and defined output. Ambiguity at any point in the workflow becomes a failure the software cannot recover from.

Failure mode 2 — it is a one-off task

Automation carries setup cost. A small workflow built properly — with error handling, a place for humans to intervene, and a way to monitor whether it is still working — costs somewhere between $5,000 and $20,000 for most SMB use cases. Ongoing tool costs and maintenance add ten to twenty per cent a year on top.

The break-even test is simple:

Runs to break even = Setup cost ÷ (Time per run × Hourly cost of the person doing it)

A $10,000 build that saves two hours per run at $80 per hour needs 62 runs to pay back. That is fine if the task happens weekly. It is not fine if the task happens twice a year — because you are looking at thirty years to break even and the tools you built it on will not exist in thirty years.

The rule of thumb we use with clients: fifty runs minimum in the first year to justify a proper build. Below that, use a template, batch the work manually, or hire a casual for the two days a year it actually takes.

Failure mode 3 — the relationship is the service

Some businesses sell process. Some businesses sell a person. When the person is what the client is buying, automating contact points removes the thing they are paying for.

A Sydney executive coach charged $12,000 for a six-month engagement. She wanted to automate her intake — thirty minutes of scheduling emails and a Typeform. The numbers worked: four hours saved per month. Then we asked what her clients said about their first contact with her. Every one of them mentioned the intake call. It was where they decided she was the right coach. Automating it would have removed the exact signal that closed the sale.

The distinction is between administrative contact and value-generating contact. A booking confirmation is administrative. A discovery call is value-generating. Invoicing is administrative. A quarterly review is value-generating. Reminders are administrative. The apology when something goes wrong is value-generating.

Automate the administrative layer aggressively — it frees time and clients rarely notice. Touch the value-generating layer carefully. If a client is paying for you, make sure they get you at the moments they are paying for.

Failure mode 4 — compliance decisions require human judgement

In Australia there are decisions that a regulator, an insurer, or a court expects a qualified person to make. The list is longer than most business owners realise. It includes legal advice under Legal Profession Uniform Law, medical decisions under AHPRA, credit decisions under the National Consumer Credit Protection Act, tax advice under the Tax Practitioners Board, financial advice under ASIC, and any employment decision touching the Fair Work Act — from performance management through to termination.

AI can help in all of these areas. It can draft the letter, summarise the case law, flag the anomaly, prepare the shortlist. What it cannot do is sign. The signature carries legal responsibility. Regulators want to see that a qualified human read the output, considered the specific circumstances, and made the call.

A common failure we see: small businesses using AI to write termination letters or warning notices without an HR review. The letters read well. The decisions underneath are often defective. The Fair Work Commission does not care how well the letter reads — it cares whether the process was followed. Use AI to draft, prepare, and accelerate. Keep a qualified human in the sign-off seat.

The test before any automation project

Before any automation project, ask two questions.

One. Can you write a clear five-step procedure for this work today, covering ninety per cent of the cases and naming the edge cases explicitly?

Two. Does the human touch at this specific step create value the client is paying for?

Yes to question one, no to question two — automate. That is the green light. Any other combination and the answer is: not yet. If the process is not documented, go document it. If the human touch matters, keep the human there and automate around them, not through them.

What to do instead

If AI is the wrong answer, there is almost always a right one.

If the process is not documented: document it. This is a two to six week job done by the person who currently owns the work, sitting with someone who asks basic questions. Do not skip to software. The document is the deliverable that unlocks everything downstream — whether that is AI, delegation, or hiring.

If the work is a one-off: use a template, a checklist, or a project you outsource to a specialist for two days. It is not a system.

If the relationship is the service: free your senior people from admin so they can spend more time on the relationship layer, not less. Automate the calendar so the coach can run more sessions — do not automate the coaching.

If compliance is the gate: use AI to prepare the pack, route it to the qualified person, and log the review. That shape scales, and it stays defensible.

For the situations where automation IS the right answer, read how to identify which process to automate first and the unit economics of an AI employee for how to size the business case before you build.