How Batteries Plus Built 10 Agents to Activate 100,000 Sales Prospects

Most enterprise AI projects fail not because the model underperforms, but because the organization cannot feed it. A recent Salesforce case study on Batteries Plus—a 700-store franchisor that deployed ten AI agents to activate 100,000 dormant B2B prospects—illustrates a pattern we see repeatedly: the technology worked, but only because years of data discipline preceded it. For the majority of mid-market companies attempting similar deployments, that foundation does not exist.

This piece is for operations and IT leaders who have been pitched agentic AI as the next unlock for their pipeline, customer service, or back-office automation. Before you sign the SOW, understand where these projects actually break—and what the successful minority does that you probably are not doing yet.

The Gap Between Demo and Production

Every vendor demo shows agents reasoning across systems, pulling customer history, triggering workflows, and surfacing insights. The demo works because the demo environment has clean, connected data. Your environment does not.

In a typical mid-market company with 100–2,000 employees, we find:

• Customer records duplicated across three to five systems with no reliable unique identifier
• Product and pricing data maintained in spreadsheets that shadow the ERP
• Historical transaction data archived in formats the modern stack cannot query
• Contact information decayed 20–30% annually with no systematic refresh

Agentic AI does not fix this. It exposes it. The first thing a well-built agent does is try to retrieve context—and when that retrieval returns garbage, the agent either hallucinates or fails gracefully. Neither outcome gets you pipeline.

What the Successful 20% Do Differently

The Batteries Plus deployment worked because the company had already invested in data unification before the AI project started. They had a single source of truth for B2B prospects, enriched with firmographic and behavioral data, accessible through APIs their agents could call. Most mid-market companies have not made that investment—and underestimate what it costs.

They Treat Data Readiness as Phase Zero

Organizations that succeed with multi-agent deployments spend 40–60% of the total project timeline on data preparation. Not integration. Not prompt engineering. Data cleaning, deduplication, enrichment, and access layer construction. If your vendor proposal shows AI development starting in month one, ask what assumptions they are making about your data.

They Scope Narrowly Before Scaling

Ten agents sounds impressive. But the Batteries Plus case involved a specific, bounded use case: reactivating dormant B2B accounts with personalized outreach. The agents had a defined input (prospect records), a defined action space (email, SMS, task creation), and a defined success metric (meetings booked). Contrast this with the typical enterprise AI pitch: “We’ll deploy agents across sales, service, and operations.” That scope guarantees failure.

They Measure What Matters Before Launch

Before deploying agents, successful organizations establish baseline metrics for the process they are automating. How many prospects does a human SDR contact per day? What is the current meeting-booked rate? What does a reactivated account cost in human labor? Without these baselines, you cannot calculate ROI—and you cannot tell whether the agents are working or just generating activity.

The Hidden Costs That Do Not Appear in the Proposal

Vendor proposals for agentic AI typically include licensing, implementation services, and maybe some training. They rarely include:

• Data remediation: Cleaning and unifying the records your agents will query. Budget 3–6 months and $50,000–$200,000 depending on your data sprawl.
• Integration middleware: Building the connectors and APIs that let agents access your actual systems, not just the CRM. Add 20–40% to the implementation estimate.
• Ongoing supervision: Someone needs to monitor agent outputs, catch failures, and tune prompts. This is not a set-and-forget deployment. Plan for 0.25–0.5 FTE ongoing.
• The rebuild in year two: Agent architectures are evolving rapidly. What you build on Agentforce, Copilot, or any platform today will likely require significant rework within 18–24 months as capabilities and best practices shift.

When you add these costs to the vendor proposal, the payback period typically extends from “six months” to “eighteen months, if adoption holds.” That changes the risk calculation.

Where to Start Instead

If you are evaluating agentic AI for your organization, resist the urge to start with the technology. Start with the data.

Run a data readiness assessment against your top three use cases. For each use case, answer:

• What data does the agent need to retrieve to act?
• Where does that data live today, and how fragmented is it?
• What is the current accuracy and completeness of that data?
• Can the data be accessed programmatically, or does it require manual export?

If you cannot answer these questions with confidence, you are not ready for agents. You are ready for a data unification project—which may still be the right investment, but it is a different project with a different timeline and budget.

The Batteries Plus case is instructive not because it shows what AI can do, but because it shows what AI requires. Ten agents activating 100,000 prospects is a compelling headline. The years of data work that made it possible is the actual story.

Multi-agent AI will eventually become table stakes for mid-market operations. The companies that benefit first will not be the ones who bought the most advanced platform—they will be the ones who did the boring, expensive work of getting their data house in order before the agents moved in. If your 2025 roadmap includes agentic AI, your 2024 should have included data unification. If it did not, adjust your timeline accordingly.

The model will work. What fails is everything around it.