Most teams do not have an AI problem. They have a handoff problem — work that falls between two tools, or sits in someone's head, or gets re-typed three times — and a model dropped on top of that makes the mess faster rather than smaller.
A lead arrives, and somewhere between the form and the follow-up it waits. Usually for a person who does not know it is waiting. You cannot automate a handoff you have not found, and the handoffs are rarely where people assume they are.
A CRM that sends email, a scheduler that also sends email, a form builder with its own automations. Each one was the right call on the day it was bought. Together they fire twice, disagree about the same contact, and nobody owns the question of which is authoritative.
The task somebody disliked most gets automated first. That is a comfort decision, not a throughput decision — and it is why so many stacks have an elaborate workflow attached to a step that happens four times a month.
An LLM writing copy, qualifying a lead or answering a customer will be wrong on some percentage of runs. Unattended, that percentage ships. The question is never whether to review the output — it is which steps get a gate and who stands at it.
GPTBot, ClaudeBot, PerplexityBot, Google-Extended and CCBot are each named separately in robots.txt, and a well-meaning blanket rule can block all of them at once. Most site owners have never read the file; fewer have read what it says to each agent individually. This is one section of the audit, not the whole of it.
None of that requires a guess. Every one of them is observable in your own tools and your own site, which is what the audit stage is for.
01
Every point where work changes hands — between two tools, two people, or a tool and a person — written down in one place, in the order it happens.
Where the work waits, and for how long. A step that takes two minutes and sits for two days is a queue, not a task.
Which steps are load-bearing for revenue and which are habit. Automating the second kind is how a stack gets busier without getting faster.
02
A recommendation per handoff, with the reason. Deterministic work belongs to plain automation; judgement under ambiguity belongs to a person; pattern work at volume belongs to a model.
Where a model is recommended, the REVIEW GATE is specified with it — who sees the output, before what, and what happens when they reject it. A step with no named reviewer does not get a model.
Where we recommend NOT automating, we say so. The honest answer for a step that runs four times a month is usually to leave it alone.
03
Read against your real stack — GoHighLevel, Make, your forms, your analytics — rather than against a reference architecture.
Where two tools overlap, which one should own the step, and what breaks if we move it.
What is missing, and whether it needs buying or can be built from what you already pay for.
04
What your robots.txt says to each AI crawler by name, whether a Content-Signal is published, and whether llms.txt exists and names you correctly.
Whether your content is server-rendered and therefore parseable, or assembled in the browser and therefore not.
We honour Disallow during the audit itself, including where we could technically route around it. A vendor who ignores the one machine-readable place you can say no is telling you something about how they will treat your site later.
05
Every section reports what it could not assess, and why.
A missing finding otherwise means either the step is fine or we never looked, and from the outside those are identical. That ambiguity is where every padded audit hides.
You get the raw readings with the date we took them, so you can check the work or re-run it later.
Two stages, priced separately on purpose. You can stop after the first one and keep the map.
If the audit concludes that the honest answer is to change three things and automate nothing, that is what it will say — and the map is still worth what you paid for it.
No. A build quoted without the handoff map is a guess priced as a fact, and the first thing we would discover is that the scope was wrong. The audit is what makes the build quote fixed.
Then it says that. A step that runs four times a month rarely justifies a workflow, and recommending one anyway would be selling you a build you do not need. You keep the map.
Because we cannot defend the number. We have no measured cross-client results to quote and will not publish one before we do. Every vendor quoting you a percentage is quoting an average of other people's businesses, or nothing at all.
Somebody on your side, named during the audit. A review gate with no named person is not a gate, and we will not implement a model-touched step without one.
For the audit, read access is enough, and the site portion needs no access at all — it is observable from outside. The build needs write access to the tools we are changing, scoped to those tools.
Five business days from the point we have read access, and we tell you up front what we could not reach.
The audit is fixed in fee and scope, credited in full against the build, and yours to keep either way. If it turns out there is less to automate than you expected, that is a useful answer too — and a cheaper one than finding out during a build.