Two platforms own the enterprise agent conversation in May 2026. Salesforce Agentforce and Microsoft Copilot Studio show up in every architecture review, every procurement deck, every CIO briefing. Both have eight-figure marketing budgets behind them. Both have salespeople telling your CFO that the platform "is the obvious choice for an enterprise like yours." Both are, in narrow circumstances, correct. Both are, in most circumstances, wrong for at least half the workloads they pitched against. The architect's job is to cut through the marketing and answer one question per workload: does this platform belong here or doesn't it? This lesson gives you the six deal-breakers (three per platform), the data-gravity rule that beats every other heuristic, and the side-by-side architecture read that survives a hostile vendor counter-pitch.

What Each Platform Actually Is

Cut the marketing first. Both vendors describe their product in language designed to feel familiar to the buyer of their flagship product. Salesforce describes Agentforce as the "agent layer of the Customer 360." Microsoft describes Copilot Studio as the place to "build, customize, and deploy your Copilots." Both descriptions are accurate. Neither is decision-relevant.

Agentforce is the agent runtime for Data Cloud-resident data

Agentforce, in the architect's read, is a vendor-native agent runtime designed to operate on data already in Salesforce's Data Cloud — the unified customer data store. The atomic units are Topics (the agent's job-to-be-done), Actions (the tools the agent can call, which are typically Flows, Apex, or Agentforce-wrapped APIs), and Instructions (the prompt-like guidance per topic). The agent runs on Salesforce-hosted infrastructure with Salesforce's prompt routing, Salesforce's eval primitives (Agentforce Testing Center), and Salesforce's pricing model. The supported model layer (Atlas reasoning engine plus OpenAI / Anthropic foundation models behind Einstein Trust Layer) is opinionated. The integration story is opinionated. The data story is the deepest moat — every workflow already living in Salesforce flows into the agent with minimal friction.

Copilot Studio is the agent runtime for M365 Graph-resident data

Copilot Studio is Microsoft's agent runtime designed to operate on data already in the Microsoft 365 Graph — Outlook, Teams, SharePoint, OneDrive, Dynamics 365. The atomic units are Topics (intent-routed conversation branches), Actions (Power Automate flows, Power Fx expressions, or custom connectors), and Generative Answers (RAG over Dataverse, SharePoint, websites). The agent runs on Azure infrastructure with Microsoft's prompt orchestration (the Copilot orchestrator behind the scenes), Microsoft's eval primitives (still maturing as of May 2026), and Microsoft's consumption-based pricing. The model layer is GPT-family (via Azure OpenAI) plus Phi for some workloads. The integration story is opinionated. The data story is the deepest moat — every workflow already living in M365 flows in.

The pattern is symmetric. Both vendors built an agent runtime on top of their existing data gravity. The choice is not really "which platform is better" — the choice is "which data gravity is the agent's primary home, and does that gravity lock the right things in or the wrong things in?"

The Data Gravity Rule

The single rule that beats every other heuristic: pick the platform that lives where the agent's primary data already is. If 70%+ of the data the agent needs is in Salesforce, Agentforce is the natural choice for that workload. If 70%+ is in M365, Copilot Studio is the natural choice. If the data is split 50/50, neither vendor is the right answer for that workload (and the third option becomes "build on a horizontal platform that talks to both").

Why data gravity wins

Every other architectural concern (latency, observability, eval, security) compounds in favor of the platform that doesn't have to cross a data boundary to do its work. If the agent has to call out to a different system every time it needs customer history, you pay in latency (50-200ms per cross-system call), tool-call reliability (each integration is a failure surface), and engineering time (one more connector to maintain). Native platforms amortize these costs to zero for their home data.

How to measure data gravity

For each agent workload, list the data sources the agent will read and write. Score each by frequency of access. The platform whose owned systems contain 70%+ of the weighted access is the data-gravity natural choice. Below 70%, the gravity argument weakens; below 50%, it disappears.

