The data layer decides what agentic AI can and can’t do

Article    September 09, 2026
The data layer decides what agentic AI can and can’t do
SHARE
BOTTOM LINE UPFRONT

Agentic AI inside a portfolio company is a data governance project before it is anything else. Salesforce alone has spent billions on that layer, and finance teams that fund the agent first end up paying for the data work twice.

Salesforce’s ambition reaches past the CRM. It wants to be the backend data and logic engine every enterprise AI agent runs on, delivered through Agentforce, regardless of which front-end tool a customer prefers to use. 

The company is backing that ambition with real money: $8 billion for Informatica, $3.6 billion for Fin to add autonomous customer service, and a pending deal for Contentful valued between $1 billion and $1.5 billion, its three largest AI acquisitions since 2025. Its partnerships with Anthropic and OpenAI extend the same strategy in a different direction, meeting users inside the tools they already work in through Headless 360 and the Model Context Protocol. 

Every one of those bets lands on the same layer: the data underneath the agent. 

How much does a fragmented data layer really cost? 

An agent that catches a bad invoice before quarter close has to connect the quote, the contract terms, the billing event, and the expected revenue as one record. That only works when every system describes the same customer and the same deal the same way. When the sales system, the billing system, and the accounting system disagree on what a contract is, the agent has nothing consistent to work from, and it still gives an answer with no sign anything was wrong. 

We worked with a fleet safety software provider that learned this directly. Moving a customer from a free trial to a paid plan meant creating separate sale and contract records. When those disagreed, billing sent out a $0 invoice. An agent reading the billing system alone would have called it valid, because within that system it was. The error lived in the gap between two systems nobody had connected. 

Multiply that gap across every step between signing a deal and getting paid, and the costs show up everywhere: manual reconciliation, billing leakage nobody catches until an audit, and a revenue figure that ships with a caveat. A unified data layer catches the error the moment it happens, not the quarter someone finally goes looking for it. 

How do you get your systems to agree? 

Readiness requires every system involved to describe the same commitment in the same terms before an agent acts on it. That level of agreement gets built during consolidation, and the earlier it happens, the less an agent inherits gaps nobody has closed yet. 

It comes down to four steps: 

  • Assign one system as the source of truth for the revenue event. Every other system references that definition instead of maintaining its own. 
  • Put controls upstream of the agent. That same fleet safety provider fixed its billing gap by holding orders the moment conditions fail validation, so the bad invoice never gets created in the first place. 
  • Join operational and financial records. An agent with CRM visibility alone can miss what an auditor cares about, and that gap turns an AI finding into an audit finding. 
  • Leave a traceable record on every output. An output nobody can trace becomes a manual investigation, which is exactly the cost the agent was bought to remove. 

An Irvine-based cloud video surveillance provider shows what it looks like once all four are in place. Every platform was reachable long before the company retired its manual transaction checks. The fix came from consolidating the operational model and integrating revenue recognition and tax into it, so every system finally described the same commitment the same way. Only then did the manual checks come off the books. 

Which means: get the order of operations wrong and the pilot stalls in finance review. Nobody signs off on output they cannot trace, so the reviewer ends up rebuilding by hand what the agent was meant to automate. Get it right and the agent extends something already earning its keep: a faster close, and an ARR figure that holds up in diligence without a workbook behind it.

Salesforce’s spending answers the question 

Billions across acquisitions and partnerships all point at the same conclusion: the data model underneath the agent is what determines whether its output can be trusted, traced, and paid for. 

For a PE-backed CFO, the read is the same. The data foundation is the AI initiative, funded and evaluated as one line item. It is what turns agentic AI spend into ROI a buyer will pay for.

We’re taking this argument to Dreamforce. On Tuesday, September 15, Accordion and Salesforce are hosting a session on how billing and revenue recognition outside the CRM block AI from doing useful work, with Mike Aaron of Salesforce Revenue Cloud and Jagan Reddy of RightRev. Seating is limited and the session is invite-only. 

RSVP FOR THE BREAKOUT SESSION

FAQ

Why is agentic AI really a data governance problem, not a technology problem?

Because an agent is only as reliable as the agreement between the systems feeding it. When the sales system, billing system, and accounting system disagree on what a contract is, the agent has nothing consistent to work from — and it will still produce a confident answer with no sign anything was wrong.

Can you give a real example of what happens when systems don't agree?

Yes — a fleet safety software provider moved a customer from a free trial to a paid plan by creating separate sale and contract records. When those records disagreed, billing sent out a $0 invoice. An agent reading the billing system alone would have called that invoice valid, because within that one system, it was.

What are the four steps to get systems ready for agentic AI?

Assign one system as the source of truth for the revenue event, put controls upstream of the agent so bad records never get created in the first place, join operational and financial records so an agent doesn’t miss what an auditor cares about, and leave a traceable record on every output so nothing becomes a manual investigation.

What happens if a company deploys an AI agent before fixing its data foundation?

The pilot stalls in finance review. Nobody signs off on output they can’t trace, so the reviewer ends up manually rebuilding what the agent was supposed to automate — meaning the company pays for the data work twice, once to build the agent and again to fix what it got wrong.

Is your data layer ready for agentic AI? Let's talk.

Our contact form is currently blocked by your cookie preferences. Please change your preferences to continue.