An assistant that knowswhat it does not know.
AI assistants grounded in your own pages, documents and records — and the automation around them.
Grounded assistants and workflow automation, attached to a build rather than sold beside one. An assistant is only as good as the material underneath it, so it is evaluated against real questions with known answers before it ever faces a visitor, refuses when it has no basis, and hands over to a person.
What this covers
Answers out of your own pages
An assistant grounded in the site itself — its pages and the documents you hand it. It answers out of what is actually written down rather than a plausible version of it, refuses where it has no basis, and hands over to a person. It never invents a price, a date or a policy.
Answers out of your records
The same thing reading the records the site already runs on — products, bookings, entries — so what it says is what the system currently holds rather than an export taken at launch. Access rules are enforced on the server: it can only read what the visitor asking could. Enquiries land in your admin with the conversation attached.
Work that runs without anyone pressing anything
An enquiry arriving becomes a record, a message and a task without retyping. Documents read, filed and passed on. Jobs on a schedule that report when they fail rather than failing quietly. You get a written map of every workflow, and you keep it.
What it costs
All of it needs hosting under it. An assistant has to run somewhere, with a database behind it.
What it costs to build and what it costs to run both come out of the scope rather than off this page. A short paid analysis settles what the assistant has to answer, what it is grounded in, and how much attention it will need once it is live; both figures are written against that.
True of all of it
Evaluated before it is seen
A set of real questions with known answers, run against the assistant before release and again after every change to the material. An assistant that has never been tested is a demo.
It refuses
Where there is no basis in your material, the answer is that it does not know and here is a person. That refusal is the feature — an assistant that always answers is an assistant that sometimes invents.
Reviewable, not opaque
The conversations are yours to read. That is how you find the questions your site does not answer, which is worth having even in the months nothing goes wrong.
A person at the end of it
Handover is designed in, not a dead end with a form. The assistant exists to answer the easy nine and route the tenth, not to stand between a buyer and you.
Your material stays yours
The documents and records it reads are yours, and so is the written map of what runs where. Nothing here is a black box you would have to rebuild from scratch elsewhere.
It attaches to something real
This runs on a build we can see and keep — which is why it sits beside the ladder rather than on it, and why it needs hosting under it.
How it runs
- [ 01 ]
Material
What the assistant is allowed to know: which pages, which documents, which records, and — just as important — what is deliberately out of scope. Written down before anything is built.
- [ 02 ]
Evaluation set
Real questions with known answers, written with you, including the ones it must refuse. This is what separates an assistant from a demo, and it comes before the build rather than after it.
- [ 03 ]
Build and grounding
Wired to the material, with access rules enforced on the server so it can only read what the visitor asking could. Then run against the evaluation set until it passes it.
- [ 04 ]
Review
Conversations read, gaps in the material fixed, the evaluation set re-run. That is monthly work rather than launch work, and it is part of what the running cost covers.
Questions
What does it need underneath it?
Hosting, all of it — the assistant runs somewhere and reads from a database, so there has to be something under it to run on. Beyond that the work does not stop at launch: there are model costs every month, the evaluation set has to be re-run whenever the material changes, and somebody has to read the conversations and fix what the site failed to answer. An assistant left alone for a year is answering out of last year’s prices.
Why is there no price on this page?
Because an assistant is not one thing with one figure behind it. What it costs to build moves with how much material it has to be grounded in and how tangled that material already is. What it costs to run moves with how often that material changes and how heavily the thing is actually used — which nobody knows, us included, until it has been answering for a while. A number printed here would be an average of jobs with nothing in common. The analysis puts both figures in writing, against a scope that is yours to keep.
Can it make something up?
It can be made much harder, and that is what the grounding and the evaluation set are for: it answers out of your material, it is tested against questions whose answers are known, and where it has no basis it refuses and hands over. What we will not tell you is that it is impossible — anyone who does is selling. What we will do is show you the evaluation results before it goes live and keep the conversations readable afterwards.
Can I have one without a website from you?
Not usefully. The assistant is grounded in a site and the records behind it, so if we did not build that we have no material to ground it in and no way to keep it current. If you have a site you are happy with, the honest answer is that the first job is the site, not the assistant.
What can the automation actually do?
The ordinary work that currently costs somebody an hour a day: an enquiry becoming a record, a message and a task without retyping; documents read, filed and passed to whoever needs them; jobs on a schedule that report when they fail instead of failing quietly. It is deliberately unglamorous, and you get a written map of every workflow so you know what runs, when, and what happens when it does not.
Where does the data go?
The material stays in your system. What goes to a model is the part of it needed to answer the question in front of it, and we tell you which model, where it runs and what is retained before anything is wired up — written down rather than described in a call.
Start with the questions you answer twenty times a week.
No demo theatre and no proposal written before we understand it. Tell us the questions you answer twenty times a week and we will say which an assistant can take.
If the problem is really an internal tool rather than an assistant, that is on the custom software page.
Custom software →