We get paid to build automation, so treat the following with appropriate suspicion. But we would rather turn down an engagement than deliver one that does not pay back, because the second one costs us a reference and costs you a budget.
Here are the four situations where we tell people not to do it.
1. The volume is not there
This is the most common one by a distance.
Automation has a fixed build cost and a low marginal cost. Manual work has no build cost and a high marginal cost. The crossover depends entirely on volume, and people consistently overestimate theirs — often by counting the annoying process rather than measuring it.
Rough arithmetic: a task taking 4 minutes, at an Ontario fully-loaded clerk rate of about $36–40/hour, costs roughly $2.50 per item. At 500 items a month that is ~$1,200/month of labour. Against a build in the tens of thousands, payback runs years — and the process will probably change before you get there.
At 5,000 items a month the same maths gives ~$12,000/month, and the conversation is completely different. Measure first. If the honest number is a few hundred items a month, a $30/month tool or a better spreadsheet is the right answer and we will say so.
2. The upstream process is unstable
Automation encodes rules. If the rules change monthly, you are not buying an asset — you are buying a maintenance obligation.
Signs to watch for: the process was redesigned twice in the last year; two people do it differently and both are "correct"; nobody can produce a written description that survives contact with a third person; the upstream system is being replaced next quarter.
Automating an unstable process locks in a snapshot of a moving target. Stabilise and document it first — often that exercise alone recovers a meaningful chunk of the time you were trying to automate away.
3. The exceptions outnumber the happy path
Automation pays when there is a dominant pattern and a manageable tail. It fails when every case is a special case.
A useful test: take 50 recent items and sort them by how they were handled. If you get a clean majority handled identically, that majority is automatable and the tail goes to a review queue. If you get 50 slightly different stories, what you have is not a process — it is judgement work wearing a process costume, and an agent will produce a review queue that costs more than the original task.
The exception queue is the hidden cost centre. At 80% accuracy on 5,000 items, review labour runs roughly $2,600/month — several times the model API bill. Automation that produces a big queue has not removed the work; it has added a machine in front of it.
4. Being wrong is expensive and the work is unverifiable
Some decisions carry consequences that dwarf the labour they consume: paying an invoice, denying a claim, releasing personal health information, sending anything under CASL, or anything affecting a person's entitlements.
These can still be automated — but only with a human approval gate before the irreversible step, and that gate eats much of the time saving. If a human must review every output anyway, be honest that you are buying a drafting assistant, not an automation. Sometimes that is genuinely worth it. Sometimes it is not, and it should be priced as what it is.
The unverifiable case is worse: if nobody can tell whether an output was right without redoing the work, you cannot build a held-out evaluation set, which means you cannot measure accuracy, which means you cannot know your operating cost. We do not take those engagements.
What we do instead
When one of these applies, the useful outcomes are usually smaller and cheaper than automation: stabilise and document the process; fix the upstream data-entry step that is generating the exceptions; buy an off-the-shelf tool; or automate one narrow slice with high volume and low consequence, and leave the rest alone.
That is most of what a readiness assessment produces — a ranked list with payback periods, and an honest recommendation that sometimes reads "not yet." Two weeks, fixed price, and half of it credits against a build if you go ahead.
If a vendor has never told you not to automate something, you are talking to a salesperson. Volume, stability, exception ratio, and consequence — check those four before anyone quotes you a number.