Once an AI system answers customers or changes records, two questions follow every incident: what instructions was it running? and what exactly did it do? In the Insight suite, those are two products. Insight Prompts governs the instructions; Insight Lense keeps the evidence.
Prompts as governed assets
In Insight Prompts, a prompt moves through four stages in one registry:
- Create: business-friendly authoring with reusable templates, variables, partials and few-shot examples.
- Test: run the prompt against curated datasets and have an LLM judge score the output against a rubric you wrote.
- Approve: multi-stage workflows with role-aware sign-off from owners, compliance and security.
- Deploy and monitor: applications pull certified prompts through the API or SDKs (Python, Node, Java), with per-prompt scorecards for latency, cost, judge scores and feedback.
Every prompt is versioned, diffable and can be rolled back. A new version can run as a challenger on a slice of live traffic and be promoted only if it beats the current one on the rubric.
Evidence, not just logs
A log of model calls says which model ran and what it cost. Insight Lense is built for the moment an agent can change a record, and keeps three kinds of evidence as one thread:
- The run: every step the agent took, in order, as parent and child spans grouped into sessions, with tokens, latency and cost on each.
- The authority decision: what the agent asked to do, what it was permitted to do, and who it had to ask. A refusal is recorded as a result, not an error.
- The business event: what actually changed, as an append-only record with old and new values, the kind of actor and the reason.
Around those sit cost and budgets derived from traces, alerts on failure rate, spend, latency and escalation, and a single score primitive for any judgement. Lense ingests standard OpenTelemetry (OTLP) and has a Python SDK with automatic instrumentation for Bedrock, OpenAI and Anthropic.
Testing the application, too
Prompts and traces govern the AI parts. Insight Test Bench covers the application around them: an AI-native QA platform that crawls every page, generates grounded test cases, runs them nightly and files pull requests.
Why did the agent set that reserve?
At a made-up specialty insurer, a claims agent raises a reserve from $50,000 to $250,000. A supervisor asks why. In Lense, one query on the claim returns the thread: the documents the agent read, the reasoning steps, the authority check (raising a reserve above $100,000 needed a human approver, who approved it), and the business event with old and new values. The trace also names the prompt version, v12, approved in Prompts two weeks earlier after beating v11 on the claims rubric.
Governance is a practice. These products record and enforce what you define: rubrics, approval steps, authority levels, retention. Deciding those, and reviewing them, is still your team's job, and regulatory expectations for AI vary by industry and region.
Five questions the records should answer
The Lense site frames observability by the questions people actually ask, each answerable with one query:
- An auditor: "Everything that happened to this account, in order."
- A supervisor: "Why did the agent settle at that number, and what was it refused?"
- Compliance: "Which actions were taken by agents rather than people this quarter?"
- Operations: "Did that event reach the next system, and how late?"
- Finance: "What did it cost to handle this piece of work?"
A useful exercise before any agent goes live is to write these five questions for your own process and check that the traces you're collecting can answer them.
Related: the reference architecture.