Enterprise AI Agent Build vs Buy: The 2026 Decision

The enterprise AI agent market in May 2026 presents a genuine build-vs-buy decision for the first time: packaged platforms — Salesforce Agentforce, Microsoft Copilot Studio, SAP Joule, ServiceNow Now Assist, and Oracle AI Agent Studio — now compete against custom stacks built on Claude Agent SDK plus Amazon Bedrock, the OpenAI Agents SDK plus Azure OpenAI, or Google Vertex AI Agent Builder, with real TCO differences that flip sign at approximately one million agent conversations per year.

That crossover matters because both sides of the market now converge on the same underlying model layer. SAP Joule runs on Claude as its primary reasoning engine. Agentforce routes through Claude under the Einstein Trust Layer. The model choice is no longer separable from the platform choice — which means CIOs evaluating buy-side platforms are simultaneously making implicit decisions about model contracts, data planes, and per-conversation pricing convexity that compound over a three-year horizon.

This guide frames the decision across twelve dimensions: the decision logic, the crossover mechanics, four buy paths, three build paths, the May 2026 vendor-consolidation wave (three of four Big Four firms standardized on Claude), and three-year TCO scenarios at 100K, 1M, and 10M sessions per year. All TCO figures are illustrative-arithmetic — every cell shows the math that produced it so you can substitute your own assumptions.

The classic build-vs-buy framing — "do we write it in-house or buy a vendor license?" — does not capture what enterprises actually face in mid-2026. Every buy-side platform now lets you customize agents with code (Joule Studio, Agentforce Builder, Copilot Studio low-code/pro-code). Every build-side path now ships orchestration primitives, hosted runtimes, and managed observability. The code is everywhere.

The meaningful question shifted to: where does the agent contract live?Who owns the agent’s surface area? Who owns the data plane? Who owns the model contract? On the buy side, the vendor owns the agent surface (you configure within their schema), the data plane runs inside their infrastructure (Salesforce Data Cloud, SAP Knowledge Graph, Microsoft Graph), and the model contract is theirs to renegotiate at renewal. On the build side, your engineering team owns all three — and absorbs the corresponding capital and ongoing ops cost.

This means the decision is fundamentally a make-or-buy of ownership, not of code. The TCO model below prices that ownership transfer explicitly across three volume bands — because the cost of ownership is volume-dependent in a non-linear way.

Per-conversation pricing grows linearly with volume. Salesforce Agentforce at $2/conversation: 100K sessions cost $200K/year, 1M sessions cost $2M, 10M sessions cost $20M before any volume negotiation. That is a convex cost curve from the buyer’s perspective — every conversation costs the same marginal dollar regardless of your cumulative spend.

Per-token pricing on the build side is concave in practice. The marginal cost of conversation number 500,001 is the same token rate as conversation number one, but engineering teams can reduce the average token footprint over time through prompt compression, model routing (Haiku 4.5 at $1/$5 per Mtok for classification steps vs Sonnet 4.6 at $3/$15 for reasoning steps), KV cache reuse, and RAG optimization. The result: total inference spend grows more slowly than linear with usage once the stack is mature.

The crossover sits roughly at one million conversations per year for a 500-seat enterprise using a complex workflow agent (cross-system order-to-cash, customer service with tool calls, multi-step HR workflows). Below that, the buy path wins on speed and the absence of engineering overhead. Above it, the three-year arithmetic favors build by a margin that widens with every additional million sessions.

The bars above make the convexity asymmetry visible. Agentforce and Claude SDK both appear close at 100K sessions — the buy path looks favorable once you factor in engineering overhead. At 1M sessions the build path begins to approach list-price parity in Year 1. By Year 3, steady-state build engineering costs shrink while per-conversation buy costs hold flat. At 10M sessions, the gap is structural.

The analysis that follows prices each vendor path at each volume band in the TCO section. First, the four buy paths deserve individual treatment because their pricing structures differ materially from each other.

