Everyone is talking about AI agents right now. I see it in every conversation I have with finance leaders, every conference I attend, every newsletter that lands in my inbox. Agents this, agents that. And I get it — the demos are impressive.
But here’s what nobody is talking about: what actually makes an agent work.
A few days ago the Wall Street Journal reported that Mark Zuckerberg — CEO of Meta, one of the companies literally building the large language models that power these agents — is constructing a personal AI agent to help him do his job. The agent retrieves answers he’d normally have to go through layers of people to get.
And the first thing I thought when I read that wasn’t “cool.” It was: why doesn’t he already have this?
That question is the whole lesson.
The Real Bottleneck Isn’t the AI
Writing a prompt is easy. Getting an AI to give you a good generic answer is easy. I can spin up an agent in 20 minutes that impresses a room full of executives — give it a focused data source, a clear set of questions, and it performs beautifully.
But the moment you try to scale that — make it useful across an entire organization, across dozens of different people asking dozens of different questions against dozens of different data sources — everything falls apart. Not because the AI isn’t good enough. Because the data isn’t ready.
I’ve seen this firsthand. Working with a client on a go-to-market story, we fed an agent 60-plus pages of data covering their entire revenue stack — pipeline, expansion, retention, the works. The insights it produced were genuinely useful. Conclusions I wouldn’t have reached as quickly on my own. But that only worked because we did the hard work of assembling and structuring the underlying data first. Without that? The agent would have given us confident-sounding nonsense.
Three Things Every Agent Needs from Your Data
When I think about what separates organizations that will get real ROI from AI agents versus those that won’t, it comes down to three things.
1. Real-Time Queryability
A dashboard can tell you “last refreshed 3 days ago” and that’s fine. An agent can’t. When a CEO asks a question and the agent quietly returns an answer based on month-old data — without flagging it — that’s a problem. The data doesn’t have to be live to the millisecond, but it needs to be stored somewhere the agent can access it reliably and on demand.
2. A Unified Data Model
This one trips up almost every organization I work with. What does “revenue” mean? Last 12 months? What’s the cutoff? What’s included or excluded? What does “retention” mean — gross or net? Logo or dollar? If two systems define the same concept differently, the agent will either ask for clarification every time or — worse — silently make an assumption and give you the wrong number with full confidence.
3. Access Controls and Data Lineage
This is the one people skip most often — and the one that matters most. A lot of enterprise security today is based on friction. Making things hard to access is the security. When I was at PwC, I could technically pull mass data from certain systems — but you had to know exactly how to do it, and nobody expected you to do it 3,500 times. An LLM connected to the same system can do that in two seconds.
The answer isn’t to trust the agent to figure out what it should and shouldn’t see. The answer is hard access controls — the agent is only able to retrieve what it’s explicitly permissioned for, full stop. It’s more work to build. It’s worth it.
What This Means for Finance Teams
Here’s the direct implication: if your organization is evaluating AI tools and nobody is talking about data infrastructure, the business case is incomplete.
The AI license is not the biggest cost in this equation. The data readiness work is. And finance teams — who are among the heaviest consumers of fragmented, multi-source data in any organization — are in the best position to lead this conversation. Audit trails, access controls, metric definitions: these are core finance competencies. They map directly to what AI agents need.
The CFO who has invested in a clean, connected data layer isn’t just running a better finance function. They’re building the foundation that makes every AI investment in the company more valuable.
Key Takeaways
- AI agents are only as good as the data infrastructure underneath them — this is why even Meta’s CEO is still building toward a fully functional agent
- The three pillars of agent-ready data: real-time queryability, a unified data model, and hard access controls with lineage
- Any AI business case that doesn’t include a data readiness budget is missing the biggest cost in the equation
- Finance teams are uniquely positioned to lead data infrastructure conversations — the skills already exist, they just need to be applied to AI readiness
- Clean, connected data is no longer just a reporting requirement — it’s the competitive moat that determines how much value your organization extracts from AI
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply