The four tests a first workflow has to pass.

Why the choice matters more than the build

An AI system is a fixed cost applied to a repeating process. If the process repeats rarely, the system is expensive. If the process has no rules, the system is unreliable. If nobody agreed what success looks like, the system is unfalsifiable: it can neither succeed nor be turned off.

Every one of those failures is decided before a line of code is written. That is why Nexvix runs a workflow audit first and prices the build second, and why the audit output is a map rather than a proposal.

The build is rarely the risk. The choice is.

Test one: does it happen often enough?

Count the instances per week, honestly, for the last month. Not the peak week, the ordinary one. A workflow running twenty times a week and taking ten minutes each is three and a half hours weekly, which is a real number to hold a system to.

Below roughly five instances a week, automation is usually the wrong instrument. A checklist, a template or a setting in software you already pay for will get there faster and cost nothing to maintain.

The exception is consequence: a workflow that runs twice a month but loses a customer when it is missed can justify a system on risk alone. Say that out loud during scoping, because it changes what you are buying.

Test two: can the rules be written down?

Ask the person who does the work to describe the decision. If they can state it as conditions (if this, then that; if unclear, ask Mo), the workflow is automatable. If the answer is “you get a feel for it”, you have found undocumented expertise, not an automation candidate.

That is not a dead end. It is a two-stage project: write the rules down first with the expert, then automate the version everybody agreed. Skipping the first stage is how a system ends up encoding one person’s habits and nobody noticing for six months.

• Conditions a person can state in a sentence
• An agreed answer for the unclear case
• A named owner for exceptions
• Something that happens when the system is unsure

Test three: is there a number that will move?

Pick one measurable outcome before pricing: replies within the hour, hours returned per week, leads not lost overnight, records entered once. It has to be a number that already exists or can be counted from next Monday.

“Efficiency”, “modernisation” and “AI adoption” are not numbers. A project with no number cannot be verified, so nobody can say whether the money worked, and the system quietly stays in place because removing it is also unmeasurable.

Nexvix writes the number into the scope document at stage two, and stage five compares against it. If it misses, the fix is ours.

Test four: what happens when it gets one wrong?

Every system gets some instances wrong. The question is what that costs. A misclassified enquiry costs a follow-up. A wrongly issued refund costs money. A wrong clinical answer is not recoverable at all.

Sort candidate workflows by the cost of a mistake, and start with the cheapest failure that still has volume. The high-consequence work is not off limits; it just has approval points designed in from the start, which is more build and more review.

Start where a mistake costs a follow-up, not a customer.

Ranking your three candidates

Score each candidate one to five on the four tests and add them up. The winner is usually obvious, and it is frequently not the one the leadership team named at the start. That one tends to be the most visible process rather than the most expensive one.

Take the ranked list to the audit. Half an hour of arguing about the scores with the people who do the work is the cheapest project insurance available.