The InsightLake suite has more than thirty products. Nobody should start with thirty. The suite's own guidance is that most enterprises start with Ask, RAG and Lense, then add Catalog, Prompts and DLM as the programme matures. This guide turns that into a plan with gates, so each phase earns the next.
Phase 1: prove value with one team
Pick one team with a clear, frequent question that today takes too long to answer: on-call engineers, a member-services desk, an underwriting team. Connect two or three sources they already trust. Deploy Insight RAG behind Insight Ask, and trace every answer in Insight Lense from day one.
- Build a test set first: forty to a hundred real questions with the document that answers each. It's how you'll know whether retrieval works, before anyone has an opinion.
- Review permissions by role with security before launch.
- Read the traces weekly: wrong answers are either a retrieval problem or a model problem, and the trace says which.
Gate: the team uses it without being asked, and the answers it gets wrong are understood.
Phase 2: make it governable
More sources and more teams mean more risk. Add Insight Catalog so every source has an owner, lineage and sensitivity classes, and so RAG's access rules have something reliable to enforce. Move prompts into Insight Prompts: versioned, tested against datasets, approved before they ship, and promoted only when a new version beats the old one. Lense links each trace to the prompt version that produced it.
Gate: no prompt or source change reaches users without review, and cost and quality are reviewed on a schedule.
Phase 3: specialise
With a governed base, go deeper where the business case is strongest:
- Insight DLM for domain-specific small models trained on your data, for tasks where a smaller model you own is faster, cheaper or more accurate than a general one, trained and served inside your environment.
- A domain solution such as Insight UW for underwriting, Carts for retail or Variant for genomics, built for the people who do the work.
- Agents that act, through Insight AI Assist, with authority levels recorded in Lense: what each agent may do alone, and what needs a person.
Gate: for every action an agent can take, you can show what it did, what it was allowed to do, and who approved it.
Eighteen months, three phases
A made-up regional bank started Phase 1 with its IT service desk: ServiceNow and Confluence, about 60 test questions, Lense from day one. Six months later, Phase 2 brought in the operations and compliance teams, the Catalog for data owners and lineage, and Prompts with a two-step approval. In Phase 3 it built a small domain model for classifying customer complaints with DLM, and only then let an assistant draft responses, with every draft approved by a person and traced.
Plans are local. Timelines depend on your data, your security reviews and your teams; the phases are an order, not a schedule. Talk to us about which products fit your first phase.
Signs you're moving too fast
- Adding sources nobody owns. Every source needs someone who can say who should see it and whether it's current.
- Changing prompts directly in code. If a prompt change can ship without review, you're not in Phase 2 yet.
- Agents with no authority levels. If you can't say what an agent may do alone, it shouldn't act yet.
- No test set. Without real questions and known answers, "it seems better" is the only measure.
Signs you're ready for the next phase
- The first team asks for more sources rather than better answers.
- Traces are being read, and wrong answers have known causes.
- Other teams are asking for the same thing for their own knowledge.
Related: the reference architecture · running AI inside your environment.