The five questions every AI audit answers

Key takeaways

• Every audit answers five questions, in order: where is the time going, where is the money going, what is fragile, what is repeatable, and what is the next dollar worth automating.
• The order matters. Most operators want to start at question five, and starting there is how AI engagements waste their first month.
• The deliverables, an operations map, an opportunity matrix, and a 12-month roadmap, belong to the client whether they hire us for the build or not.

The frame

The two-week audit we run at the start of every engagement, the first phase of [the audit-build-operate model](/blog/audit-build-operate-model), is built around five questions. They look obvious written down. In practice almost nobody has the answers at the resolution that matters, because the answers require evidence most operators have never had a reason to collect.

Here they are, in the order we ask them.

1. Where is the time going?

Not "what does your team do." That answer is always a role description: one person handles tenant comms, another handles maintenance, the owner handles the owners. What we're after is closer to a calendar. In the last week, how many hours actually went to lease renewals? To inquiry triage? To the recurring escalation that has been bouncing around the office for a month?

You don't get this by asking. You get it by sitting next to the team and watching, counting, timing things. It looks slow, and it's the cheapest time in the whole engagement, because every question after this one gets evaluated against what it finds.

2. Where is the money going?

This sounds like a finance question. It's mostly a question about leverage.

The most expensive hours in a small business usually belong to the most constrained person, and that person is usually spending a large slice of their week on work that wouldn't need them at all if it were shaped differently. The owner assembling reports a coordinator could assemble. The senior person retyping what a system should have captured. The cash cost of this is invisible, because nobody invoices for it. The opportunity cost is the biggest number in the audit.

So the money question really asks: whose hours are worth the most, and what are those hours being spent on?

3. What is fragile?

Every operation has workflows that work only because one specific person remembers how to do them. No document, no system. Someone has been doing it for three years and holds the whole thing in their head.

The audit names these, bluntly, and puts them on a list, because fragile workflows are where automation goes wrong. If a process exists only as one person's memory, you can't build a reliable system around it until that memory has been written down. A model can't infer it, and an integration can't compensate for it. Automate a fragile workflow without documenting it first and you ship something that looks right for a month, then breaks on the first case that only ever lived in someone's head.

So the fragile list has two uses. It tells us what to write down before we build anything, and it tells us what not to build first.

4. What is repeatable?

The opposite of fragile: consistent inputs, stable decision rules, predictable outputs, bounded failure modes. Repeatable doesn't mean a model can handle it perfectly. It means the workflow has enough structure that you can put a system around it and measure whether the system is working.

The maintenance ticket flow is usually repeatable. The lease renewal cadence is usually repeatable. The owner who emails at 11pm with a request that needs judgment is not, and shouldn't be a build target.

The repeatable list is where the build will live. Confusing it with the fragile list is the standard failure of "let's automate everything" pitches. Most of the everything isn't ready.

5. What is the next dollar worth automating?

The question every operator wants to start with, and the one that can only be answered after the first four.

A good answer isn't a list of impressive things that could be built. It's a ranked roadmap: cost to build, expected saving, dependencies, sequence. Ranked by return on the constrained person's time, not by how well it demos. The vendor demo is always at the technological frontier, because the frontier demos well. The highest-return build, for almost every operation we look at, is boring: a database, a clear schema, structured notifications, a few carefully written prompts, a dashboard. (Choosing the tool for each piece is its own decision, covered in [Make vs n8n vs custom Claude](/blog/make-vs-n8n-vs-custom-claude).)

Answering question five well mostly takes the discipline to recommend the boring thing.

What you get at the end

Three documents:

1. An operations map. The calendar from question one, drawn out: where each role's hours actually go in a typical week.
2. An opportunity matrix. The workflows scored on time cost, fragility, repeatability, and model fit, with the math shown.
3. A 12-month roadmap. The answer to question five, ranked and sequenced: what gets built first, what comes next, and what's deliberately deferred.

The documents are the client's, full stop. Take them into a build with us, take them to another firm, or use them as a hiring plan. Writing the audit as if the client might walk away with it is exactly what keeps it honest.

The real deliverable

The roadmap is the artifact. The clarity is the product. Most operators have never seen their own business written down with numbers attached, and that view, more than any single automation, is what the two weeks buy.