Marketing and revenue operations teams spent the past year racing to deploy AI agents into their CRM workflows. New survey data shows most of them skipped a step: making sure the data those agents act on is actually trustworthy.
What happened
LeanData, which builds go to market orchestration software used by more than 1,000 B2B companies, surveyed 157 revenue operations, marketing operations and sales practitioners in May 2026 for its AI GTM Customer Survey. The results, published on the company’s own newsroom, found that 79% of respondents were already scaling or deploying AI agents inside their go to market motions, while 70% said poor data quality was undermining the results those agents were supposed to deliver.
The breakdown of causes points squarely at infrastructure rather than the AI itself. Asked why AI initiatives stall or underperform, 45% of respondents blamed inconsistent or incomplete CRM data, 37% pointed to undocumented business processes, and 32% cited siloed teams. None of those three root causes is a shortcoming in the underlying model or agent. They are the plumbing problems that surface only once an agent starts acting autonomously on records a human used to eyeball first.
“AI models are now both powerful and incredible. Yet, we all know they work only as well as the infrastructure that lies beneath them,” said Katy Keim, CEO of LeanData. “Executives told GTM teams to move fast on AI, and they did. Now those same teams are reporting that their data is not clean, their business processes live in people’s heads and everyone is thin-slicing different views of the customer.”
The fear is control, not capability
The most telling number in the survey is not about performance at all. Sixty percent of respondents said they were afraid their AI agents would act on inaccurate data or violate ownership rules, and when asked what they wanted most instead of more AI capability, 31% said a complete audit trail of every action taken on every record. That is a governance request, not a technology request, and it lands differently than the usual martech complaint about model quality.
LeanData’s own response arrived on September 17, when it shipped a platform release built explicitly around that gap rather than around adding another autonomous agent to the stack. The centerpiece is an AI Inference Node that lets administrators call an AI model in the middle of an existing routing flow to classify, extract, summarize or infer attributes from a record, then route on the result, instead of replacing the deterministic rules teams already trust with a fully autonomous agent making the call end to end. A companion AI Assistant inside the company’s FlowBuilder tool lets admins describe what they want in plain language and get help building or debugging the routing logic itself.
“As GTM becomes increasingly AI agent-driven, LeanData’s role is to make AI outcomes fast and trusted,” Keim said of the release. “This release makes AI a building block, not a tool bolted onto the stack. Ops teams keep control and gain simplicity in how they administer their workflows. Everyone’s racing to ship the flashiest agent. We’d rather fix what’s happening under the hood, because that’s what decides whether the AI actually works.”
Early customer use backs up the pitch. “The AI Inference Node in LeanData has been an incredible tool for our team,” said Linzy Cote, revenue operations manager at Traliant, a workplace compliance training company. “It has reduced the amount of manual work needed to research incoming leads, review cases and assign them out.”
A pattern beyond one vendor
The specifics belong to LeanData, but the shape of the problem does not. Marketing and sales stacks across the industry have spent 2026 layering agentic features onto systems that were never built to be acted on autonomously, and adding another agent on top of an already sprawling tool stack has repeatedly turned out to compound the coordination problem rather than solve it. The complaint LeanData’s customers are voicing, that nobody can see how an agent’s decision actually got made, is the same trust gap that has pushed some publications on this beat to argue agents need auditable trust layers before they need broader access to systems of record.
What it means for the marketing leader
The practical takeaway is not to slow down AI adoption. It is to audit what the agent is standing on before extending what it is allowed to touch. A marketing or RevOps leader evaluating an AI agent vendor in Q4 should be asking for the same thing LeanData’s own customers said they wanted: a record-level audit trail, a documented process for what the agent is permitted to change, and a way to trace a routing or scoring decision back to the data that produced it. Vendors that can answer those questions concretely, rather than with a roadmap slide, are the ones likely to survive the next round of budget scrutiny once the current wave of agent pilots reports back on actual data quality incidents.
The 60% figure on fear of ownership violations also has a procurement implication. Any AI agent that writes to a CRM or MAP record, not just ones marketed as “autonomous,” should come with a clear answer to who owns a field when the agent and a human disagree about it. Teams that wrote that governance question into their RFPs before this survey came out are, per LeanData’s own numbers, still a minority.
How to evaluate it
Before adding the next agent, run an audit of the data it will read from and write to: how current, how documented, and how siloed. LeanData’s finding that 55% of respondents named bad data and unreadiness as their main blocker, more than any complaint about the AI layer itself, suggests that budget spent on data hygiene and process documentation will outperform budget spent on a flashier agent for most teams still working through this gap.
Source: LeanData
