# Governed-data agent pilot

We build a working AI agent that answers business questions from your governed data, and we measure its accuracy and its model cost per answer.

- **Scope:** One business area
- **Price:** From $45,000
- **Typical duration:** 3 weeks
- **Delivery:** Remote, in your environment

## What you get

We work with your team to pick one business area, such as finance or operations, and the questions people there need answered. The agent works with your data where it is today, and if you don’t have a data platform yet we can set one up as part of the work. At the end you have the agent running in your environment, an evaluation you can run again, and a written readout of how it did and what production would take.

- An agent deployed in your environment, built as a custom Claude agent or on a platform’s own agent service, such as Snowflake Cortex or Databricks Genie
- An answer key of 40 to 60 business questions, with the correct answers agreed on by your business owner
- An evaluation that records which answers the agent got right and wrong against that key
- Measured model cost per answer for the configuration tested
- A written readout with the findings and recommended next steps

## How the result is measured

Before the agent is tested, your business owner agrees on the questions and the correct answer to each one. The agent then answers all of them, and the evaluation records each answer as right or wrong along with the model cost of producing it.

We go through the wrong answers with you and group them by cause, such as a metric with two definitions in the source data, a table the agent couldn’t reach, or an ambiguous question. The readout lists the causes and what fixing each would involve.

The numbers describe the agent on that answer key, in the configuration tested. A production decision also depends on your security review, your operating plan and an owner’s approval.

- [View a sample pilot readout](https://ironcondorconsulting.ai/sample-readout)

## When it fits

- Teams with governed data in a data platform, such as Snowflake, Databricks or DuckDB, and a business area that wants answers from it
- Teams whose last AI pilot ended in a demo and no decision
- Teams that need to know how often an agent is right and what each answer costs before they ask for a production budget

The starting price assumes the data is already governed and in a platform the agent can reach. If it isn’t, we can do that work too, for example setting up Snowflake, Databricks or DuckDB and modeling the data with dbt, and we’d include it in the proposal. The pilot also needs a business owner who can agree on the correct answers.

## What we need from you

- Data for one business area, in a data platform or in the systems where it’s kept today
- A business owner who can agree on the questions, confirm the correct answers and review the findings
- Access to a non-production environment, or agreement that we set one up

## Working in your environment

The agent and the evaluation run inside your own cloud account, under the access your team approves. The model services that process the data are agreed on during scoping.

## After the engagement

You receive the agent, the answer key, the evaluation and the readout. From there you can extend the agent, prepare it for production with your own team or with us, and run the evaluation again whenever the data or the model changes.

## Questions

### Which platforms do you build AI agents on?

We build custom agents with Claude that work with the data platform you have, such as Snowflake or Databricks. We also build on a platform’s own agent service, such as Snowflake Cortex or Databricks Genie, when that suits you better. In both cases the pilot delivers a working agent in your environment, with its accuracy and model cost per answer measured.

### Do you build custom Claude agents?

Yes. We build custom agents with Claude, along with the tools they call and the evaluation that measures them, and we deploy them in your environment.

### What if we don’t have a data platform or governed data yet?

We can set that up as part of the engagement. That can mean a new Snowflake or Databricks account or a lighter option such as DuckDB, with the data modeled in a tool such as dbt. It adds scope, and our proposal prices it.

### What does a governed-data agent pilot cost?

The pilot starts at $45,000 for one business area. Our proposal gives a fixed price before work starts. Any later change in scope is agreed on in writing first.

### How long does a governed-data agent pilot take?

The pilot typically takes three weeks from kickoff, once the prerequisites are in place.

### What if the agent does poorly on the answer key?

The readout reports the score and groups the wrong answers by cause, so a low score still shows you what needs fixing.

### Can we run the evaluation again after the pilot?

Yes. The answer key and the evaluation stay with you, so you can run them again when the data, the prompt or the model changes and compare the results with the pilot.

Tell us which business area you have in mind and where its data lives. [Discuss a project](https://ironcondorconsulting.ai/contact)

---

Source: https://ironcondorconsulting.ai/services/governed-data-agent-pilot