Microsoft’s enterprise agent offering has two pricing surfaces that stack. The base Microsoft 365 Copilot seat license — anchored at $30/user/month at last published list price (verify current pricing on microsoft.com) — gives users access to Copilot in all M365 apps plus opt-in to the Frontier program, which includes early access to capabilities like Microsoft 365 Copilot Cowork (released March 30, 2026). The seat cost is the floor; for a 500-seat deployment that is $180K/year before any agent-specific charges.

Copilot Studio is a separate metered product. The widely-cited list anchor is approximately $200/month for a 25,000-message pack with $10/1,000-message overage — plus a pay-as-you-go per-message SKU (verify on the Microsoft Copilot Studio pricing page before signing; Microsoft has revised pack sizes). At 100K messages per month, the metered cost is approximately $40K/year in overage above the included 300K messages from the pack.

The multi-model tax matters here. Our analysis of Copilot Cowork’s enterprise agent workflows found that Critique mode adds approximately 20% premium over a single-model baseline, while Council mode (multi-agent deliberation) runs approximately 2.5× the baseline cost. High-complexity agentic workflows on Copilot Studio can therefore land well above the per-message list price on an effective-per-conversation basis. Verify your specific workflow’s multi-agent mode usage before modeling TCO.

The structural advantage of the Microsoft path is deep M365 integration — agents that need to act on SharePoint, Teams, Outlook, or Dynamics 365 workflows carry zero additional integration engineering cost. For organizations already operating at M365 E3/E5, the incremental cost of Copilot Studio agents is lower than any alternative.

Salesforce Agentforce’s published list anchor has been $2 per conversation — "a complete interaction session between the agent and a user" as Salesforce defines it. A conversation is not a message or a token; it is the entire session. At $2/conversation, a 1M-session/year deployment costs $2M/year in list-price conversation charges, plus Salesforce Data Cloud capacity for grounding.

Salesforce has signaled a move toward Flex Credits pricing (~$0.40/credit, approximately one credit per message rather than per conversation) — verify the current model on salesforce.com/agentforce/pricing before any TCO modeling, as this structural change would materially alter the per-unit economics. Our Salesforce Agentforce CRM automation guide uses the $2/conversation anchor as of its publication date; reconcile with the live pricing page before signing.

Volume discounts are negotiated via Salesforce account executives; large deployments (5M+ conversations/year) typically negotiate to $1.00–$1.50/conversation per our understanding of published ranges. Even at $1.25/conversation, 10M sessions/year is $12.5M — still a wide gap above the build path at the same volume.

The compelling case for Agentforce is its native CRM integration. For customer service automation where agents need to read and write Salesforce objects — cases, contacts, orders, knowledge articles — the zero-integration cost is a genuine advantage that the TCO model partially captures but cannot fully quantify. If 80% of your agent interactions are CRM-native, the Agentforce path avoids months of integration engineering.

SAP announced the Autonomous Enterprise on May 12, 2026 at Sapphire Orlando: 50+ Joule Assistants orchestrating 200+ specialized agents across Autonomous Finance, Spend, Supply Chain, HCM, and CX — with Claude as the primary reasoning and agentic capability via a direct SAP-Anthropic partnership announced the same day. The SAP AI Agent Hub will be generally available Q3 2026 at no additional charge inside SAP Business AI Platform.

The SAP Joule path has an important prerequisite: it only makes sense if your organization is already on RISE with SAP or SAP GROW. The agent capability is bundled into those subscription tiers; for a non-SAP shop, Joule is not available as a standalone purchase. The Takeda case study (vendor-reported) cites up to 10% productivity gains and up to 25% reduction in stock-out revenue loss — treat these as directional signals from SAP’s own materials, not audited benchmarks.

As Holger Mueller of Constellation Research observed on May 12, 2026: “It’s the first time on this side of the millennium that SAP has a vision for ERP. And while the Autonomous Enterprise is not a unique vision, SAP has a compelling start to deliver it.” The vision is real; verify execution timelines against SAP’s published roadmap before committing.

