Agentforce charges $2 per conversation. A 30-seat support team handling 50,000 interactions a month hits a $100,000 monthly line item - on top of existing Service Cloud seat fees. That bill was not in last year's budget, and most procurement teams did not see it coming. That number is not a reason to build instead of buy. It is a reason to understand what you are actually deciding when you pick a platform.
The debate running through every engineering leadership offsite right now is framed wrong. "Build or buy?" treats the AI agent stack as a single thing, and it is not. It has at least three layers - the foundation model, the orchestration logic, and the workflow integrations - and the right answer can be different at each one. Teams that treat it as a binary are either paying a vendor to own something they should control, or they are building something a vendor already does better.
Why the "just buy it" advice keeps failing in production
The headline case for buying is strong. A prototype comes together in weeks. But moving from prototype to production - handling edge cases, building governance, ensuring reliability across thousands of interactions - stretches to 6-12 months. That gap is where most in-house builds bleed budget. Building entirely in-house requires 5-10 AI engineers, ML infrastructure specialists, and compliance architects - and McKinsey's State of AI 2025 found that 68% of enterprise leaders cite talent scarcity as their primary barrier.
Forrester predicts that 75% of companies attempting to build their own agentic systems will fail, citing the complexity of requiring "diverse and multiple models, sophisticated retrieval-augmented generation stacks, advanced data architectures, and niche expertise."
That is a real data point, and it deserves to be taken seriously. Vendors like Agentforce and Copilot Studio exist because the undifferentiated infrastructure - auth, audit logs, connector libraries, human-in-the-loop plumbing - is genuinely hard to build and maintain. Nobody's CRM integration logic is a competitive moat.
But "buy" usually gets interpreted as "buy the whole stack," and that is where teams hand over something they will eventually need back.
What you are actually buying when you pick a platform
Neither Agentforce nor Copilot Studio lets you self-host the agent runtime in 2026. That is the thing worth sitting with. When you buy a platform, you are not just buying convenience - you are accepting that the vendor controls the model, the reasoning engine, the pricing model, and the upgrade path. Each platform's data layer is essentially mandatory: skipping it produces agents that hallucinate or fail to act on the right record.
Agentforce assumes your customer graph lives in Salesforce. Copilot Studio assumes you have already bought the productivity graph - Microsoft 365 and Dataverse. If your workflow crosses both, you are paying connector overhead on at least one side. Every connector hop is an additional governance surface, an additional latency cost, and an additional source of staleness.
The deeper issue is inference economics. A single chatbot API call might cost $0.001. A multi-step agent that plans, retrieves context, invokes tools, reflects on output, and self-corrects can cost $0.10 to $1.00 per task completion - a 100x to 1,000x multiplier. When a vendor bundles that cost into a "per conversation" price, they are also taking the margin on every model call your agent makes. Vendors often lure customers with generous pilot credits, yet scaling to production routinely reveals 500-1,000% cost underestimation.
Beagle in action#cs-ops, Thursday afternoonThe steelman for buying everything, honestly stated
A fair reading of the data is that most teams should not be running their own orchestration layer at all. The KPMG AI Pulse reports that 42% of companies have successfully integrated agents into their workflows, up from just 11% in early 2025. The teams closing that gap fastest are mostly buying, not building.
The choice itself is less predictive of success than the ability to embed agents inside governed, end-to-end business processes. A team with no agent governance running on LangGraph is in worse shape than one with clear process ownership running on Copilot Studio. The platform is not the variable that matters most.
And the market is moving. Sierra AI, which charges only when an agent successfully resolves an issue without human intervention, hit $100M ARR in 21 months and crossed $150M+ ARR by early 2026. Outcome-based pricing from a vendor that eats the inference cost risk is a genuinely different proposition from a platform that charges per action regardless of result.
Where the line should actually be drawn
Enterprises can buy commodity agents when the process is standardized, the data already lives inside the vendor system, and speed matters more than workflow ownership. Differentiating agents need more control when the workflow crosses systems, uses proprietary data, or encodes domain behavior that competitors should not be able to rent from the same vendor.
That distinction is the real decision. Not build versus buy - buy the layer you do not want to maintain, own the layer that encodes how your business actually works.
In practice, this usually means: buy foundation model access and a connector library, buy the hosting and compliance scaffolding, but own the orchestration logic that decides what your agent does when a workflow fails, when it escalates, and what it logs. Most enterprises do both - buy foundational AI infrastructure, then build proprietary orchestration layers on top.
Engineering leaders who succeed in 2026 and 2027 will be the ones who built a repeatable decision process for where new capabilities belong in the stack, and a platform function that owns the seam - the integration layer between what you built and what you bought - with the same rigor they apply to production infrastructure.
The question your team should answer before the next vendor call is not "build or buy?" It is: which layer, specifically, are we buying? The answer to that question will determine whether your Year-2 renewal is a renegotiation or a trap.
Pricing figures cited from Salesforce Agentforce pricing, Kovil's Agentforce breakdown, and MindStudio's agentic pricing analysis. Production adoption figures from Agentic AI Knowledge Base.