Do you have to clean up your CRM before using AI agents?
No. Fix what the first piece of work depends on, match duplicates instead of deleting them, and let corrections flow back to the CRM over time.
No, not first. A full CRM cleanup is slow, expensive and never finished, and most of the mess that breaks agents is not the kind a cleanup fixes. Fix the few things the first piece of work depends on, match duplicate and outdated records to one identity instead of deleting them, and let corrections flow back to the CRM once agents are allowed to write. The CRM gets cleaner because the work runs, not before it.
Why the question comes up
Almost every company that looks at AI agents looks at its CRM and winces. Records are outdated, the same customer exists three times, and contacts left years ago. On a recent call, a consultant who works with mid sized suppliers described a client with two CRMs in parallel, each team using the one it prefers, and a plan to put a copilot on top so it could search both. His instinct was to fix the basics first. The client's answer was that there is no budget for a cleanup.
Both sides have a point. The data really is bad: in Validity's 2025 survey of 602 CRM users, 76% said less than half of their CRM data is accurate and complete. In Salesforce's 2024 State of Sales survey, only 35% of sales professionals completely trust the accuracy of their organization's data. But a cleanup project that has to finish before anything else starts usually never starts.
What kind of mess breaks agents
It helps to split the mess into three kinds, because only one of them needs fixing up front.
Duplicates and outdated names. The same company spelled three ways, a supplier still recorded under a legal entity that has since been acquired, an old email address next to a new one. This is the mess people mean when they say "clean the CRM". It is also the kind a matching step handles well: records that describe the same thing are linked to one identity, clear cases by rule and unclear ones by a person who knows the customers. Nothing has to be deleted first. The method is in how to build a context graph from the systems you already have.
Missing relationships. The CRM knows the account and the deal. It rarely knows that the buyer moved to another team, that the deal waits on a supplier, or that a regulation on one call touches ten other accounts. No cleanup adds that, because it was never in the CRM. It comes from connecting the CRM to email, call notes, the ERP and outside sources.
Undefined work. Nobody agreed what a qualified lead is, which system owns the customer's address, or when a deal counts as lost. This is the mess that does need fixing before you build, because an agent cannot follow a standard nobody has written down. It is a decision, not a data job. See agree the standard before you build.
Why a copilot on top of two CRMs doesn't fix it
Pointing an AI assistant at both systems sounds like the cheap way round. It searches both, finds two versions of the customer and has no way to choose between them, so it picks one or blends them and answers with the same confidence either way. That is the pattern behind confident wrong answers: more access to messy tools does not give an agent a view of the business. The two CRMs need to resolve to one picture of each customer, with a decided source of truth per field, before an agent can act on them.
What to fix first, and what to leave
| Fix before the first agent | Leave for later, or let the work fix it |
|---|---|
| The definitions the first piece of work depends on | Old records nobody uses |
| Which system is the source of truth, per field | Duplicates the matching step can link |
| The fields the first agent reads and writes | Fields no current workflow uses |
| Who approves an agent's changes to records | A full migration to one CRM |
Start from one piece of work, such as preparing account meetings or chasing open quotes, and check only the data that work touches. That keeps the cleanup small enough to finish.
How the CRM gets cleaner afterwards
Once records are matched to one identity, the agent works from the matched picture while the CRM stays where it is. Where your team grants write access, agents can update records through scoped tools, so a merged duplicate or a corrected contact can go back to the CRM, with a person approving the changes that need one. Over time the system people use every day gets cleaner as a side effect of the work, instead of as a separate project.
How Ortelian approaches it
Ortelian builds a pipeline for each of your systems and keeps every record. Nothing is migrated. The world model holds each company, person and deal once, however many systems name it, with how they relate and the source of every fact. Agents work from that model, in the tools your team already uses, and are read only until a write is granted. Actions that reach the outside world wait for a person's approval by default, and every run, action and approval is recorded. The audit at the start maps the work and sets the standard the first build must meet. More on the platform and how we work.
Questions people ask
Should we move to one CRM before we start with AI agents?
Only if you were going to anyway. Salesforce found that 53% of sales teams that had fully implemented AI first consolidated their tech stack, so it can help. But a migration is a project of its own. Matching records across two CRMs gives agents one picture of each customer without waiting for it.
Can AI clean up our CRM data for us?
It can do much of the matching: spotting that three records are the same company, or that a contact changed jobs. The unclear cases still need a person who knows the customers, once. What AI cannot do is decide your definitions, such as what counts as a qualified lead.
What if our sales team doesn't enter data properly?
That is a process problem more than a data problem. Agents can take over much of the logging, from call notes and email, so the record stays current without asking people to type more. Decide what has to be recorded and who owns it, then let the agent do the entry.
How much data cleanup does the first agent need?
Only what its piece of work touches. If the first agent prepares account meetings, the accounts, contacts and open deals it reads need to resolve to one identity. The rest of the CRM can wait.
Sources
- Validity, The State of CRM Data Management in 2025, July 2025
- Salesforce, State of Sales, 6th edition, 2024
- Ortelian, Platform, Trust and How we work
