There is a pattern emerging across Agentforce deployments that deserves more attention than it is getting.
Organisations invest in the platform. They design the agents, map the topics, wire up the actions and configure the Trust Layer. Then, somewhere in the later stages of the programme, a question surfaces that does not have an obvious answer: what happens when the default large language model is not the one the business is allowed to use?
This is not an edge case. For any organisation operating under serious governance constraints — regulated data obligations, enterprise supplier frameworks, or public sector compliance requirements — the model question is not a programme detail. It is a programme-level decision. And too many organisations are hitting that wall only after significant time and budget have already been committed.
Tightening regulation, increasingly specific procurement frameworks and the competitive pressure to get agentic AI into production are all accelerating the moment when that wall gets hit. The EU AI Act, the FCA's AI guidance and sector-specific requirements in automotive, financial services and healthcare are all moving in the same direction: greater accountability for which models are used, on what data, with what audit trail. Enterprise supplier frameworks are specifying AI model providers at a contractual level. And the cost of a deployment that stalls in procurement — or has to be rebuilt because the model governance was not right — is not just a technology problem. It is a business problem.
The framing that matters here is this: Agentforce is not a product you implement. It is an enterprise AI platform you configure. That distinction has significant implications for how programmes should be designed and how partners should be evaluated. When the right expertise is in the room, the constraints that look fixed are often not. Custom LLM support, alternative model providers, governance-aligned data architecture — these are legitimate enterprise configurations, not workarounds. The question is whether your delivery partner has the access, the relationships and the technical capability to unlock them.
A deployment that could have stalled
One of North America's largest automotive manufacturers needed Agentforce running in production across its Customer Care function. The programme was substantial: a Master Agent orchestrating seven specialist agents covering customer, vehicle, service, goodwill, recall, case and dealer interactions. Over 660,000 annual advocate research interactions were in scope, with a 60% reduction in case-resolution cycle time as the target.
The data foundation came first. Three siloed environments — Salesforce CRM, an internal Customer Insights Hub and a separate VIN and Customer 360 portal — were unified through Salesforce Data Cloud using zero-copy ETL from Databricks. The Einstein Trust Layer was configured with data masking, zero retention and a full audit posture, because recall, warranty and goodwill workflows carried regulated obligations that could not be compromised.
Then the model question surfaced. The manufacturer's enterprise governance and supplier relationships required a specific choice of model provider. The default Agentforce planner LLM was not sanctioned for use. The alternative — Anthropic Claude — was on the Salesforce roadmap as a pilot feature, not yet generally available. Without a path forward, the deployment could not proceed.
The solution was approached on three fronts simultaneously. The pilot pathway inside Salesforce was identified and the client's engineering team were brought into it. The client engaged the Salesforce account team commercially, the pilot contract was signed, and the feature was enabled in the relevant sandbox. A structured evaluation then ran alongside Salesforce's product team in weekly working sessions — covering every topic and action in the advocate agent, batch regression testing at scale and root-cause investigation on topic failures, with findings fed directly back into Salesforce's product development process.
The outcome was a production deployment running on Anthropic Claude — with the ability to switch between models and Claude set as the default — two to three months before the feature reached general availability. The evaluation fed directly into the hardening of that feature for the GA release. This is what it looks like when a governance blocker becomes a programme unlock, and when that unlock advances the platform for every organisation that comes after.
What this means in practice
The organisations getting the most from Agentforce are not the ones that moved fastest. They are the ones that treated governance as a first-class programme decision rather than an implementation detail.
That means having the model conversation at the start — before a single agent is designed, the question of which LLM will run the planner and whether it is sanctioned under your enterprise supplier framework should be answered. It means ensuring the data foundation is resolved before the first agent is built; the Trust Layer configuration, data masking posture and audit obligations shape everything that runs on top of them. And it means choosing a delivery partner whose relationship with Salesforce's product team runs deeper than a partner badge — one that can tell you what is on the roadmap and help you get into it when what you need is not yet generally available.
The gap between what is in general availability and what is possible with the right access and relationships is significant. Where in your Salesforce estate is the platform default holding your programme back? Where is the governance constraint that is quietly slowing your rollout? Those are the questions worth asking — and the sooner they are asked, the better the answer tends to be.
