An assistant that answers from your systems, not from the internet.
A question-answering interface grounded in your own records, scoped to what each person is allowed to see, and honest about the limits of what it knows.
Why most assistants get quietly abandoned
A general model answers confidently about your business and is wrong, because it has never seen your business. Staff test it twice, catch it inventing a supplier term, and go back to asking a colleague. The failure is not intelligence — it is that the assistant was never connected to anything true.
- The same operational questions are asked of the same three people every week.
- Answers depend on who you ask and how long they have worked there.
- A general-purpose chatbot was trialled and nobody uses it now.
- Information exists, but only in documents and screens nobody can query.
What we build
An assistant wired to your systems of record, which retrieves before it answers, cites what it used, and declines when the records do not support a response.
Read: data readiness is the real prerequisite →- Retrieval over your documents, database records and system data
- Permission-aware answers that respect the access each user already has
- Citations back to the record or document the answer came from
- A refusal path: when evidence is thin, it says so rather than inventing
- Deployment in your tenancy or on-premises where confidentiality requires it
- Usage and accuracy monitoring so quality is observed, not assumed
How it runs
Retrieval before generation, and permissions before retrieval.
- 01Scope the questions
We collect the questions people actually ask, then check which are answerable from records that exist today.
- 02Connect the sources
Documents, databases and system data are indexed with their access rules carried through, not stripped away.
- 03Retrieve, then answer
Each question triggers retrieval against the user’s permitted scope. The model composes an answer from what came back.
- 04Cite and constrain
Answers carry their sources. Where retrieval returns nothing adequate, the assistant says so instead of filling the gap.
- 05Watch and tune
Unanswered and poorly answered questions are reviewed, which usually points at missing data rather than a model problem.
What changes once it is running
What an assistant grounded in real records changes about a working day.
Answers stop depending on tenure
Knowledge held by a few long-serving people becomes reachable by everyone entitled to it.
Answers can be checked
Every response carries its source, so a user can verify rather than trust.
Permissions still hold
The assistant cannot become a route around access control, because retrieval runs inside it.
Gaps become visible
Questions the assistant cannot answer are a precise map of where your data is missing or unreachable.
How an engagement is shaped
Grounding is proven on a narrow domain before the scope widens.
Question audit
One to two weeks establishing which questions matter and which are answerable from current records. Some will not be, and that is a useful finding on its own.
Grounded pilot
A working assistant over one domain and one user group, measured on answer quality against questions with known answers.
Extend and operate
Additional sources and audiences added, with ongoing monitoring of quality as the underlying data changes.
Common questions
The things buyers ask before they commit. If yours is not here, it is a good first question for the assessment.
- Will it leak information across departments?
- Not if retrieval is permission-aware, which is how we build it. The assistant searches only what the asking user could already open. An assistant that ignores this is a data-loss incident waiting to be discovered.
- Can it run without sending data to a third-party model provider?
- Yes. Where confidentiality or regulation requires it, we deploy models inside your environment. That trades some capability for control, and we will be explicit about which capability.
- What stops it making things up?
- Retrieval grounding plus an explicit refusal path, and citations so a user can check. No architecture removes the risk entirely, which is why we scope assistants to informational work rather than to unsupervised decisions.
- How do we know it is any good?
- A question set with known correct answers, scored before launch and re-scored as data changes. If we cannot measure it, we do not claim it.
Bring the question your team asks every week.
One real operational question. We will trace what answering it reliably would actually take.