ServiceNow Now Assist is sold as a Pro Plus edition layered on existing Now Platform subscriptions (ITSM, CSM, HRSD, SecOps). Pricing is typically a percentage uplift on the base subscription, negotiated per contract — verify the current uplift band on servicenow.com. For organizations running ServiceNow at scale for ITSM or employee service management, Now Assist is the lowest-friction path to agentic automation because the agents operate natively inside the existing Now Platform workflow graph. For organizations not already on ServiceNow, there is no economic case for Now Assist as an AI entry point.

Oracle AI Agent Studio (announced at Oracle CloudWorld 2025) is the build-and-deploy surface for AI agents inside Oracle Cloud HCM, ERP, CX, and SCM. Pricing is bundled inside the underlying Fusion Cloud Application subscription with a separate “AI consumption” line for high-volume usage — verify the AI consumption unit on oracle.com/artificial-intelligence/ai-agents before modeling TCO. As with SAP Joule, Oracle AI Agent Studio is a coherent choice only if you are already a Fusion Cloud Applications customer.

Both paths follow the same structural logic: the agent capability is a value-add inside an existing vertical ERP/ITSM subscription, not a standalone AI platform. This makes the marginal cost assessment straightforward — what matters is the uplift, not the base subscription cost your organization already pays.

The Claude Agent SDK (renamed from Claude Code SDK on September 29, 2025) is the canonical Anthropic build surface for production agents. Available as npm @anthropic-ai/claude-agent-sdk and PyPI claude-agent-sdk, it ships orchestration primitives, tool-call scaffolding, and tracing hooks. Amazon Bedrock with Cross-Region Inference is the recommended production deployment path — it provides geographic profile options and the managed infrastructure to serve frontier models at enterprise scale.

Current Claude model pricing (verified May 2026 via platform.claude.com/docs/en/about-claude/models): Sonnet 4.6 at $3/$15 per million tokens (input/output), Opus 4.7 at $5/$25 per million tokens with 1M flat context, Haiku 4.5 at $1/$5 per million tokens at 200K context. A data-residency inference_geo: "us"flag adds a 1.1× multiplier on 4.6+ models. Fast Mode (beta) runs at 6× the standard pricing tier on Opus 4.6/4.7 — relevant for latency-sensitive workflows.

The build path is the 2026 default for three reasons: (1) Anthropic reached 34.4% business AI adoption in April 2026 — the largest market share among foundation-model providers in enterprise tooling; (2) three of four Big Four firms built their enterprise AI stacks on Claude, creating a practitioner knowledge pool that accelerates custom implementation; (3) the per-token economics of Sonnet 4.6 at a 30,000-average-token conversation cost approximately $0.54/conversation at blended input/output rates — one-quarter of Agentforce list pricing for comparable complexity. Explore our detailed analysis of Claude Agent SDK production patterns for the implementation architecture.

The OpenAI Agents SDK (Python package openai-agents) is the production successor to OpenAI’s archived Swarm project. It ships multi-agent orchestration via “handoffs,” guardrails, and tracing as open-source primitives. Azure OpenAI is the recommended production deployment surface for organizations with Microsoft enterprise agreements or data-residency requirements tied to Azure regions.

Current OpenAI model pricing (verified May 2026): GPT-5.5 at $5/$30 per Mtok standard (with a long-context surcharge above 272K input tokens); GPT-5.4 at $2.50/$15; GPT-5.4 mini at $0.75/$4.50; GPT-5.4 nano at $0.20/$1.25 — all via developers.openai.com/api/docs/pricing. At GPT-5.4 rates ($2.50/$15), a 30,000-token conversation runs at approximately $0.49/conversation in blended input/output cost — comparable to Claude Sonnet 4.6.

