How a $9B+ enterprise software leader cut analytics turnaround by 75% with a governed multi-agent assistant
One hour per analytics request, reduced to 15 minutes using agents, for accelerated data product delivery. A two-person phData pod, embedded in the customer’s Enterprise Data & Analytics team, built a governed multi-agent assistant on Claude and Amazon Bedrock AgentCore and took it to production in 10 weeks.
At a glance
| Industry | Technology |
| Scale | About 21,000 employees, $9B+ annual revenue, about 150 certified internal data products across 25 data domains |
| Challenge | Each analytics request took about one hour of manual SQL, analysis, charting and report writing, and metric definitions were duplicated across systems |
| Technology | Anthropic Claude (Opus, Sonnet, Haiku), Amazon Bedrock, Amazon Bedrock AgentCore (Runtime, Gateway, Identity, Memory), Strands Agents, Amazon EKS, Amazon S3, AWS Secrets Manager, Amazon CloudWatch, AWS X-Ray, Snowflake, Atlan, Okta |
| Timeline | Working system in production in 10 weeks |
| Result | 75% reduction in time per analysis, from 1 hour to 15 minutes, with about 93% accuracy against a 90% release threshold |
| Phdata Service | AI Workforce, Generative AI, AI & Machine Learning, Anthropic, AWS |
Every analytics request cost an hour, and metric drift eroded trust
The customer’s Enterprise Data & Analytics (ED&A) team owns about 150 certified data products. Getting an answer out of them was still slow. Every request meant roughly an hour of hand-written SQL, spreadsheet analysis, charting, and report writing. Turnaround time capped how many questions the team could answer.
Trust was the second cost. Metric definitions were repeated across overlapping systems and drifted over time, so analysts spent effort reconciling numbers before they could share them.
Without the help of AI, request volume would outgrow analyst capacity across CX, GTM, Sales, and Success, and there would still be no governed, measured foundation to answer faster without sacrificing accuracy. Earlier attempts had run for six months across two architecture iterations without reaching a production design.
What phData did
phData embedded a two-person pod of ML Solutions Architects in the ED&A AI Engineering team. The pod covered data and semantics, agent orchestration, and evaluation and governance, and worked backward from two targets: a working system in 12 weeks or less, and 15 minutes per analysis. It built on the customer’s existing Snowflake, Atlan, and Okta platforms rather than replacing them. Five decisions shaped the delivery.
One question in, a finished analysis out
A single natural-language question now starts a multi-agent workflow. A Claude Opus orchestrator splits the question into retrieval, analysis, reporting, and insight steps, and hands each to a specialized sub-agent running on Strands and Amazon Bedrock AgentCore. Claude Sonnet handles the reasoning-heavy steps, and the lighter Claude Haiku handles analysis and insight for speed and cost. The output is a set of plots, reports, tables, and SQL. Sessions and artifacts persist, so analysts reopen earlier work rather than redo it.
A single agent would have had to retrieve, analyze, chart, and report in one pass, and one failing capability would have blocked the rest. Separate sub-agents keep each step simple and the platform reusable.
Every answer comes from governed definitions
The assistant does not write free-form SQL over raw tables. It answers from governed Snowflake semantic views and Atlan metadata and metrics, reached through MCP servers. The customer’s existing data products were reused as-is, with no new data layer, thereby shortening the path to production. The customer’s data governance team continues to own metric definitions. The pod owns the automated evaluation and readiness checks around them.
Accuracy is measured before anything ships
Speed only helps if answers stay right. The pod built a Golden Query Bank of more than 100 verified ground-truth queries across multiple data domains, and a repeatable Semantic Evaluation Workflow. Every new or refined semantic model is paired with gold prompts and expected behavior. A new semantic model or agent version ships only if it matches or beats the production accuracy on the Golden Query Bank. Testing of governed-metric grounding reached about 93%, against a deployed threshold of 90%.