A worked example. An expense-report assistant agent. Data sources by weighted access: SAP Concur (expense submissions, 40%), Outlook (receipts in email, 25%), SharePoint (policy docs, 15%), Dynamics 365 (cost centers, 10%), Slack (notifications, 10%). M365 share: 25% + 15% + 10% = 50%. Salesforce share: 0%. SAP share: 40%. No single platform owns 70%. Native vendor lock-in is not justified. Build on a horizontal platform that can talk to all three.

Another. A customer-account-summary agent for an account manager. Data sources: Salesforce account history (60%), Salesforce opportunity data (20%), Salesforce service cases (10%), Outlook recent emails (10%). Salesforce share: 90%. Agentforce is the obvious choice for this workload. Building on a horizontal stack just to avoid Salesforce lock-in costs you engineering for no architectural benefit.

Agentforce: Three Deal-Breakers

Three architectural facts about Agentforce that decide whether it fits a workload. If any one of them is a hard no, Agentforce is not the right choice no matter how good the demo was.

Deal-breaker 1: Data Cloud gravity

Agentforce is built on Data Cloud. Workloads where the data is not in Data Cloud (or cannot be cleanly ingested into Data Cloud) lose most of Agentforce's value. The platform's "this just works" promise depends on the data already being unified in Data Cloud's schema. If your customer data lives in Snowflake, BigQuery, or a custom data lake — and you have not Zero Copied or replicated it into Data Cloud — you are paying for an agent that has to traverse boundaries Salesforce specifically designed itself to avoid.

What this means in practice. An architect at a mid-market company with Salesforce CRM but no Data Cloud subscription cannot deploy Agentforce meaningfully — the agent will have access to standard Salesforce objects but not the unified customer profile. The Data Cloud SKU is the prerequisite, not the optional add-on. Budget for it accordingly: Data Cloud enterprise pricing in 2026 starts at $250K/year for credit packs and climbs steeply with row count and consumption.

Deal-breaker 2: The $550/seat floor

Agentforce service edition lists at $2 per conversation for the unlimited model, but enterprise contracts usually settle into a per-seat or hybrid model. The architect's practical number is the $550/user/month enterprise seat that anchors most multi-year deals (Service Cloud + Agentforce bundle). That is the floor for the human user who interacts with the agent via the Salesforce UI. For agent-to-customer interaction (Sierra-style deflection), the conversation-based pricing applies. Both can stack.

What this means in practice. A 200-seat service org running Agentforce in seat-plus-conversation model lands in the $1.4M-$2.2M annual range before you count Data Cloud. If you do not have a workload that returns multiples of that in value (deflection of human ticket cost, productivity uplift, revenue lift on agent-driven leads), Agentforce will be cut at the next budget review. The pricing demands a workload where the math is unambiguous.

Deal-breaker 3: The Salesforce-stack assumption

Agentforce assumes Salesforce is the system of record. Agent actions are most naturally Flows, Apex methods, or Agentforce-native APIs. Integrations to non-Salesforce systems are possible via MuleSoft, external services, or HTTP callouts — but each is a step away from the platform's happy path. If your operational stack is heterogeneous (Salesforce CRM but Dynamics ERP, or Salesforce CRM but Workday HR), every cross-platform action becomes an integration project. The cost-of-integration shows up in tool latency, runbook complexity, and the eval set you have to maintain to ensure cross-system calls work.

What this means in practice. Agentforce is a great choice when Salesforce is 80%+ of the operational stack. It is a mediocre choice when Salesforce is 30-50%. It is a poor choice when Salesforce is the CRM but every other operational system lives elsewhere.

Copilot Studio: Three Deal-Breakers

Deal-breaker 1: The 25K-credit pack at $200

Copilot Studio in 2026 is metered on Copilot Credits. The pricing anchor that matters for architects: 25,000 credits at $200 per month, billed per-tenant. Generative responses, prompt actions, and autonomous agent runs consume credits at different rates (generative answers around 1 credit, full autonomous runs at 10-50 credits depending on complexity). Pay-as-you-go is available at higher per-credit cost.

