Agentforce vs. Microsoft Copilot Studio: a stack-honest comparison for enterprise buyers

The evaluation meeting usually goes the same way. Salesforce demos an agent that resolves a service case end to end — reads the customer’s history, checks entitlement, issues the refund, logs everything. Microsoft demos an agent that drafts the account review from last quarter’s emails, the Teams thread, and the proposal deck nobody can ever find. Both demos land. Both vendors claim the enterprise agent crown. And the committee leaves the room no closer to a decision.

Here’s the uncomfortable truth neither sales team volunteers: they’re both right, in different worlds. Agentforce and Copilot Studio are not two implementations of the same product. One is the agent layer of a CRM platform; the other is the agent layer of a productivity suite. Most head-to-head comparisons miss this and end up scoring a fish on tree-climbing.

We work on the Salesforce side of this fence, so you know our vantage point — but this comparison only helps you if it’s honest about where Microsoft wins, and it wins in more places than Salesforce marketing admits. This post compares the two platforms on grounding, agent-building model, pricing at scale, governance, and lock-in, then gives you the tiebreaker question that settles most real evaluations.

Two platforms with different centres of gravity

Agentforce is Salesforce’s agent platform, launched in September 2024 with customer-facing service agents priced at $2 per conversation and expanded since into employee-facing and voice use cases. Agents are built from topics — bundles of natural-language instructions and executable actions drawn from Flow, Apex, prompt templates, and APIs — and orchestrated at runtime by the Atlas Reasoning Engine, which plans which actions to run and in what order. In October 2025 Salesforce rebranded the stack as the Agentforce 360 Platform and shipped Agent Script, a declarative language for hybrid reasoning that lets builders pin down deterministic logic where the LLM shouldn’t improvise. The centre of gravity is the CRM: the agent lives next to the customer record, and its actions read and write that record under the platform’s security model.

Copilot Studio grew out of Power Virtual Agents and sits in the Power Platform. It’s the maker environment for building agents that publish into Teams, Microsoft 365 Copilot, SharePoint, websites, and other channels. One clarification that saves buyers real confusion: Copilot Studio is not Microsoft 365 Copilot. The latter is the $30-per-user-per-month assistant embedded in Word, Excel, Teams, and Outlook; Copilot Studio is the factory where you build your own agents, and Microsoft now bundles access to it with Microsoft 365 Copilot licences for internal scenarios. Its centre of gravity is the Microsoft Graph — the tenant-wide index of your files, mail, chats, and sites.

Here’s the comparison in one view before we take each dimension apart:

Dimension	Agentforce	Microsoft Copilot Studio
Natural home	Salesforce CRM — Service, Sales, Experience Cloud, Slack	Microsoft 365 — Teams, Copilot Chat, SharePoint, plus web and custom channels
Grounding	Data 360 retrievers over unified CRM and federated data (RAG, zero-copy)	Microsoft Graph — SharePoint, Dataverse, connectors, tenant graph semantic search
Agent model	Topics + instructions + actions (Flow, Apex, prompts); Agent Script for deterministic control	Topics + tools + knowledge sources; generative orchestration plans across them
Pricing meters	$2/conversation or Flex Credits (~$0.10/action); per-user add-ons	Copilot Credits ($200 per 25,000/month); zero-rated for M365 Copilot licensed users internally
Governance	Einstein Trust Layer — inline masking, zero retention, toxicity detection, audit trail	Microsoft Purview — audit, DLP, sensitivity labels, eDiscovery, retention
Best first workload	Customer-facing service and sales on CRM data	Employee-facing productivity across M365 content
Watch out for	Data 360 credits billed separately; consumption forecasting	Credit burn once grounding and actions stack up; Purview licensing

Grounding: depth in the CRM vs. breadth across Microsoft 365

Grounding — what data the agent can actually reason over — is where the two platforms diverge most, and it’s the difference that matters longest.

Agentforce grounds agents through Data 360 (formerly Data Cloud) using retrieval-augmented generation over indexed knowledge and harmonised customer data, with retrievers you can tune and filter, and zero-copy federation for data that stays in your warehouse. The depth advantage isn’t really the retrieval, though. It’s that the same agent can act on what it retrieves — update the case, apply the discount, schedule the technician — because its actions are Flows and Apex running inside the org, subject to field-level security and sharing rules that already exist. Two decades of platform automation become the agent’s toolbox on day one.

Copilot Studio grounds agents in knowledge sources spanning SharePoint, Dataverse, public websites, uploaded documents, and Microsoft Copilot connectors, with optional tenant graph grounding that runs semantic search over the entire Microsoft Graph — including third-party data synced in through Graph connectors. Knowledge sources authenticate as the agent user, so the agent only surfaces content that specific person can already access. That breadth is genuinely hard for Salesforce to match: no Agentforce agent natively reads your workforce’s email, chat history, and document estate. If your institutional knowledge lives in SharePoint and Teams — and in most enterprises, an enormous amount of it does — Copilot Studio starts with a corpus Agentforce simply cannot see.