The OpenAI Agents SDK path benefits from the OpenAI Deployment Company ecosystem: the $4B+ professional services vehicle (launched May 11, 2026, 19-firm consortium led by TPG) creates an implementation partner network for organizations that need forward-deployed engineers rather than internal build capacity. For organizations whose deployment partner is OpenAI-aligned, this may compress implementation timeline enough to shift the build-path economics toward Year 1 parity with buy paths.

Google Vertex AI Agent Builder is the Google Cloud platform-level surface for building enterprise agents — search, conversational, RAG, and multi-modal. Pricing is per-query for search-based agents and per-token for generative agents, with separate grounding charges. The grounding cost is material for RAG-heavy workflows: every document retrieval call adds a per-grounding-query charge on top of generation token costs. Verify current rates on cloud.google.com/agent-builder/pricing before TCO modeling.

The Google path makes structural sense for organizations already operating significant GCP workloads. Vertex AI Agent Builder integrates natively with BigQuery (for structured data grounding), Cloud Storage, and Google Search grounding — advantages that are hard to replicate on AWS or Azure without additional integration engineering. Gemini 3 Flash (GA December 17, 2025) and Gemini 3.1 Flash-Lite (GA May 8, 2026) are the cost-optimized models for high-volume agent deployments on Vertex; Gemini 3.1 Pro (preview as of February 19, 2026) is the frontier option.

The Vertex path is currently the weakest in terms of practitioner ecosystem size relative to Claude Agent SDK and OpenAI Agents SDK. Google’s Gemini Code Assist (free tier from March 2026) and the google-gemini/gemini-cli are building developer mindshare, but the dedicated agent-orchestration community is smaller. Factor onboarding time into the Year 1 TCO.

The most consequential buy-side signal of May 2026 is not a vendor announcement — it is a client-side co-selection effect. When three of four Big Four professional services firms standardize on a foundation model, every Fortune 1000 company that relies on Big Four implementation partners inherits that choice in practice.

The vendor co-selection effect works as follows: a Fortune 500 CIO whose primary implementation partner is Deloitte, PwC, or KPMG will receive Claude Agent SDK patterns, Claude-trained forward-deployed engineers, and Claude-native tooling as the default recommendation from their partner. The EY path (Microsoft-first) similarly favors Copilot Studio. For most organizations, the question “which build stack?” is partly answered by “which Big Four firm do we use?”

This is a structural shift in how enterprise AI decisions get made. The May 2026 wave compressed the timeline: Deloitte (October 2025) set the first Big Four precedent; PwC and KPMG followed within five days of each other in May 2026. For CIOs who have not yet made a vendor commitment, their implementation partner’s choice is now a meaningful input into the build-vs-buy decision — and specifically into the choice of build stack if they go custom.

All figures below are illustrative-arithmetic: every cell shows the formula that produced it so you can substitute your own assumptions. Engineering costs assume a loaded rate of $20K/month per senior engineer. Token costs use published model pricing. Do not treat these as audited industry averages — verify vendor pricing pages before any procurement decision.

Three scenarios: (A) simple Q&A agent / internal HR knowledge, 100K conversations/year; (B) complex workflow agent / cross-system order-to-cash, 1M conversations/year; (C) high-volume customer-facing CX agent, 10M sessions/year.

The TCO math above resolves cleanly at the volume extremes. The middle band (500K–2M conversations/year) is where the decision depends on factors the math cannot fully capture: integration complexity, data sovereignty requirements, internal engineering capacity, and deployment timeline. The framework below organizes those factors.

The MCP server build-vs-buy TCO framework extends this analysis to the infrastructure layer — the choice of MCP server hosting (self-managed vs vendor-managed) follows a similar crossover logic at a lower dollar threshold. For organizations navigating the full stack decision, combining the agent-level and MCP-level TCO models gives a more complete picture of true three-year cost.

For the implementation side once you have made the decision, our AI transformation services and CRM automation services cover both the build (Claude Agent SDK + Bedrock architecture and deployment) and the buy (Agentforce and Copilot Studio configuration and integration) paths.