What this means in practice. A back-of-envelope conversion: 25K credits supports roughly 5,000-10,000 generative-answer chat interactions, or 500-2,500 autonomous agent runs, depending on the mix. A department-scale workload (10,000+ daily autonomous actions across an HR or IT org) consumes credits faster than the pricing tier suggests; budget for 5-20 credit packs ($1K-$4K/month) for moderate adoption and 50+ for true department-wide use ($10K+/month). The per-credit rate at scale is competitive but only when measured against the right baseline. Compare against $0.04-$0.12 of LLM token cost per equivalent action on a horizontal stack — Copilot Studio is paying for orchestration, M365 integration, and the Copilot orchestrator. You are not getting the raw model cost.

Deal-breaker 2: M365 graph dependency

Copilot Studio's value proposition depends on Microsoft Graph being the source of truth for the data the agent needs. Outlook, Teams, SharePoint, OneDrive, Dynamics 365 — the M365 graph is the agent's home. Workloads where the primary data lives outside the M365 graph (Salesforce, Workday, SAP, custom data lake) lose most of Copilot Studio's value.

You can connect external systems via custom connectors (Power Platform supports 1,000+) or Power Automate flows. But the moment you do, you are paying the cross-boundary tax that Copilot Studio was specifically designed to avoid. The connector layer is mature but introduces latency, complexity, and a separate eval and observability surface (Application Insights for the agent, Power Platform Admin Center for the connectors, Microsoft Purview for the data). The platform is at its best when the data is already in the graph.

What this means in practice. Copilot Studio is the obvious choice for an M365-native organization (most large enterprises, especially in regulated industries that standardized on Microsoft a decade ago). It is a mediocre choice for organizations where Salesforce or Google Workspace owns the day-to-day collaboration layer and M365 is just the email backplane.

Deal-breaker 3: The weak voice story

Copilot Studio's voice agent capabilities, as of May 2026, are behind the leaders. Microsoft has voice via the Speech Services and the Voice agents in Dynamics 365 Contact Center, but the unified Copilot Studio voice experience for general-purpose agent deployment lags Sierra, Decagon, and even some of the horizontal-platform alternatives (LiveKit + Vapi, Pipecat) on latency, interruption handling, and naturalness. For text-only or in-app chat workloads, voice is not a factor. For workloads that require high-quality voice (customer-facing phone agents, IVR replacement, voice-driven internal assistants), Copilot Studio is not the right choice in May 2026.

What this means in practice. If voice is a non-functional requirement, exclude Copilot Studio from the shortlist for that workload. Microsoft is closing the gap — the gap is real today.

Side-by-Side: The Architecture Read

The architect's side-by-side. Same dimensions, same lens, no marketing language.

Data plane

Agentforce: Data Cloud is the unified data plane. Zero Copy from Snowflake, Databricks, BigQuery is available but adds latency and limits some real-time scenarios. Best when data is natively in Salesforce objects.

Copilot Studio: Microsoft Graph + Dataverse is the unified data plane. Connectors for external systems are mature (Power Platform's 1,000+ connector library) but each connector is an integration to maintain. Best when data is natively in M365 / Dynamics.

Tooling and actions

Agentforce: Actions are Flows, Apex methods, Agentforce-native APIs, MuleSoft Composer, or HTTP callouts. Action authoring is in the Flow Builder or Apex IDE. Tool reuse across topics is high; tool reuse across non-Salesforce systems requires MuleSoft or custom integrations.

Copilot Studio: Actions are Power Automate flows, Power Fx expressions, custom connectors, or AI Builder prompt actions. Authoring is in the Copilot Studio designer or Power Platform tooling. Tool reuse across topics is high; the connector library covers most third-party SaaS at the cost of consumption credit overhead.

Model layer

Agentforce: Atlas reasoning engine plus pluggable foundation models (OpenAI via Einstein Trust Layer, Anthropic, on-platform fine-tunes). Model selection per topic is supported. Cost is bundled into pricing.

