Claudeforce, the expanded partnership between Salesforce and Anthropic, was announced alongside Salesforce’s Q2 results at the end of August and has been centre stage at Dreamforce this week. Most of the commentary has, understandably, focused on the AI. The data underneath it deserves equal attention, because that is where much of the real implementation effort will sit.
What Claudeforce actually is
The partnership runs in two directions. Salesforce in Claude is a plugin for Claude Cowork with 37 prebuilt sales skills, including meeting preparation, deal health, pipeline review, email drafting and record updates. It is underpinned by AIforce, Salesforce’s harness for exposing business data and workflows to agents through MCP servers, APIs and CLI tools. Actions are routed back through Salesforce, so existing permissions and business rules still apply. It is currently with pilot customers, with an open beta due this month.
Claude in Salesforce makes Claude available as a reasoning model within the platform, including through Agentforce.
Both sit within Salesforce’s wider Headless 360 direction: exposing platform capabilities to authorised agents through APIs and MCP, rather than solely through the Lightning UI.
The implication is straightforward. When AI becomes the interface to Salesforce, the quality of the data becomes part of the interface too.
The workarounds stop working
Anyone who has spent time in a mature org knows that users quietly compensate for the data. Sellers know that two Accounts are really the same company, that the email on a Contact is two jobs out of date and that a parent–child relationship was never set up in the Account hierarchy. They work around it. An AI answering on their behalf has no access to that knowledge.
Take a meeting brief. Claude pulls together the account, opportunities and recent activity for “Acme Ltd”. Meanwhile, the open escalation and the renewal conversation sit against “ACME Limited”, a duplicate created by a web-to-lead process two years ago. The seller walks into the meeting unaware of either. The real risk isn’t that Claude visibly fails. It’s that Claude produces a fluent, confident brief that masks the ambiguity. When a user looks at three records in the UI, they can at least see the problem. A polished summary removes that cue.
A capable model may spot contradictions and infer likely answers. But some questions are governance decisions, and no model can reliably make them on the organisation’s behalf if the organisation hasn’t made them itself:
- Which records represent the same customer?
- Which system owns a given attribute?
- Which value survives when sources disagree?
- When should a match go to a human for review?
Permissions are not the same as trusted data
Credit where it’s due: Claudeforce’s governance model is sound. Salesforce in Claude respects the user’s permissions and because actions go through Salesforce, sharing, validation rules and Flows still apply. That governs what the AI can see and do. It doesn’t tell the AI whether there should be threeAccounts for one customer, which of them is authoritative or which of two conflicting phone numbers to trust.
Enterprise AI needs two layers:
- Platform and action governance controls what the AI can see and do.
- Data governance controls which customer identity, attributes and relationships the AI should trust.
Claudeforce strengthens the first layer. The second remains the customer’s responsibility.
The same applies to Claude in Salesforce. Better reasoning doesn’t compensate for unresolved data and more context is not the same as more certainty. More conflicting context simply creates more ambiguity.
Headless 360 also extends the reach of your data beyond the Salesforce UI.
Well-governed data now improves more experiences. Poorly governed data now travels further.
Deduplication is the start, not the end
Matching tells you that two records probably represent the same customer. Agentic use raises further questions:
- Which record should processes use?
- Which values survive?
- Which source owns each attribute?
- Which relationships matter?
- Which decisions are safe to automate, and which should go to a data steward?
This is the territory of operational MDM. It provides a governed Golden Record with persistent identity, survivorship, source ownership, lineage, relationships and stewardship. That doesn’t mean forcing everything into one simplistic record. Survivorship can reasonably differ by attribute, source, brand or market. The aim isn’t less data; it’s data you can trust.
Where Data 360 fits
The obvious question from any Salesforce architect is whether Data 360 already handles this through identity resolution.
Identity resolution is valuable, but it solves a different problem. Data 360 brings in transactional, behavioural and engagement data and resolves identities into unified profiles. Through the Data 360 MCP Server, it makes that context available to agents. Those unified profiles, however, are built alongside the source records. They don’t consolidate the duplicate Accounts and Contacts in the CRM itself, and those CRM records are what your sellers, Flows, integrations and Salesforce in Claude work with day to day.
The two approaches are therefore complementary:
- Operational MDM establishes which customer record the business trusts within the CRM.
- Data 360 enriches that customer with wider context for insight and activation.
Agents reason best with both.
Govern it where the work happens
If Salesforce is where sellers, service teams, automation and agents take action, the trusted record should be maintained there too. The same governed customer can then serve:
- users;
- Flows;
- Agentforce;
- Claude, in both directions;
- Data 360;
- reporting;
- downstream integrations.
Governing the data in place also turns data quality from a pre-project clean-up into an ongoing discipline. A one-off dedupe before go-live starts to degrade the day the next web form, list import or integration sync runs.
Where to start
A full MDM programme doesn’t need to be in place before joining the beta. A pragmatic sequence:
- Scope to what the sales skills will touch. Start with the Accounts and Contacts linked to open pipeline and to the users in the pilot.
- Agree survivorship and source ownership for the attributes that matter most to those skills: account names, hierarchy, key contacts and contact details.
- Put matching and merging on a continuous footing for that scope, with ambiguous matches routed to a steward rather than auto-merged.
- Widen coverage as the rollout widens.
- Along the way, it is worth checking whether consent and other sensitive attributes are governed appropriately, and whether users, automation and AI will all see the same trusted customer.
How clearMDM helps
clearMDM is a Salesforce-native operational MDM and data quality platform. It supports a continuous lifecycle within the org:
Cleanse > Match > Merge > Govern > Maintain
That lifecycle delivers persistent Golden Records with governed matching, survivorship, lineage, relationship context and stewardship.
clearMDM masters the existing Salesforce records themselves rather than a copy held elsewhere and it records each record’s MDM status as attributes on that record. The governed Account or Contact is therefore the same record that users, Flows, Agentforce and Salesforce in Claude already work with, so there is no separate AI data pipeline to build. Its mastering status sits on the record too, visible to any user, process or agent with the right field access.
clearMDM complements Data 360 rather than replacing it. The clearMDM Data Steward Agent for Agentforce helps stewards review match decisions, understand the AI’s reasoning and automate the lower-risk ones. That is the same human-in-the-loop principle agentic AI needs more broadly.
The question has changed
Permissions tell Claude what it may see. Workflows tell it what it may do. Neither tells it which conflicting customer record the business actually trusts, unless that decision has already been governed.
So the question for Salesforce teams is no longer just “What data can our AI access?”
It’s “What data should our AI trust?“