In 2024, enterprises split their AI investments almost evenly - 47% build, 53% buy. By 2025, that reversed to 24% build and 76% buy, according to Menlo Ventures' annual enterprise AI survey. The commentary around that shift has been almost entirely vibes: "packaged tools are maturing," "internal builds are expensive," "speed wins." What almost no one states plainly is where the math actually flips - and what the 95% failure rate tells you about why most teams are asking the wrong question in the first place.
The volume threshold nobody prints on a slide
The build-vs-buy AI agent question is not primarily about philosophy - control, lock-in, customization. It is about volume and unit economics, and there is a specific crossover point.
The packaged platforms - Salesforce Agentforce, Microsoft Copilot Studio, SAP Joule - compete against custom stacks built on SDKs like Claude or OpenAI. The TCO crossover sits at approximately one million agent conversations per year. Below that, buy-side platforms win on time-to-value and integration simplicity. Above it, per-token build economics begin to undercut per-conversation list pricing on packaged platforms.
Run the arithmetic yourself using Agentforce's public pricing. Agentforce in 2026 offers four billing meters: Flex Credits at $0.10 per action, $2 per conversation, $2 per resolved service case, or per-user licenses from $5 to $550 per month. On the conversation meter: 1 million conversations a year is $2 million. A typical conversation involves 5 to 15 actions. At $0.10 per action, that puts the effective cost between $0.50 and $1.50 per conversation via Flex Credits, compared to a flat $2 on the per-conversation plan.
Meanwhile, a custom stack running Claude Sonnet-class models on a major cloud provider costs roughly $3 per million input tokens and $15 per million output tokens - and a typical customer-facing conversation consumes well under 10,000 tokens. At scale, that matters. Before scale, the overhead of building, evaluating, maintaining, and iterating on custom agent infrastructure eats the savings whole.
Rapid model evolution, ongoing retraining costs, and model drift are making internal AI development far more expensive to maintain than many teams originally expected.
The crossover is real, but it is a different number for every team depending on average conversation complexity, how many parallel agent types you are running, and whether your stack needs private data access that would require custom retrieval layers regardless of which deployment model you choose.
Why the 95% failure rate is not an argument against buying or building
MIT's Project NANDA released a July 2025 report finding that despite $30-40 billion in enterprise investment, 95% of generative AI projects yield no measurable business return. That number gets cited constantly as a caution against ambitious AI projects. It is actually an argument against picking the wrong workflow.
The statistic is real. The popular reading - "AI does not work for business" - is not what the study says. MIT measured something narrow and brutal: whether integrated pilots moved the P&L. Most did not.
The pattern the data reveals is simpler than the coverage suggests. Teams are deploying agents against workflows that are generic enough that the agent adds no differentiation, but complex enough that they fail without significant investment in tooling and prompt engineering. AI products may be easy to start building, but they are high-maintenance software. They require continuous model upgrades, prompt tuning, evaluation, infrastructure optimization, and UX iteration. For most non-core workflows, buying makes far more sense because the cost of innovation and maintenance is amortized across many customers instead of being borne by a single company.
That last sentence is the actual answer. If the workflow is generic - customer support triage, internal lookup, meeting summaries - someone else is already paying to tune the model for exactly that use case. Your build adds cost, not advantage.
Beagle in action#cs-ops, 2:47pmThe steel-man for building
The buy-first consensus deserves a challenge. There are workflows where building pays, and they share three traits.
First, the agent operates on genuinely proprietary data - not just internal data, but data whose patterns represent a real competitive edge. A commercial lender's credit decisioning logic, a specialist insurer's claims-routing rules, a manufacturing firm's equipment fault taxonomy. Build custom when the agent runs on your proprietary data and workflows and becomes a competitive moat.
Second, volume. Businesses that build custom AI agents see slower initial ROI - typically 12 to 24 months - but achieve better long-term returns because they own the infrastructure and can scale without increasing vendor costs. At a million conversations a year, the per-conversation meter starts hurting. At five million, a well-built custom stack can be materially cheaper.
Third, regulatory or data residency requirements that a cloud-hosted vendor platform simply cannot satisfy. Some regulated industries - healthcare, finance, government - require that data processing occurs within specific geographic or network boundaries. An on-device or on-premises model satisfies these requirements without custom infrastructure for a frontier LLM.
Most successful enterprises in 2026 run hybrid patterns, using vendor platforms for 80% of workflows and custom agents for the critical 20%. That is not a cop-out answer. It is the right unit of analysis: workflow by workflow, not company-wide.
What the pricing complexity of packaged agents actually signals
One genuinely non-obvious insight: the fact that Agentforce now has four billing models - the original $2 flat per conversation is no longer the only option; in May 2025 Salesforce introduced Flex Credits, a consumption-based system where businesses pay per action rather than per conversation
- tells you something about the maturity of the market, not just the pricing.
When a vendor goes from one simple pricing model to four, it is usually because usage patterns turned out to be wildly more varied than the initial model assumed. Complex multi-step workflows (say, a refund that touches four backend systems) cost the same $2 as a one-message FAQ answer under conversation pricing. That is a subsidy from complex to simple, and Salesforce discovered enterprises would rather have granularity than simplicity.
The implication for the build-vs-buy decision: what counts as a single "conversation" when one customer query triggers several backend processes was never entirely clear. Before you model the TCO of any packaged agent platform, you need to map how many actions your average workflow actually triggers. The number almost always surprises teams on the upside.
Build vs buy AI agents: common questions
What is the main factor in the build vs buy AI agent decision?
Volume is the dominant factor once you control for workflow type. Below roughly one million agent conversations per year, packaged platforms like Agentforce or Copilot Studio win on time-to-value and lower operational overhead. Above that threshold, per-token build economics on a custom stack typically undercut per-conversation list pricing. Data residency and proprietary workflow requirements can override the volume math entirely.
Why do most enterprise AI agent pilots fail?
MIT's Project NANDA found that despite $30-40 billion in enterprise investment, 95% of generative AI projects yield no measurable business return. The study identifies a divide where only 5% of pilots deliver ROI, with success hinging on adaptive, learning-capable systems rather than static tools. The more common failure mode is simpler: teams automate workflows that are generic enough to buy but too narrow to justify the maintenance cost of building, so the pilot stalls in production limbo.
When does building a custom AI agent actually pay off?
Build earns its complexity when three conditions hold simultaneously: the workflow operates on genuinely proprietary data that creates competitive advantage, volume is high enough that per-token economics beat per-conversation vendor pricing, and someone in your organization will own and maintain the system long-term. Missing any one of those three, the build almost certainly underperforms a bought alternative within its first 18 months.
How does Agentforce pricing affect the build vs buy calculation?
Agentforce pricing offers four ways to pay: Flex Credits at $0.10 per action, $2 per conversation, $2 per resolved service case, or per-user licenses from $5 to $550 per month. The choice between meters matters as much as the headline per-conversation rate. A workflow that triggers 20+ actions per conversation is effectively the same price under Flex Credits as the flat conversation rate - at that point, the packaging advantage of a vendor platform starts narrowing, and a custom build's economics look more competitive.
Is the hybrid approach (buy some, build some) actually viable?
Yes, and it is where most serious deployments land. Buy a platform that exposes its logic layer and build your differentiated agent behavior on top of it. You get deployment speed without sacrificing the workflow control that matters. This is where most serious enterprise deployments land in 2026. The mistake is treating buy-vs-build as a single company-level choice rather than a workflow-by-workflow judgment repeated every time a new automation case surfaces.