Not a language model. A dependency graph somebody had to write by hand.

Most of what gets sold as an AI audit is a prompt with a logo on it. This is worth explaining properly, because the difference is the entire product and it is invisible from the outside.

A deterministic engine, not a probabilistic one.

Your answers never pass through OpenAI, Anthropic, or any other third-party model. The engine walks a graph that we authored, node by node, over years of doing this work by hand. The same answers always produce the same findings in the same order. A language model cannot promise that, because generating text is not the same operation as resolving a dependency.

91questions in the restaurant set
60nodes in the dependency map
91named software platforms
80documented failure modes, across 57 of 60 nodes
42–76findings returned, depending on your answers
35shared core nodes across all industries

It answers the question underneath the one you asked.

Findings enter your report that no answer triggered. They are there because something you did flag cannot be built without them first. This is the part that separates a graph from a lookup table, and it is easiest to show than to describe.

Tell a general AI “my inventory lives in a spreadsheet”

It recommends inventory software. Maybe three platforms, maybe a guide to connecting them. It answered your question correctly and completely.

The problem is that inventory software cannot calculate depletion without a baseline, and you do not have one.

The engine returns four items

  1. Complete one full manual count
  2. Write recipe cards with verified yields
  3. Start reading the reports your POS already makes
  4. Connect inventory software to the POS

Only the fourth came from what you said. The first three are groundwork you did not raise, and the report labels every item as one or the other so you can see which is which.

Some findings exist only because it reads every answer at once.

Manual re-entry turns up in five separate questions. Sales typed into the books. Hours re-typed into payroll. Invoices in a pile. Inventory counted on paper. A POS that connects to nothing. Answer those individually and you have five problems.

Two or more of them together is one integration problem wearing five costumes, and the engine emits it as a single structural finding rather than five tasks. This is the argument I would make first if I only had one. A conversation that answers one question at a time cannot reach it, no matter how good the model is, because the finding does not exist in any single answer.

Three of the four automation states are refusals.

The engine resolves the graph into phases, then sorts within each phase by how many downstream items each one unlocks. Every item carries an automation state, and most of them are the engine telling you not to.

yes

Automate it. The judgment involved is mechanical and the data underneath is reliable.

partial

Automate the mechanics, keep the judgment human. Usually where a person needs to see something before it goes out.

not_yet

Correct eventually, wrong today, with the gate named. You get told what has to be true first.

never

The human judgment is the point. Automating it would remove the thing that makes it work.

Automated par-level ordering is the item most restaurant owners want first. It sits fifth in its chain and is gated on roughly eight weeks of real depletion history, and the report says so in the words it says everything else in: par levels set from guesswork produce automated orders of the wrong quantities, delivered faster and with less human review than the phone call they replaced.

A product whose business model rewards recommending maximum automation does not write that sentence.

Two fair criticisms, answered directly.

“An architect reads it but does not write it, so the advisory value is zero.”

The review is subtraction and correction, and it is worth being specific about what that means rather than leaving it as a reassuring word. It checks each finding against what you actually answered. It resolves items that contradict each other, which happens when answers to different questions imply different things about the same system. It confirms the sequence holds for your situation rather than for the average business in your industry. And it cuts anything that reads as generic, because a finding that could apply to anyone is not a finding.

If a report comes out wrong, that is where it gets caught. One correction and regeneration is included after delivery, so if something is wrong that the review missed, it gets fixed without another fee.

“If I do most of the work myself, I am paying for a checklist.”

Knowing which parts you can do, and what order they go in, is the finding. A typical restaurant report breaks down like this:

17you can do yourself
13someone on your team can do
9one free call with a vendor you already pay
<10need API and database skills

The nine vendor calls are the clearest case. Those are features you are already paying for and nobody turned on, and the report names the vendor and what to ask for. You would not have known to make the call, which is a different thing from not being able to make it.

What this does not do.

A page that concedes nothing gets discounted wholesale, and rightly. Four limitations worth knowing before you buy rather than after.

The credit only has value if you use it$500 credits toward a consultation, the Business Systems Audit, or implementation, within a year. A buyer who never engages further has paid $995 for a document.
There is no technical implementation in itIt does not show you how to write a webhook or debug a sync failure. It identifies which items need those skills and says so.
The AI prompt blocks are a convenienceThey carry the phase, its order, its prerequisites and its limits, and deliberately exclude later phases so a model cannot range ahead. Useful. Not proprietary engineering.
The individual steps are ordinaryWalking the building to inventory your equipment is not a novel idea. The claim is about the sequence, the prerequisites, the refusals and the failure modes.

The things people ask before buying.

Is the report generated by AI?

No. The report is produced by a deterministic rules engine over a hand-authored dependency graph. Your answers are not sent to OpenAI, Anthropic, or any other third-party model. The same answers always produce the same findings in the same order, which is not true of a language model.

Why can I not just ask ChatGPT this?

A general model answers the question you asked. Ask it about inventory on a spreadsheet and it recommends inventory software. The engine returns four items and only the fourth is the one you raised, because software cannot calculate depletion without a baseline count and verified recipe yields first. It also reads all 48 answers at once, which is how it consolidates five separate symptoms of manual re-entry into one integration finding. A conversation answering one question at a time cannot reach that.

Does a human review it, and what does that change?

A systems architect reads the complete roadmap before it is delivered. The review checks findings against your answers, resolves anything contradictory, confirms the sequence holds for your specific situation, and removes items that read as generic. It is subtraction and correction rather than rewriting. If a report comes out wrong, the review is where that gets caught, and one correction and regeneration is included after delivery.

How many findings are typical?

A restaurant report returns forty to eighty-plus depending on industry and answers, grouped into eight phases. The dependency map has 60 nodes, 91 named software platforms, and 80 documented failure modes across 57 of those nodes.

Can I do the work myself?

Most of it. A typical restaurant report has about 17 items you can do alone, 13 for someone on your team, 9 that take one free call with a vendor you already pay, and fewer than 10 that need API and database skills. Knowing which is which, and what order they go in, is the finding.

What does the $500 credit actually get me?

It credits toward a consultation, the Business Systems Audit, or implementation work, for one year, with no project minimum. If you never engage further, you have paid $995 for the document. That is worth saying plainly rather than burying.

What is the difference between this and the $4,500 audit?

This one is built from what you tell us. The Business Systems Audit is built from what we inspect and verify ourselves in your systems, contracts and data, and it produces dollar figures per finding. This one estimates time where it can and says so.

Run it on your own business and judge the output.

No email, no sales call. Twelve to sixteen sequenced items back. If the free one does not understand your operation, the paid one will not either, and you will have found that out for nothing.

Take the free scan →