How it works
From public information to an answer you can stand behind.
Anyone can point a language model at a website. The distance between that and an answer a regulated company is willing to put its own name on is where the work sits. We build both the layer and the agent that runs on it.
Read the colours
Runs automatically, continuouslyA person decides, ours or yoursFive stages, and most of the work is not code
Collect
We read the public web the way a careful guest would, then add what only you have.
Public sources
Your own material
- robots.txt honoured per RFC 9309, crawl rate kept polite, scoped to the right market
- Re-crawls detect what changed, not just what exists
- Which sources are worth reading, and how deep to go
Why it's harder than it looks
Every site is built differently, and half of them quietly break your extraction next quarter.
Verify
Two tracks, handled differently, kept apart. They meet inside an answer, never in a database.
- Both tracks first: menus, banners and legal boilerplate stripped, only the answer-bearing text kept
- Your content is structured with the facts that decide whether an answer is right: products, credit process, credit bureau, markets
- Material that was never on your website is connected with you
- No other company's chat can retrieve your content, and it cannot become shared knowledge without both your opt-in and an explicit promotion step of ours
- Shared knowledge is written, checked against primary sources and signed by an industry expert before it can be served
Why it's harder than it looks
Keeping the two apart is an architecture decision you cannot retrofit. Blending them is easy, cheap, and the reason most of these systems cannot be sold to a bank.
Organise
One order of precedence, applied to every question, every time.
- Your content and the shared layer are indexed separately, per company
- Written and indexed once in English, answered in 50 languages
- Structured facts about your company, products and market role, held as ground truth
- The order of precedence is agreed with you, and any tier you do not want is switched off
Why it's harder than it looks
Reaching for the wrong shelf is worse than an empty shelf. A confidently mis-sourced answer is the one that ends up in a screenshot.
Answer
Every answer is checked before anyone sees it.
- Written from retrieved passages, with the source shown next to it
- A second model reads the answer back against those passages and blocks anything unsupported
- Guardrails: no invented prices or rates, no advice outside scope, no personal data repeated back
- The AI model is a component, not the product. We test and swap the underlying LLM as better ones arrive; your knowledge, rules and voice stay put
- The tone, the limits and the escalation rules are set with you, per company
Why it's harder than it looks
Knowing when to refuse is a tuned behaviour, not a switch. Too cautious is useless; too willing is a liability.
Deliver
Wherever your customers already are, in both directions.
- Chat widget on your own site, in your own branding
- Portal for your team: what was asked, what was answered, what was missing
- MCP for other AI assistants, and email and voice in and out, are coming
- A person approves the message and who receives it before anything goes out
Why it's harder than it looks
Anyone can send a message. Handling the reply is the hard part, and it only works when the message and the answer come from the same verified source.
The answer order
Three tiers are served. Three exist and are refused.
- 1
This customer's own case
A document they shared, or your own systems
Primary
- 2
Your own content
Your site, your documents, what you gave us
Primary
- 3
The shared verified layer
Our verified articles first, authority sources after
Fallback
- 4
Another company's private content
Never shared
- 5
Unverified or expired content
Never served
- 6
The open internet
Never served
If none of the first three carries the answer, the agent says so. It does not fall through to the ones below.
Stage 2 in detail
The verified knowledge layer, and why it is the slow part
Every company's own content stops at the edge of what that company publishes. Customers ask past that edge constantly: what a payment remark actually does to you, what the enforcement authority can and cannot take, how a credit check shows up two years later. Nobody's website answers that, because it is nobody's job to.
Across 864 scored answers, four in ten drew wholly or partly on this shared layer. On credit-bureau questions it was closer to seven in ten. The thinner a company's content is on a topic, the more of the answer this layer carries.
Watch
Authorities and consumer bodies are crawled continuously. Not to answer from, but to catch change: a new statute number, an edited amount, a page quietly replaced.
Draft
An article is written to answer a question a real person asked, drawing several sources together. One authority page almost never answers what was actually asked.
Verify
An industry expert checks it against the primary sources: the statute itself, the authority's own wording, the amount as it stands this year.
Sign and date
The article carries who verified it and when. An unsigned draft cannot be served, however good it reads. This is enforced by the system, not by discipline.
Re-review
When the watch step sees the ground move, the article goes back to a person. A rule change is a re-verification, not a note in a backlog.
Retire
When a newer article covers the same ground, the older one is withdrawn rather than left to drift and quietly contradict its replacement.
Running underneath, always
The quality loop
Answer quality is a number here, not an impression. Every answer is scored against a reference a person wrote, and the score has to hold before anything reaches a customer.
Benchmark
Runs draw whole conversations from a bank of 781 questions across eight product types, by a fixed seed, so results stay comparable after a data change.
The facit
Every question carries a reference written and reviewed by a person: the direction a correct answer takes, or what a good answer must do and must never invent.
Judge
An independent judge model, never the one that wrote the answer, at temperature zero so a re-run gives the same number. Twelve axes, scored one to five.
Triage and gate
Each company gets a verdict on whether its chat is fit to turn on at all. A change that lowers a score does not ship. That call is ours, not the model's.
What the judge weighs
Six categories on a 0–100 scale. All of it judged from the question and the answer alone, with no access to what the chat retrieved, which means the same measure can be run on any chat, including the one you have today.
- Answer correctness
- Are the facts right, and right for Swedish conditions
- Hallucination control
- Are there claims with no support anywhere
- Compliance
- Is advice kept inside its limits, and is personal data never repeated back
- Source transparency
- Does the reader see a source where a claim needs one
- Answer relevancy
- Does the answer cover what was actually asked
- Tone of voice
- Does it sound like you, and is the length right for the question
Compliance is a behavioural check on how the chat speaks, not a legal audit against a statute. The categories are weighted against each other, but the weights are a tool for us rather than a promise to you.
Three measures that require us to own the index
All three are quantitative and measured continuously, 0 to 100. They cannot be computed for an external chat, because they need visibility into exactly which sources an answer was built on.
- FaithfulnessGreen ≥ 95
- The share of claims in the answer actually carried by the sources retrieved.
- Citation accuracyGreen ≥ 85
- Of the claims that show a source, the share where that source honestly carries the claim.
- Data qualityGreen ≥ 70
- How well your own content carries your own customers' questions, or whether we need to fetch more.
- ProvenanceSums to 100
- Where the answer actually drew from: your own content, the shared layer, both, or neither.
The thresholds differ per measure. Faithfulness carries the strictest of them all, because a chat answering from our own index has little excuse for a claim the sources do not carry.
~100
Swedish companies whose content the knowledge base already covers
4 in 10
answers draw on the shared verified layer, rising to 7 in 10 for credit bureaus
50
languages answered from one canonical English source
92%
of the claims in an answer are carried by a source we retrieved
781
benchmark questions, each with a reference answer written and reviewed by a person
What does it look like for your company?
The knowledge base already covers close to a hundred Swedish companies, and yours is likely one of them. Ask a question in the demo and see what it answers about you today, with sources. For the full measurement we run your own question set and show the result as it came out, including the parts that did not hold.
