AI ERP with agents inside the systems you run
AI phone and voice agents for real calls
LLM.coPrivate, self-hosted LLM deployments
Legal AI infrastructure for firms
AI RFP discovery and response drafting
Automatic.coBusiness process automation
CyberAttack.aiThreat detection and defense
Secure AI virtual data rooms
DigitalSign.coAI-designed digital signage for any screen
hardware.coCustom AI hardware, software, and engineers
Data.coPrivate AI infrastructure and governance
Build vs Buy AI Agents and the Financial Break Even
Per-seat pricing on an AI agent platform is rarely the largest line in a three-year total cost of ownership; it is the sticker. Variable usage, integration work, and exit cost are the bill.
Four line items drive the three-year number for an agent deployment: per seat, per token or per credit, per integration, and the cost of leaving. Each has a published rate on Copilot Studio, Salesforce Agentforce, or Glean that a build can be measured against at your own engineering rate.
What a Platform Really Charges For
Platforms publish a seat price, but that price rarely covers an agent that does work. Copilot Studio shifted in September 2025 to a credit model where a generative answer costs 2 credits, an agent action costs 5, and tenant graph grounding costs 10, drawn against a $200 pack of 25,000 credits or pay-as-you-go at $0.01 per credit. Agentforce offers three paths: $2 per conversation, $0.10 per action against Flex Credits, or a $50 Coworker seat plus a $125 add-on for sales or service. Glean does not publish rates, but reported enterprise contracts cluster around $75 per seat with roughly a 100-seat minimum. The seat is the floor. Volume drives the real bill, and the vendor controls which actions count toward it. A classification change that re-prices "agent actions" as "premium reasoning" can move the effective rate 2x to 10x against the same workload, with no new contract required.
Line Items That Only Appear on a Build
A build carries line items a platform folds out of sight. There is the model bill from OpenAI, Anthropic, or self-hosted inference. There is orchestration, meaning the workflow engine, retries, and tool schemas. There is observability with per-call cost attribution, identity and secrets for the agent's machine actor, and the engineering salary that keeps the whole thing from drifting. McKinsey's 2026 survey of 1,719 respondents found that 32% of organizations opted to build functionality internally with agentic coding tools rather than buy an off-the-shelf SaaS replacement. The 32% reflects a buy-side decision reversal: when a tokenized, self-run version of a specific workflow costs less than the seat price of a generic one, the economics flip. The project-scope breakdown for a custom AI agent cost tracks the same ratio.
A Three Year TCO Model With Numbers
A working model needs four or five line items on each side and a 36-month horizon. For a 100-seat customer-facing agent, the buy side on Agentforce Coworker is roughly $60,000 per year in seats, plus Flex Credits burn: at 300 agent actions per seat per month, that is $360,000 at $0.10 per action. Integrations through MuleSoft or API integration work add $30,000 to $120,000 depending on the number of systems touched, and premium support runs around 20% of license. The build side has the engineering ticket up front — a four-engineer pod for three to four months at US senior-engineer rates lands between $350,000 and $500,000 — then inference at published model rates of roughly $0.50 to $3 per 1,000 interactions, plus $80,000 to $150,000 a year for hosting, observability, and maintenance. Break even on seats alone falls between month 14 and month 22 for most profiles. On token spend it arrives earlier, because platforms mark up the underlying model call. The private versus proprietary AI cost breakdown walks the token math per workload.
The Exit Cost Nobody Models First
Exit cost is a line the buyer tends to see in month 25, and it has four components. The custom logic sits in the platform's authoring tool. The integration adapters are written against its APIs. The conversation history and audit data live in its store. The retrieval index or fine-tuned model is bound to its backbone. On Copilot Studio, topics and knowledge sources live in Dataverse; on Agentforce, flows and actions live in Salesforce metadata; on Glean, the index lives in a proprietary graph against a connector catalog. None of this ports cleanly. A fair planning estimate is 30% to 60% of the original integration build cost to migrate, plus a year of parallel-run licensing on both ends. That figure belongs in year-one TCO, not year three, because the decision to avoid it compounds. Builds have their own version of this, concentrated in the choice of model and vector store, but the surface area is smaller and sits inside code the buyer owns.
How Governance Failure Resets the Break Even
Every TCO model assumes the agent survives governance review. Enough of them do not that the risk belongs in the math. S&P Global Market Intelligence's 2025 survey of more than 1,000 respondents found that 42% of companies abandoned most of their AI initiatives, up from 17% a year earlier, and the average organization scrapped 46% of proof-of-concepts before production. A failed pilot still carries its sunk seat and integration cost. Deloitte's 2026 survey of 3,235 leaders put mature governance for agentic AI at 21% of organizations, which means most buyers lack the runtime visibility to notice when a platform reprices its action catalog or an in-house agent drifts. Platforms often include audit, approval, and identity plumbing out of the box, shortening the path to sign-off. Builds reach parity when the governance layer is a first-class deliverable, not a month-nine patch. The approach to running AI agents in production treats that layer the way it should be treated.
What Changes This Calculation Over the Next Year
The number most likely to move this decision in 2027 is not a platform list price. It is the cost per million tokens on open-weight reasoning models that match frontier quality for the subset of work most agents actually do: tool calls, retrieval-augmented answers, and bounded multi-step workflows. If that curve keeps bending the way it did through 2025 and 2026, the token-side of the build case gets cheaper faster than the seat-side of the buy case. Platforms will respond by re-pricing credits, renaming tiers, and moving reasoning behind premium SKUs. Model that compression in year one rather than year three, and the break-even window against Copilot Studio, Agentforce, or Glean closes months earlier than the current spreadsheet suggests. The buyer who re-runs the math every quarter catches the inflection; the one who signs a three-year commit at today's rates pays the markup.