Copilot Studio: GPT-4o / GPT-5 family via Azure OpenAI, Phi for some workloads, increasing model choice in 2026. Model selection per agent is supported. Cost is metered via Copilot Credits.

Eval and observability

Agentforce: Agentforce Testing Center for regression suites. Agent Monitor for production traces. Reasonable out-of-the-box; missing some of the depth a Braintrust or LangSmith provides for non-platform users.

Copilot Studio: Application Insights integration, Power Platform Admin Center, Microsoft Purview for data governance. Eval primitives are improving but the explicit "eval suite" pattern is less mature than Agentforce's Testing Center.

Pricing model

Agentforce: Per-seat ($550/user/month enterprise floor common), per-conversation ($2 / conversation list, lower at scale), Data Cloud consumption credits. Multi-year discounts standard.

Copilot Studio: Copilot Credits ($200 / 25K credits/month per tenant) plus M365 Copilot per-seat ($30/user/month) for many embedded scenarios. Pay-as-you-go available at premium per-credit rate.

Exit cost

Agentforce: high. Topics, Actions, Instructions, and trained models are proprietary. Migration off requires prompt rewrites, tool re-wrapping, and eval set rebinding (see lesson 4). 6-12 months for a department-scale program.

Copilot Studio: high. Topics, flows, connectors, and Power Platform assets are proprietary. Migration off similarly requires rewrites. 6-12 months for a department-scale program.

Three Workloads, Three Different Answers

The framework is per-workload. Three illustrative scenarios.

Workload 1: Service deflection for a B2C company with Salesforce Service Cloud as the system of record

Data gravity: Salesforce Service Cloud (cases, customer profile, history) ~85%. Outlook ~10% for case-related email. Other ~5%.

Answer: Agentforce is the natural choice. The data is in Salesforce. The seat licensing covers the human supervisors. The deflection economics work because the agent is acting on data it already owns. Copilot Studio for this workload would require building connectors to pull case data out of Salesforce on every call — a non-starter at scale.

Workload 2: Internal IT helpdesk for an M365 shop

Data gravity: SharePoint (KB articles) ~30%, Intune / Endpoint Manager ~25%, Active Directory / Entra ID ~20%, ServiceNow (ticket system) ~15%, Teams (user notification) ~10%.

Microsoft-native share: 30% + 25% + 20% + 10% = 85%. ServiceNow is the one significant outsider but Power Automate has mature connectors. Answer: Copilot Studio is the natural choice. The agent is acting on the Microsoft graph 85% of the time. The Copilot Credit math is reasonable for the ticket volume (a 500-employee org typically runs 3K-6K IT tickets/month, well under the credit-pack ceiling for moderate use).

Workload 3: GTM outbound and lead enrichment for a B2B SaaS

Data gravity: Salesforce CRM ~30%, Apollo / ZoomInfo / external enrichment ~25%, LinkedIn ~15%, Marketo or HubSpot ~15%, Outlook (email outreach) ~15%.

No platform owns 70%. Salesforce share 30%. Microsoft share 15%. The data is genuinely fragmented across data brokers, social platforms, and CRMs. Answer: neither platform is the natural choice. Buy Clay (the vertical agent for this workload — see lesson 1). The horizontal-build alternative on a platform like Lindy or n8n + Claude is the second choice. Agentforce or Copilot Studio for this workload would mean fighting the platform on every action.

The Marketing Counter-Pitch to Be Ready For

Both vendors have practiced counter-pitches. Be ready.

"But Agentforce is the only place to act on the Customer 360"

Translation: data gravity argument restated. If your customer data is in Salesforce, yes. If it isn't, the Customer 360 is an aspiration and a multi-month Data Cloud ingestion project, not an existing reality. The right pushback is: "show me the actual data flows for the workload I am evaluating; if they're 30% Salesforce I am not paying the Agentforce premium to operate on that 30%."