Access follows each user's own permissions
Okta authenticates users. AgentCore Gateway passes a user-scoped token to Snowflake, so every query runs with that person’s own access. The assistant has read-only access to governed data and takes no write actions. Tokens are held in AgentCore Identity and AWS Secrets Manager, never in agent code. The API and the agent runtime use separate IAM identities, and all components run inside the customer’s VPC. Every agent invocation and query is logged with user ID, content, and timestamp, and agent calls are traced in AWS X-Ray and Amazon CloudWatch.
Managed services, three independent components
The platform is split into a React UI, a backend API, and the agent runtime, each deployed separately. The UI and API run on Amazon EKS to match the customer’s container platform. The agents run on Amazon Bedrock AgentCore Runtime, so session isolation and scaling are managed rather than built and operated by the team. Claude inference comes from Amazon Bedrock. Memory and artifacts live outside the pods, in AgentCore Memory and Amazon S3, so a replaced pod loses neither conversation context nor finished analyses. The managed services are multi-AZ, which fit an internal analytics assistant better than the added cost of multi-region.
An earlier Snowpark Container Services approach was set aside because it did not fit the customer’s AI platform design.
A governed agent assistant cut each analysis from an hour to 15 minutes
Time per analytics request: about 1 hour with manual work reduced to under 15 minutes, a 75% reduction
Time to a working system: 6 months of earlier iterations without a production design, 10 weeks with phData against a 12-week target
Accuracy: about 93% on governed-metric grounding, above the 90% deployed threshold
Coverage: a Golden Query Bank of 100+ verified queries now gates every release
The orchestrator-and-sub-agent pattern and the Snowflake and Atlan integrations are reusable across other ED&A AI initiatives, so the next agent starts from a working, governed foundation.
reduction in time per analysis, from 1 hour to 15 minutes
to a working system in production, against a 12-week target
accuracy on governed-metric grounding, above the 90% release threshold
Ready to see what a similar engagement could look like for your organization?
What's next
The customer is evaluating governed metric registries to replace duplicated metric definitions with a single source of truth and eliminate the rework they cause. The same platform is set to support additional use cases across CX, GTM, Sales, and Success.
More customer stories
Scaling efficiency at Ultradent through Claude training
See how phData’s hands-on Claude training helped Ultradent cut month-end close from 12–14 days to 5–7 days and expand Claude use across teams.
Frequently asked questions
What is agentic analytics?
Agentic analytics uses AI agents to carry out the steps of an analysis, not just answer a single question. Given a business question in plain language, agents retrieve data, run analyses, build charts, and write up findings. Analysts review and decide rather than assembling every query and report by hand.
How is agentic analytics different from traditional BI or a chatbot on top of data?
Traditional BI shows answers to questions someone already built a dashboard for. A basic chatbot translates one question into one query. Agentic analytics plans a multi-step workflow, uses tools, checks its own work, and returns finished artifacts such as SQL, tables, charts, and/or reports. It handles new questions without waiting in a backlog.
Why does a semantic layer matter for AI analytics agents?
A semantic layer defines metrics, relationships, and business terms once, in a governed and versioned registry. Agents that answer based on those definitions often return answers more quickly and with greater accuracy and at a lower token cost. Agents that write free-form SQL over raw tables can interpret "revenue" or "active customer" differently each time. For most enterprises, getting metric definitions right is the main driver of trustworthy AI answers.
Where should a company start with agentic analytics?
Start with one high-volume type of question on data that is already governed. Record how long those requests take today, set a target, and build the evaluation set before the agent. A narrow first use case builds trust, produces a measured result, and leaves a reusable foundation for the next team.
Ready to build a governed agent assistant for your analytics team?
Talk with a phData expert about turning your certified data products into a measured, production-ready multi-agent assistant.
- Free scoping conversation, no commitment
- Certified engineers embedded with your team
- Proven delivery from pilot to production