Scaling Analysis with AI Agents: How One Team Cut Manual Work by 75 Percent

What if you could turn a time-consuming, slide-by-slide manual analysis into a repeatable workflow that generates dealer-ready data stories in minutes?

That's exactly what FordDirect and OneMagnify accomplished. Brendan Sullivan, director of advertising analytics at FordDirect, and Stephanie Butterbrodt, BI consultant at OneMagnify, shared the playbook in their Domopalooza 2026 session, The Dealer Winback Hack: Scaling Analysis with AI Agents in Domo. The session walks through how the team replaced a labor-intensive, dealer-by-dealer performance review with a governed AI agent workflow that equips sales reps with consistent, compelling narratives to re-engage terminated or at-risk dealers.

The results speak for themselves: a 75 percent reduction in manual work hours, 90 percent faster turnaround from analysis to story generation, and 10 dealers successfully won back. Whether you're exploring AI agents for sales enablement, operational analytics, or customer engagement, this session offers a practical blueprint you can adapt to your own workflows.

Build an AI-ready data foundation first

Pre-aggregation and standardization drive speed and consistency

AI agents work best when they're not doing heavy lifting on calculations or aggregations. The team emphasized that data readiness and pre-aggregation were key enablers for speed, accuracy, and consistency in AI agent outputs. They moved from unstructured Excel narrative text to standardized monthly metrics, running everything through an ETL (extract, transform, load) pipeline that performs calculations and pivots before the agent ever sees the data.

As one presenter explained, "Ultimately, the goal was that we wanted to get to show one row per dealer [...] And we weren't relying on the AI agent to do those calculations or to do the aggregations."

This approach reduced inconsistency and made QA straightforward. When the agent receives clean, pre-aggregated data, it can focus on analysis and narrative generation rather than wrestling with messy inputs.

Here's how you can apply this principle to your own AI agent projects:

• Run deterministic calculations (aggregations, pivots, filters) in your ETL before passing data to the agent.
• Structure your output so each entity (dealer, customer, account) appears as a single row.
• Standardize formats and bring in all relevant metrics at a granular level (such as monthly) to support flexible future analysis.
• Use your ETL output both to feed the agent and to store agent results for downstream dashboards.

This foundation doesn't just improve agent reliability. It also makes it easier to QA results and adapt the workflow for new use cases.

Scale multi-entity analysis with workflow orchestration

A looping pattern keeps each AI run focused and accurate

Scaling analysis across dozens or hundreds of entities introduces complexity. The team tested several approaches, including pushing all dealers through the agent at once and using a multi-agent setup. Neither delivered consistent results.

The solution: a looping workflow pattern that assigns an index to each dealer and routes them through the AI agent one at a time. After each dealer is analyzed, the results are appended back to a dataset, and the ETL refreshes to update dashboard outputs.

This pattern offers several advantages:

• Each agent run stays focused on a single entity, reducing the risk of errors or hallucinations.
• Results are stored incrementally, so you can track progress and troubleshoot issues.
• The workflow supports both batch runs (all dealers) and ad hoc runs (single dealer), giving you flexibility as business needs change.

One presenter described the end result: "So within a matter of minutes, a user can request a story for a dealer, and it generates on a dashboard for them to look at."

If you're building multi-entity AI workflows, consider structuring your orchestration so each run is self-contained. This makes it easier to scale, debug, and extend the workflow over time.

Use guardrails and examples for consistent outputs

Prompt engineering is an artifact, not an afterthought

AI agents can produce inconsistent or off-target results if you don't give them clear instructions. The team treated prompt structure, examples, and guardrails as control mechanisms for consistent, business-aligned narratives.

Their agent instruction design broke down into five key areas:

• Context and role framing: Tell the agent what role it's playing and what the goal is.
• Data and tools: Specify which datasets and tools the agent should use.
• Training examples: Provide sample outputs so the agent understands the expected tone, structure, and content.
• Task logic and guardrails: Define prioritization rules (for example, surface high-impact stories first), story limits (three stories per dealer), and edge cases (dealers with no story or improved performance).
• Restrictions: Ensure the agent only uses provided data and avoids assumptions or hallucinations.

This structured approach reduced variability and made it easier to QA outputs. The team also kept humans in the loop at the point of external consumption. Sales reps review the dashboard in real time during dealer conversations and can flag anything unusual back to the analytics team.

As one presenter put it, "So this is the final step where we really left the human in the loop."

If you're deploying AI agents for customer-facing or high-stakes workflows, consider these practices:

• Write your prompt instructions as a documented artifact, not a one-off experiment.
• Include examples that reflect the tone and structure you want.
• Set explicit guardrails to prevent the agent from straying beyond your data.
• Build in a feedback loop so people can validate outputs before they reach external audiences.

What you can take away

The FordDirect and OneMagnify case study offers a clear path for anyone looking to scale analysis with AI agents. The core principles apply regardless of your industry or toolset:

• Pre-aggregate and standardize your data so the agent can focus on analysis, not calculation.
• Use a looping workflow pattern to scale multi-entity analysis while keeping each run focused.
• Treat prompt design as an engineering artifact, complete with examples, guardrails, and restrictions.
• Keep humans in the loop at the point of external consumption to maintain trust and catch edge cases.
• Quantify your impact in both business outcomes (winbacks, revenue) and operational metrics (time savings, resource reallocation) to justify expansion.

The team reported running 97 dealer stories, generating 24 opportunity stories, and winning back 10 dealers. They also reallocated two team resources to other high-priority projects. Those numbers make a strong case for investing in AI-ready data foundations and governed workflows.

Scaling analysis with AI agents doesn't require a massive team or a blank-check budget. It requires a thoughtful approach to data, orchestration, and governance, and a willingness to iterate as you learn.

Ready to see the full breakdown? Watch the session on Domo's Domopalooza resources page to get the complete story, including workflow diagrams, agent configuration details, and Q&A with the team.