Evidence-backed sales answers in a conversation, without waiting for SQL

Executive summary

Chatwit’s leaders and business users kept asking the same kind of question: why did sales performance change? Every answer depended on technical staff to find the right data, write SQL, run several queries and interpret the results. DevOps TechLab deployed QueryDash, its agentic investigation platform built on Amazon Bedrock, so users can ask that question in plain language and receive a step-by-step, evidence-backed explanation. Every query is validated as read-only before it touches production data, and DevOps TechLab runs and improves the service as a managed engagement.

Client

Chatwit · Sales analytics

Business problem

Every answer to “why did sales change?” depended on technical staff writing and running SQL.

What we delivered

An AI investigation agent (QueryDash) on Amazon Bedrock with Amazon Nova Pro, connected read-only to Chatwit’s business data.

Engagement

Discovery, build and onboarding, then managed operations and continuous optimisation by DevOps TechLab.

Outcome

Business users investigate sales performance themselves, in plain language, with evidence for every finding.

The business challenge

Chatwit wanted its users to understand business and sales performance without waiting for a technical team. Getting to an answer meant identifying the right data, preparing SQL, running multiple queries and interpreting the results before anyone could act.

Why it mattered

• Decisions waited on analysts

  Each new question had to go to technical staff, which slowed the move from a business question to an action.

• Skilled people on repeat work

  Technical staff spent time finding data, writing SQL and running several queries for every request.

• Dashboards answer “what”, not “why”

  Fixed reports show that a number moved. Explaining the cause meant connecting several data points by hand.

• Every follow-up started again

  There was no way to keep the context of an investigation, so each follow-up question meant another request.

What Chatwit set out to achieve

1. Let business users investigate performance in plain language, with no SQL and no ticket queue.
2. Return explained, evidence-backed findings, not just raw numbers.
3. Keep the context of an investigation so users can ask follow-up questions.
4. Remove any risk to production data from AI-generated queries.
5. Use one platform for new questions, instead of a new analytics project for each.

The solution: an AI analyst that investigates, not just queries

Users describe a business concern in plain language, for example “Investigate why sales performance changed and guide me through the findings.” The agent plans the investigation, finds the relevant data, writes and validates each query, analyses the results and decides the next step, recovering on its own when a query fails. Users watch the investigation unfold in real time and can ask follow-up questions in the same conversation.

Key design choices

• Managed generative AI

  Amazon Bedrock with Amazon Nova Pro handles reasoning and tool use. There are no models to host and no GPU infrastructure to run.

• No servers to manage

  The application runs on Amazon ECS on AWS Fargate behind an Application Load Balancer, so it scales with demand.

• Works on the data where it lives

  The agent queries Chatwit’s existing operational databases directly, read-only, instead of copying data into a new warehouse.

• Conversations that remember

  Amazon DynamoDB keeps the conversation and investigation history, so follow-up questions build on what the agent already found.

Security and governance

For leadership, the first question about AI on production data is risk. The platform is built so that AI-generated queries never touch the database directly.

• Read-only by design

  Only read statements (SELECT, WITH, SHOW, DESCRIBE, EXPLAIN) are allowed. Any write or schema change is rejected before it runs.

• Validated before execution

  Every query the AI writes passes through a server-side validation layer. The model never has a direct line to production data.

• Protected performance

  Results are capped, so a broad question cannot turn into a runaway query on a production database.

• Managed credentials

  Database credentials sit in AWS Secrets Manager, and access is governed by AWS IAM roles, not stored in code.

Architecture on AWS

Click to enlarge

How we delivered

1. 01Discover

   Mapped the questions business users ask, the data behind them and how it relates, and defined secure, read-only access to production.

2. 02Build & onboard

   Connected Chatwit’s data, configured Amazon Bedrock and Nova Pro, set up read-only validation and conversation memory, deployed on AWS and tested against real investigation scenarios.

3. 03Run

   DevOps TechLab operates the service: monitoring the application, the AWS environment and the agent, managing credentials and access, and resolving production issues.

4. 04Improve

   Ongoing tuning of agent planning, query generation and recovery, plus reviews of performance, cost and security, and new investigation scenarios as Chatwit’s needs grow.

Business outcomes

Chatwit’s users now move from a business question to an explained answer in a single conversation, instead of a chain of requests to a technical team.

• ✓
  Self-service answers

  Business users investigate sales performance themselves, in plain language, without SQL or a ticket queue.

• ✓
  Decisions backed by evidence

  Each finding comes with the data, metrics and reasoning behind it, so leaders can act on it with confidence.

• ✓
  Technical teams freed up

  Less manual SQL analysis means engineers spend less time on repeat data requests.

• ✓
  Zero write access to production

  AI-generated queries are validated and read-only, so production data is never changed by the agent.

• ✓
  A platform, not a one-off

  New investigation scenarios are added on the same platform, without a separate analytics project for each.

What’s next

Because the architecture is reusable, Chatwit can extend the agent to new business questions without building a separate analytics workflow for each. DevOps TechLab continues to advise on new conversational analytics scenarios and further opportunities for AI-driven automation.