Why Your Salesforce Data Isn't Ready for AI Agents
- AI
- 8 min
You’ve started evaluating Agentforce, or Copilot, or one of the dozen AI tools now plugged into your GTM stack. The demo looked great. The pilot got greenlit. And somewhere in week three, things started going sideways. Wrong recommendations, missed context, an agent confidently citing a contact who left the company eight months ago.
Before you conclude the AI isn’t ready, it’s worth asking a different question: Is your data ready for AI?
Get our latest insights into your inbox
The Problem isn't the Model
Across the Salesforce ecosystem right now, a consistent pattern is emerging in post-mortems on stalled AI deployments. It is rarely the algorithm. A widely cited industry estimate puts the figure starkly: 88% of enterprise AI agent pilots fail to reach production, not because the agents are weak, but because the CRM data underneath produces confidently wrong outputs at scale. That’s a different failure mode than what most teams plan for.
Bad data has always been a CRM annoyance. Duplicate records, an outdated phone number, a stale job title. Humans navigate around these problems instinctively. A sales rep glancing at an incomplete contact record fills in the blanks from memory. A sales manager catches an obviously wrong forecast before it reaches the board deck.
AI agents don’t do that. As one analysis of Salesforce data quality puts it, garbage in, garbage out was the old principle. The 2026 version is sharper: garbage in, confidently wrong out. Agents do not pause to verify a stale record the way a human would. They act on it, then propagate the action across thousands of records before anyone notices.
Salesforce’s own product marketing has converged on the same message. As the company’s Tableau product marketing director put it, an AI strategy without a data framework is just a wish list. Attempting to deploy AI agents without one leads to inconsistent results, security risks, and a lack of user trust.
What "data readiness" actually means
It’s tempting to treat data readiness as a vague hygiene goal. “Clean up the CRM” without a concrete definition. Salesforce’s own guidance on the topic is more precise, and worth using as a working checklist before evaluating any agent deployment:
Is your data unified and harmonized? If your data is fragmented across Sales Cloud, Service Cloud, spreadsheets, and a dozen point tools, the agent will deliver fragmented and inconsistent experiences. Unification isn’t optional. It’s the precondition.
Have you resolved identities and is the information current? The same contact often exists as three different records: full name, abbreviated name, email-only. And each one tells the agent something slightly different. Old, incorrect data leads to frustrating experiences for customers and unreliable outcomes, including outright hallucination.
Do you have governance and security in place? An agent should only access the data it needs to do its job, and that access needs to be auditable.
Can you activate the data in real time? Data sitting in a warehouse, updated weekly, doesn’t power an agent that needs to act now.
Is there a feedback loop? Agents need humans in the loop checking whether they’re acting on the right information, not a “set and forget” deployment.
Separately, a widely referenced breakdown of what “good” CRM data looks like for AI purposes narrows it to three properties: data needs to be complete (the full picture, not partial context), structured, and effective for the specific task the agent is meant to perform. Without completeness, AI models miss vital context: what stage a contact is at, what previous interactions occurred, who else is involved in the decision.
The numbers behind the problem are larger than most teams expect
This isn’t an edge-case concern. Recent industry data paints a fairly stark picture of how unprepared most enterprise data actually is for agentic AI.
Fewer than one in five companies has a high level of data readiness, and only 9% are fully prepared for the data integration and interoperability that AI requires, according to a 2025 Capgemini report on AI agents.
A separate analysis found that 81% of companies say fragmented data is preventing them from unlocking AI’s potential. Service agents miss complete customer histories, sales agents miss signals because marketing interactions aren’t visible, and analytics agents produce unreliable insights that undermine decision-making.
The trust problem compounds this. Industry surveys cited by Salesforce found that nearly six in ten AI users say it’s difficult to get what they want out of AI right now, with over half saying they don’t trust the data used to train the systems they’re working with. Separately, a survey found that 90% of high-level data professionals believe company leadership isn’t paying enough attention to bad or inadequate data, even as AI initiatives accelerate. Only 9% of organizations report fully trusting their data which directly affects their confidence in CRM reporting.
