Every agency has now tried to use an LLM for onboarding, and most have quietly stopped. The reason is always the same: it produced something plausible, someone shipped it, and it was wrong.
That is not an argument against AI in onboarding. It is an argument against a specific, lazy implementation of it — the one-shot prompt.
The one-shot problem
Ask a model “what are Acme Chiropractic’s opening hours” with nothing else in context, and it will answer. It has to answer; that is what it does. The answer will be a well-formed, confident, entirely fictional set of hours.
The model was never given a way to check. So it did the only thing available: it produced the most likely-looking string.
Grounding, then tools
There are two fixes, and you need both.
Ground it before it generates. Crawl the client’s actual website into a local knowledge base and pass it as context. Pass every other field on the client record too, so a generated value is consistent with the values a human already approved. Now the model is summarizing evidence rather than inventing it.
Give it tools, and let it loop. A field-generation call should be a short agentic loop — up to five iterations — with real tools available:
- a Google Business lookup, which returns the client’s actual address, hours, phone and review presence;
- a domain lookup, which returns real DNS and registrar facts.
Now “what are their hours” is a lookup, not a guess. The model calls the tool, reads the result, and fills the field from it. When a tool’s key is not configured, the loop degrades gracefully instead of failing — and reasoning models that do not support forced tool selection fall back to a single-shot call automatically.
The value still has to be approved
This is the part people skip, and it is the part that makes the whole thing safe to use.
A generated value should land in the field with its status intact. It is “supplied”, not “approved”. It moves through the same review flow as anything a human typed, and it reaches the CRM through the same durable write queue. A person still signs off.
What the AI removes is the typing. It does not remove the judgment, and any tool that claims otherwise is selling you a liability.
What that looks like in a day of work
Paste the client’s domain. Brand Fetch returns their colors, fonts, tagline, about-us copy, socials and logo. The crawl builds a knowledge base of their site. Field generation fills the rest of what is publicly knowable, checking itself against Google Business and DNS as it goes.
Your team opens the client and reviews thirty fields instead of researching and typing them. The kickoff call is about the four things a website genuinely cannot tell you.