Before You Automate It, Ask These Three Questions
The most expensive automation isn't the one that breaks. It's the one that works perfectly — on a task that never deserved automating.
We see it constantly: a business excited about AI picks the most interesting task to automate instead of the most costly one, spends real money, and ends up with a clever system nobody uses. So before we build anything for a client — or for ourselves — we run the same three questions. They kill about half of the ideas that reach us. That's the point.
Question 1: Does this happen often enough to matter?
Automation pays back per repetition. A task that eats 20 minutes but happens fifty times a month is a genuine gold mine — that's over 16 hours of the same work, done the same way, every month.
A task that eats a painful half-day but happens twice a year is a bad automation target no matter how annoying it is. Annoyance is not frequency. The half-day version will also have changed shape by the time it comes around again, which means your automation greets it like a stranger.
The quick test: write down the last five times the task actually happened, with dates. If you can't, it doesn't happen enough.
Question 2: Is the process stable, or does it change every time?
Automation is a bet that tomorrow's task looks like today's. Some processes honor that bet: every incoming CV needs the same fields extracted, every portal enquiry needs the same qualifying questions answered, every month's report pulls the same numbers.
Others don't. If every instance needs a judgment call about what the process even is — special-case clients, one-off negotiations, "it depends" at every step — you don't have an automation problem. You have a process-definition problem, and no amount of AI will automate a process you can't describe.
There's a middle case, and it's where modern AI actually earns its keep: the process is stable but the inputs are messy. Reading CVs is like this — every CV is worded differently, but what you're extracting from them never changes. Old-style automation choked here; language models thrive here. That's a genuine change in what's automatable, and it's why this question no longer kills as many ideas as it used to.
The quick test: could you write the steps down clearly enough that a smart new hire could do the task right on day one? If yes, it's automatable. If the honest answer is "they'd need to sit with me for a month," define the process first.
Question 3: What happens when it's wrong?
Everything fails sometimes — people and machines both. The question isn't whether your automation will produce a wrong output. It's whether a wrong output gets caught or gets shipped.
This is the question that decides the shape of the automation, not whether to build it:
- Cheap, visible mistakes → automate fully. A misfiled document gets noticed and fixed in a minute; let the machine run.
- Expensive or invisible mistakes → automate the work, gate the decision. This is why our own outreach system drafts every email but a human approves every send, and why Nexu AI ranks candidates but recruiters make the calls. The machine does the hours; a person spends seconds deciding.
- Catastrophic mistakes (money leaves, legal exposure, a customer relationship dies) → keep a human on the trigger, full stop, and let automation only prepare the action.
If a vendor pitches you full automation on a task whose failures are expensive and quiet, walk away. That's not confidence in their AI — it's indifference to your downside.
Run your own list through this
Take the three tasks that annoy your team most. Score each: frequent? stable? cheap-to-catch failures? The task that clears all three questions is your first automation — and it's usually not the flashy one. It's the boring, repetitive reading-and-typing work that nobody wanted to defend in the first place.
That boring task is where the payback lives.
NodalNexus builds AI systems and the automation behind them — starting with the workflow audit, not the code. If you want a second pair of eyes on your list, start a conversation.