Each side can technically reach the other’s data. Copilot Studio has connectors into Salesforce; Agentforce can federate external sources through zero-copy and MuleSoft. In practice both cross-stack paths add latency, integration effort, and a second security model to reconcile. Native grounding wins, and which platform is “native” depends entirely on where the data your agent needs already sits.

Topics, tools, and orchestration: how you actually build agents

Both platforms organise agents around topics, which trips up buyers into thinking the build experience is similar. It isn’t — and the difference shapes who on your team can do the work.

An Agentforce topic packages a classification description, natural-language instructions, and a set of actions. The Atlas engine classifies the request into a topic, then plans and executes actions until the request is fulfilled. Actions are first-class platform artefacts: an invocable Apex method, a Flow, a prompt template. That means the build skillset is your existing Salesforce team, and the hard work is mostly the same work Salesforce projects always involve — clean automation, sensible permissions, well-described interfaces. Agent Script added a programmable layer for teams who need the agent to behave deterministically in regulated or high-stakes flows: conditionals, variables, and explicit control flow compiled into an agent graph that the reasoning engine executes.

Copilot Studio inherited authored dialog trees from Power Virtual Agents — in classic orchestration, a topic fires when the user’s message matches its trigger phrases, and the conversation follows the nodes you drew. The platform’s current default is generative orchestration, where an LLM planner selects topics, tools, knowledge sources, and even other agents based on their descriptions, chains them for multi-intent requests, and auto-generates questions for missing inputs. Tools are where Microsoft’s ecosystem shows: the Power Platform connector catalogue, agent flows for deterministic sequences, and — from Build 2025 — multi-agent orchestration, bring-your-own-model from the Foundry catalogue, and computer use for legacy UIs. Model flexibility is a real Microsoft advantage as of this writing; Salesforce manages model selection for you.

Notice the convergence, because it’s telling. Salesforce started with an LLM planner and is adding determinism (Agent Script). Microsoft started with deterministic dialog trees and added an LLM planner (generative orchestration). Both vendors learned the same lesson from production: pure prompt-driven agents are hard to trust, and pure decision trees are hard to scale.

Pricing at scale: two meters that don’t translate

This is where naive comparisons do the most damage, because the two pricing models don’t share a unit. You’ll see claims that Copilot Studio is “100x cheaper.” At parity of work, it usually isn’t — but the cost curves do bend differently, and the difference is decision-relevant.

Salesforce runs two consumption meters. Conversations price customer-facing engagement at a flat $2 per conversation, however simple or complex. Flex Credits price the work itself: $500 per 100,000 credits, with a standard action consuming 20 credits — $0.10 at list — and voice actions consuming 30. The pricing page adds per-user options: add-ons at $125 per user per month and Agentforce 1 Editions from $550 per user per month with a bundled credit pool. Two traps to know: the meters can’t be mixed in one org, and Salesforce’s own pricing examples note that Data 360 credits — where grounding and knowledge indexing get billed — are a separate line. In practice, the metered agent usage is often not the biggest number on the invoice.

Microsoft switched its common currency from messages to Copilot Credits on September 1, 2025, sold as prepaid packs of 25,000 credits at $200 per month or Azure pay-as-you-go at the same rate. The published rates: a classic answer is 1 credit, a generative answer 2, an agent action 5, tenant graph grounding 10, and agent flow actions 13 per hundred. A credit works out to $0.008 at pack list — which sounds like rounding error until you stack the meters. Microsoft’s own example prices a single grounded response at 12 credits: roughly $0.096. Read that twice. A Graph-grounded Copilot Studio answer and an Agentforce action both land at about ten cents. The platforms’ unit economics converge at equivalent work; what differs is the shape of the bill.

Two structural differences matter more than the rates. First, employee-facing usage by Microsoft 365 Copilot licensed users is zero-rated for classic answers, generative answers, and tenant graph grounding in Microsoft 365 surfaces — so if you’ve already bought M365 Copilot for your workforce, internal agents ride at close to zero marginal cost, within fair-use limits. That’s a genuinely strong position Salesforce answers only with its own per-user add-ons. Second, expiry and overage behave differently: unused Copilot Credits don’t carry over month to month and technical enforcement can deny service on sustained overage, while unused Flex Credits don’t roll over across terms but overage simply bills at your contracted rate. Model your single biggest workload on both meters before you believe anyone’s per-unit price — ours or theirs.

Governance: an inline trust pipeline vs. a tenant-wide compliance estate

Both vendors have serious governance stories. They’re just aimed at different failure modes, and it’s worth being precise about which one keeps your risk team up at night.

The Einstein Trust Layer sits inline on every request between Salesforce and the model providers. Before a prompt leaves the platform it passes through secure data retrieval, dynamic grounding, PII masking, and zero-data-retention agreements with LLM partners, with toxicity detection on the way back and an audit trail — including the masked prompts and safety scores — stored in Data 360. The design philosophy is a guarded request path: the sensitive thing never reaches the model, and everything that happens is logged where you can query it. It’s default-on infrastructure, not a policy you have to remember to configure. Its scope, fairly stated, is what flows through Salesforce’s AI stack — it doesn’t govern the rest of your information estate.