The forecasting impact is direct and measurable. Inaccurate forecasting tied to poor data quality affects a meaningful share of sales organizations, and several industry analyses tie data quality directly to financial loss. Duplicate or incomplete customer records cause missed opportunities, double-booked engagements, and wasted marketing spend when AI-driven outreach unknowingly targets the wrong contacts or duplicates effort.
Why this is structurally different from past CRM problems
Traditional CRM issues included duplicate records, missing fields, outdated contact info. A salesperson could work around a few mistakes in a report. A direct mail piece sent to an old address was a minor, contained error.
When an AI agent built on top of that same data starts making autonomous decisions, the stakes change entirely. The agent doesn’t know it’s working from a flawed record. It acts with full confidence on whatever it’s given. The moment AI starts acting on it, a small inaccuracy in CRM data gets magnified, not corrected.
This is also why simply buying a better AI agent product doesn’t solve the underlying issue. As one technical breakdown of Salesforce AI failures put it plainly: it’s not the algorithms, it’s not the lack of ideas. It’s the data. When Salesforce data is inconsistent, poorly documented, or siloed across organizations, AI doesn’t have the clarity and context it needs to learn and reason, regardless of which agent platform sits on top.
The three data problems that most commonly break enterprise agent deployments before they reach production are consistent across analyses: duplicate account and contact records, stale or unstructured contact data, and incomplete account hierarchies. Specifically, incomplete buying committee and opportunity contact role data, which leaves agents working from a fraction of the actual deal context.
What good looks like in practice
It’s worth being concrete about the gap between “data exists in Salesforce” and “data is usable by an agent.” In one detailed account of fixing this problem for an Agentforce deployment, service agents that were routing cases incorrectly during testing reached over 90% accuracy in production after a focused data governance and unification pass without any change to the agent configuration itself. The AI didn’t improve. The foundation under it did.
That single example captures the core insight worth taking away from all of this research: the fix for an underperforming AI agent is very rarely a better model, a different vendor, or more sophisticated prompting. It is almost always the data the agent was given to work with in the first place.
Where Nektar fits
Most data quality problems in Salesforce don’t originate from bad intent. They originate from the fact that customer interactions happen across email, calendar invites, calls, and meetings. And none of that activity reaches the CRM unless someone manually logs it. Reps forget. Meeting attendees go uncaptured. A buying committee member who’s been on four calls never gets created as a contact, because creating that contact was never anyone’s job.
A governance project can clean up what’s already in Salesforce. It can deduplicate records, standardize fields, and resolve identities. What it can’t do is keep the system populated with the engagement signal that’s happening outside of Salesforce every single day, for every rep, on every deal.
Nektar sits between your engagement systems and Salesforce, automatically capturing every customer interaction, resolving identities, mapping contacts to the correct accounts and opportunities, and continuously writing structured, complete data into the CRM, without requiring a single rep to change their behavior. It also backfills historical interactions, so when a rep leaves or a deal changes hands, the institutional context doesn’t disappear with them.
This is the distinction worth drawing for any team currently evaluating Agentforce or Copilot and starting to worry about data quality: cleaning your CRM once is necessary but not sufficient. The data needs to stay complete, continuously, as new interactions happen. Because an AI agent making decisions next quarter needs the same level of completeness as the one making decisions today.
Salesforce is the system of record. Nektar is the system that ensures the record actually reflects reality. So that whatever AI agent you deploy on top of it, today or eighteen months from now, has something reliable to act on.
Enjoyed our content? Follow Nektar on LinkedIn
See it in production
Mimecast’s VP of AI Acceleration built 8–10 AI applications on top of Nektar’s data layer generating $2M in directly attributed expansion revenue and surfacing $80M+ in additional pipeline within 80 days