"Copilot Studio is included in your M365 enterprise agreement"

Translation: the entry-level Copilot Studio is bundled but the agent capabilities you need (autonomous agents, generative answers at scale, custom connectors with external data) burn through credits fast and require additional packs. The right pushback is: "model the credit consumption for the workload at full adoption and show me the true monthly cost — don't assume the entry tier covers it."

"Our roadmap addresses your concern"

Translation: the feature you need does not exist today. The roadmap may be real, may not. The right pushback is: "I am buying production capability today, not a roadmap. If the feature ships, we'll evaluate the platform again. Until then, the workload goes to the alternative that solves it today."

"You'll save money long-term"

Translation: vendor TCO modeling that assumes you stay on the platform for five years, adoption scales as projected, and renewals do not include price increases. Real TCO modeling (lesson 3) usually shows the vendor model is optimistic by 30-50% on the cost side and pessimistic on the buyer's exit cost by an order of magnitude.

The Decision Template for Each Workload

Use this for each workload in the program plan.

1. Map the data sources. Where does the data live? Weight by frequency of access.
2. Calculate data-gravity share. Microsoft graph share, Salesforce share, other share. If one is 70%+, that platform is the natural choice. If none is, native platforms lose their main advantage.
3. Check the three deal-breakers for each candidate. Data Cloud subscription for Agentforce, credit-pack math for Copilot Studio, voice requirement, stack heterogeneity.
4. Estimate the all-in annual cost. Seats + consumption + data + integration. Compare to the alternative.
5. Confirm exit story. If you have to migrate in 24 months, what does that cost?
6. Decide and document. One workload, one platform, with the data-gravity ratio and cost math as the recorded justification.

The Five Failure Modes to Avoid

The "we standardized on X" trap

"We standardized on Microsoft" or "we standardized on Salesforce" is not a technology argument; it is a procurement argument. The agent decision is workload-by-workload. If the standardization gives you a discount, great. It does not override the data-gravity math.

The "the future is one agent platform" trap

No, the future is not one agent platform. The 2026 reality is two large vendor-native platforms, a half-dozen credible vertical agents, and many workloads that don't fit any of them. Plan for compose (lesson 1), not consolidation.

The "we'll wait for the roadmap" trap

Workloads ship now. Roadmaps slip. Choose for today; evaluate again in 18 months.

The "the demo solved it" trap

Vendor demos use vendor-controlled data and vendor-cherry-picked scenarios. Your customer base will be 15-30% different on the dimensions that matter. Pilot with real production data on a meaningful slice before the multi-year commit.

The "we got a great discount" trap

A great discount on the wrong platform is still the wrong platform. The integration cost, latency tax, and engineering overhead of fighting data gravity will eat the discount in 18 months.

Key Takeaways

• Agentforce is the agent runtime for Data Cloud-resident data. Copilot Studio is the agent runtime for M365 Graph-resident data. The pattern is symmetric.
• The single rule that beats every other heuristic: pick the platform that lives where the agent's primary data already is. 70% data-gravity share is the threshold for native-platform justification.
• Agentforce's three deal-breakers: Data Cloud gravity (no Data Cloud, no Agentforce), $550/user/month enterprise seat floor, the Salesforce-stack assumption.
• Copilot Studio's three deal-breakers: 25K-credit pack at $200/month math (real cost at scale is much higher), M365 graph dependency, the weak voice story as of May 2026.
• Eval and observability: Agentforce Testing Center is more mature; Copilot Studio is closing the gap via Application Insights + Power Platform Admin Center.
• Exit cost is high on both. 6-12 months for a department-scale migration on either platform.
• The decision is per workload. Service deflection on Salesforce Service Cloud = Agentforce. Internal IT on M365 graph = Copilot Studio. GTM with fragmented data = neither (buy Clay, build on horizontal).
• The marketing counter-pitches to be ready for: "Customer 360," "included in M365," "the roadmap," "long-term savings." Pushback is always the data-gravity math and the workload-specific TCO.
• Failure modes: "we standardized," "future is one platform," "wait for roadmap," "the demo solved it," "great discount." None of these override the workload analysis.