Microsoft’s answer is not one component but the Purview suite applied to agents. Purview supports auditing, data classification, sensitivity labels, DLP, insider risk management, communication compliance, eDiscovery, and retention for Copilot Studio agent interactions. Prompts and responses are captured in the unified audit log and stored in the user’s mailbox, which means legal can put them on hold and search them like any other communication. Sensitivity labels carry through: an encrypted document is only used if the agent user holds the extract right, and responses display the highest-priority label of their sources. The honest caveats: much of this assumes corresponding Purview licensing, and governing agents published to non-Microsoft channels requires enabling Purview’s pay-as-you-go billing.

If your board’s question is “prove no customer PII reaches a foundation model,” the Trust Layer is the more direct answer. If it’s “produce every AI interaction mentioning this deal for the regulator,” Purview’s eDiscovery and retention machinery is ahead. Neither vendor is weak here. They’re strong at different questions.

Lock-in cuts both ways

Nobody’s agent is portable, and you should price that in on both sides. Topics, instructions, actions, tests, and analytics are platform artefacts; moving from one platform to the other is a re-implementation, not a migration. The lock-in compounds through three quieter channels.

• Data gravity. Agentforce gets better the more you consolidate into Data 360 — which is metered separately and deepens the Salesforce commitment. Copilot Studio gets better the more your content and third-party data flow into the Graph and Dataverse. Grounding quality is the moat on both sides.
• Skills gravity. Agentforce work lands on Salesforce admins and developers; Copilot Studio work lands on Power Platform makers and M365 admins. Whichever bench you staff becomes an argument for the next agent living on the same platform.
• Commercial gravity. Credit pools, pre-purchase discounts, and bundle pricing on both sides reward concentration. Salesforce’s Flex agreements let you convert between user licences and credits; Microsoft’s commit units spend across eligible products. Both are designed to make the marginal agent cheapest on the platform you already pay.

The mitigations are the same regardless of vendor: keep business logic behind APIs the agent calls rather than inside the agent, keep instructions and prompts in version control, and design channels so the front door isn’t welded to the brain. You won’t eliminate lock-in. You can keep it to the orchestration layer instead of your entire process logic.

The honest tiebreaker: where your record of truth and your workforce live

Strip away the demos and most evaluations settle on one question: where do your system of record and your workforce already live? If Salesforce holds the customer master and runs your revenue and service processes, customer-facing agents belong on Agentforce — the grounding is native, the actions already exist as Flows and Apex, and the security model is the one your auditors have already signed off. If your workforce lives in Teams and Outlook and you’ve bought Microsoft 365 Copilot, employee-facing agents belong in Copilot Studio — the corpus is already indexed, the surface is already open in front of every employee, and the marginal cost is close to zero.

That’s why the largest enterprises we see aren’t choosing. They’re drawing a boundary: Agentforce for the customer-facing processes anchored in CRM, Copilot Studio for internal productivity anchored in M365, with integration handling the handoffs. The failure mode isn’t picking the “wrong” platform — it’s forcing one platform to fight on the other’s home turf, then blaming the technology for the integration bill. If the Salesforce side of that boundary is yours to build, that’s what our Agentforce practice does, and our readiness assessment will tell you in a few minutes whether your org’s data and automation are actually prepared for it. Walk into the next vendor meeting with the boundary already drawn, and both demos suddenly get much easier to judge.

Understanding the basics

What is the difference between Agentforce and Microsoft Copilot Studio?

Agentforce is Salesforce’s platform for building AI agents grounded in CRM data, with actions built from Flow and Apex and governed by the Einstein Trust Layer. Microsoft Copilot Studio is Microsoft’s agent-building environment grounded in the Microsoft Graph — SharePoint, Teams, Outlook, and Dataverse — publishing into Microsoft 365 surfaces and governed through Microsoft Purview. Agentforce is strongest for customer-facing work on CRM records; Copilot Studio is strongest for employee-facing work across Microsoft 365 content.

Is Agentforce or Copilot Studio cheaper at scale?

Neither, universally — the units differ. Agentforce lists at $2 per conversation or roughly $0.10 per action on Flex Credits. Copilot Studio credits cost about $0.008 each, but a Graph-grounded answer consumes around 12 credits (~$0.10), so equivalent work often costs similar money. Copilot Studio’s decisive edge is that Microsoft 365 Copilot licensed employees use internal agents at no extra metered cost; Agentforce counters with flat per-user add-ons. Model your actual workload on both meters.

Can Agentforce and Copilot Studio be used together?

Yes, and large enterprises with both stacks commonly do. The workable pattern is a boundary by workload: Agentforce runs customer-facing agents on the CRM system of record, Copilot Studio runs employee-facing agents on Microsoft 365 content, and integration — APIs, middleware, or connectors — handles handoffs between them. What doesn’t work well is forcing either platform to ground primarily in the other’s data, which adds latency, cost, and a second security model to every request.

---

Weighing Agentforce against Copilot Studio for a real workload? Talk to us — stack-honest platform advice is what we do.