AI assistants
Useful answers from your own documents, with sources
Give staff or customers answers drawn from an approved collection of your documents, with a link to the page each answer came from and a clear limit on what the assistant will attempt.
Who this is for: Teams who answer the same questions repeatedly from policies, procedures, product information or contracts.
Working system available
A working in-house system in this family is available to demonstrate. Its case study shows representative behaviour, safeguards and limitations, together with any measured result and the limits of the run that produced it. No client outcome is claimed.
You will recognise this if…
- The answer exists in a document somebody wrote, and nobody can find it quickly.
- New staff ask the same questions for months, and answering them falls to whoever is nearest.
- A general chatbot would answer confidently and you would have no way to check whether it was right.
What we deliver
- Assistants that answer from an approved collection of documents, with citations
- Internal policy and procedure search
- Customer support and website chat assistants
- Onboarding and training assistants
- Document question answering
- Knowledge-base ingestion and refresh workflows
- Access controls, citations and escalation to a person
- Assistant evaluation and ongoing improvement
Common builds
Concrete systems in this family, described the same way the free Snapshot describes them.
Internal knowledge assistant
Staff ask a question in plain English and get an answer from the documents the business already holds.
A private search and answering layer over your drives, policies and internal documentation. It respects existing file permissions, so nobody sees through it what they could not already open.
AI customer support assistant
Common enquiries answered from your own documentation, with anything unusual handed to a person.
An assistant grounded in your actual policies, product information and past answers, so replies come from your material rather than a general model. It escalates rather than guesses, and every answer cites where it came from.
How we work on this
01
Choose the material
We agree which documents are in scope. A narrow, reviewed collection produces better answers than everything you have ever written.
02
Set the boundary
We decide what the assistant declines to answer, and what happens instead when it does.
03
Build and cite
Every answer carries the source it came from, so a reader can check it rather than trust it.
04
Evaluate
We test it against a reviewed question set, including questions it should refuse, before anyone relies on it.
Systems we connect for this
- OpenAI
- Anthropic
- Qdrant
- Microsoft 365
- Google Workspace
We connect the systems your team already uses. Naming a platform here means we can integrate with it, not that we are partnered with or endorsed by it.
Evidence
Working in-house systems in this family, shown with representative behaviour, safeguards and limitations.
Questions
- What is RAG, and is that what this is?
- Retrieval-augmented generation is the technical name for it. In practice it means the assistant searches an approved collection of your documents and writes its answer from what it finds, rather than from general knowledge picked up in training. That is what makes a citation possible.
- Can it make things up?
- The design reduces it substantially and the citation makes it checkable, but no assistant of this kind is immune. That is why we build them to decline questions the material does not cover, and why we evaluate refusal behaviour as carefully as answer quality.
Next step
Describe the process that prompted you to read this page. If it is a good fit we will say so, and if it is not we will say that too.