MAVIS — Built in
We wanted ECHO’s users to ask the software itself. So we built MAVIS.
MAVIS is a custom AI assistant that lives inside ECHO. She answers from the product’s own written knowledge, shows you the article she used, and passes you to a human the moment she is out of her depth. She is the evidence behind the claim that this studio builds AI assistants.
What it is
MAVIS — Marketing Assistant for Voice, Ideas & Strategy — sits in a panel on every signed-in screen in ECHO. A member types a question in their own words and gets an answer grounded in ECHO’s help corpus: thirty-odd articles, a full product walkthrough, and a reference sheet.
She knows which screen the person is on, so "how do I do this?" has a subject. She links the article behind every answer, so it can be checked. And when the question is outside what she has been given, she says so and offers to send the conversation to a person in one tap.
She is a new front door to knowledge that already existed — not a new source of it. That distinction is the entire design.
Why it was built
Every SaaS product accumulates the same problem: the answer is written down, in an article nobody found, using a word the customer didn’t search for. Support tickets arrive that the help centre already answers.
The obvious fix is a search box, and search boxes have been failing at this for twenty years — they match words, not questions. What people want is to ask.
The decision was to let them ask the product itself, but to bound it hard: MAVIS can read and she can talk, and that is all. She cannot see an account, cannot change anything, and cannot answer from outside the corpus. A support assistant that invents a policy is worse than no assistant at all.
What’s inside
The product itself.
Drag or scroll Swipe
A third mode inside the help panel ECHO already had — ask, or browse. She opens by saying what she cannot do, which is the whole design in one sentence.
She answers from ECHO’s own written help rather than from a general model, so she can explain a feature and then tell you when not to use it. Every answer carries the article it came from.
The handover, and the source chip from the previous answer sitting above it. One tap sends the whole conversation to a person, so nobody has to explain themselves twice.
How it was built
Strategy, product, build, launch.
Decide the boundary
Scope came before architecture. In: the written corpus, the current screen, links to sources, handover to a human, and a log of every question asked. Out, deliberately: account data, actions of any kind, and anonymous public access. Naming the out-list first is what made the build small.
The knowledge
The help articles moved server-side so the corpus could be assembled into the request rather than shipped to the browser. It is small enough to hold whole, which means MAVIS can answer across articles instead of retrieving one — and there is no retrieval layer to get wrong.
The interface
She was built into the help widget already shipping on every signed-in screen, as a segmented toggle beside the existing article browser. A new assistant that needs a new place to live is an assistant nobody opens.
The safeguards
Grounded strictly in the corpus, so an unanswerable question produces "I don’t know, here’s a human" rather than a plausible invention. Behind authentication, so every question is attached to a known person. Rate-limited per user. Bounded timeout with a single retry. Every exchange logged, so what people actually ask becomes product evidence.
The outcome
What it did.
MAVIS ships inside ECHO as part of the product rather than as a bolt-on, and the question log has turned into a standing list of what the documentation still explains badly — which is arguably worth as much as the answers themselves.
For this studio she is the working example behind a claim other people make loosely. An AI assistant trained on your own material, wearing your name, bounded so it cannot embarrass you, built into the product you already have — that is a job with a finished example attached, not a proposal.
Visit ECHO, the platform MAVIS lives inside
MAVIS has no address of her own — she ships inside ECHO, behind sign-in.
Questions
What people ask about MAVIS.
Can AI be trained on my own content?
Yes, and for most businesses it should be — an assistant answering from your material is useful in a way a general chatbot is not. MAVIS answers only from ECHO’s own help articles, product walkthrough and reference sheet.
"Trained" is usually the wrong word, though. For a corpus this size nothing is retrained: your material is supplied to the model as the source it must answer from, which is faster to build, cheaper to run, and instantly correctable — fix the article and the answer changes.
How do you stop an AI assistant making things up?
By bounding it. MAVIS is instructed to answer only from the material she has been given, and to say she does not know when the answer is not in there — with a route to a human attached.
The other half is what she cannot do. She cannot see account data and cannot change anything, so the worst failure available to her is an unhelpful answer with a visible source next to it, rather than a confident wrong action.
Every answer links the article it came from, which makes her checkable. An assistant you cannot audit is an assistant you cannot trust.
Can I control what the assistant is allowed to say?
Completely. The knowledge is a set of documents you own and edit; the tone is specified; the topics it will refuse are specified. MAVIS coaches on method — how to record a better voice note, how to get more out of one recording — and declines to have opinions on anything outside that.
If a policy changes, you change the document. There is no retraining cycle and no waiting.
Can an AI assistant be built into the software I already have?
That is the usual case, and it is the cheaper one. MAVIS was built into a help panel ECHO was already shipping — a new backend route, the corpus moved server-side, a chat view inside existing furniture, and a logging table.
Most of the cost in this work is not the AI. It is deciding the boundary, writing the knowledge properly, and building the handover for the moment it fails.