Build vs Buy AI Agents: Why Mid-Market Businesses Choose Custom Development

The buy-vs-build question for AI agents is rarely answered with clear numbers. Vendors lead with per-user pricing that obscures the true cost at scale. Custom development firms lead with capability arguments that obscure the upfront investment. The result: businesses make six-figure decisions based on incomplete data.

This guide provides the complete total cost of ownership (TCO) analysis across the main options — Microsoft Copilot Studio, Salesforce Einstein, and custom LangGraph.js development — so mid-market operations leaders and CTOs can make the decision with actual numbers.

The Landscape: What You're Choosing Between

COTS AI Platforms (Commercial Off-The-Shelf)

Microsoft Copilot Studio is Microsoft's agent-building platform, sitting within the Power Platform ecosystem. It allows non-developers to build conversational agents connected to Microsoft 365 data.

• Pricing: £30/user/month (Copilot Studio standalone); included in Microsoft 365 Copilot at £30/user/month
• Add-on: Message capacity packs at £200/month for 25,000 messages
• Data residency: Processed within Microsoft's cloud; UK/EU regions available but data flows through Microsoft infrastructure

Salesforce Einstein is Salesforce's AI layer, embedded within the CRM platform. Capability ranges from predictive scoring (included in Sales Cloud) to autonomous agents (Einstein Copilot, Agentforce).

• Einstein Copilot: Included in Enterprise and above (Sales Cloud Enterprise starts at ~£150/user/month)
• Agentforce (autonomous agents): £2/conversation or bundled at £50+/user/month on higher tiers
• Practical cost for a 100-user sales team with full agent capability: £60,000–£90,000/year licensing alone

Other platforms: Microsoft Azure AI Studio, AWS Bedrock agents, Google Vertex AI agents — these are builder platforms, not finished products; they require significant development effort regardless, making them more comparable to custom build than COTS.

Custom LangGraph.js Development

Custom agents built with LangGraph.js on Azure infrastructure, developed by a specialist firm (Techseria fixed-fee model):

• Build cost: £18,000–£65,000 per agent system depending on complexity
• Infrastructure: £200–£800/month (Azure Container Apps, GPT-4o API, storage, monitoring)
• Annual maintenance: £2,000–£8,000 per agent system
• Per-transaction cost: Marginal (LLM API tokens + compute), typically £0.01–£0.05 per complex workflow run

Three-Year TCO Comparison

Scenario: AP Automation for a 200-Employee Manufacturing Business

Processing 500 invoices/month, needing 3-way match validation, approval routing, and ERPNext posting.

Option A: Microsoft Copilot Studio

Cost Item Year 1 Year 2 Year 3

Copilot Studio licences (20 finance users) £7,200 £7,200 £7,200

Message capacity packs £2,400 £2,400 £2,400

Implementation/configuration £15,000 £2,000 £2,000

ERPNext connector development £8,000 £0 £0

Internal IT overhead (ongoing) £4,000 £4,000 £4,000

Annual total £36,600 £15,600 £15,600

3-year total £67,800

Note: Microsoft Copilot Studio does not have a native ERPNext connector. Connector development is required regardless.

Option B: Salesforce Einstein / Agentforce

Only relevant if already on Salesforce. If your primary system of record is ERPNext, Salesforce-based agents add significant integration complexity and cost.

Cost Item Year 1 Year 2 Year 3

Agentforce licences (conversation-based, 500 inv/month) £6,000 £6,000 £6,000

Salesforce Enterprise (required base) £36,000 £36,000 £36,000

ERPNext ↔ Salesforce integration £20,000 £0 £0

Configuration and deployment £18,000 £4,000 £4,000

Annual total £80,000 £46,000 £46,000

3-year total £172,000

This is the worst option if you're not already a Salesforce customer — you're paying for a CRM to access its AI layer.

Option C: Custom LangGraph.js (Techseria Fixed-Fee)

Cost Item Year 1 Year 2 Year 3

Build (AP automation agent, ERPNext integration) £28,000 £0 £0

Azure infrastructure £4,800 £4,800 £4,800

Annual maintenance and updates £3,000 £3,000 £3,000

LLM API costs (500 invoices/month) £600 £600 £600

Annual total £36,400 £8,400 £8,400

3-year total £53,200

Custom development reaches cost parity with Copilot Studio by month 14 and costs 22% less over three years — while providing full process fit and data ownership.

Where COTS Platforms Win

The TCO table doesn't tell the whole story. COTS platforms genuinely win in these scenarios:

You're deep in the Microsoft 365 ecosystem and your processes fit the out-of-the-box model. If your workflows involve Microsoft Teams, SharePoint, and Outlook — and don't require complex external system integrations — Copilot Studio is the fastest path to productivity. The platform is designed for this context.

You need to deploy quickly with limited development resource. Copilot Studio can be configured by technically capable non-developers. A custom LangGraph.js agent requires TypeScript developers and takes 8–16 weeks minimum.

Your processes are simple and stable. Rule-based chatbots and FAQ agents don't need custom development. If the agent primarily answers questions from a knowledge base without complex multi-step reasoning, Copilot Studio's cost is justified.

You're a Salesforce-first business with standard CRM workflows. Einstein Copilot adds value within the Salesforce context without additional integration complexity.

Where Custom Development Wins

Process fit is paramount. COTS platforms are built for common processes. Your AP workflow involves a bespoke approval hierarchy, a custom ERPNext implementation, and supplier-specific validation rules that have accumulated over eight years. No platform accommodates this without extensive customisation — and platform customisation within Copilot Studio or Agentforce is constrained, expensive, and creates upgrade risk.

Data residency and compliance requirements. GDPR Article 28 requires a Data Processing Agreement with every sub-processor. Copilot Studio processes data through Microsoft's infrastructure; Agentforce through Salesforce's. For financial services, healthcare, or businesses with UK data sovereignty requirements, custom agents deployed within your own Azure tenant keep data entirely within your control.

High volume makes per-user licensing uneconomical. Copilot Studio at £30/user/month for 150 finance, operations, and procurement staff = £54,000/year in licensing alone. Custom agents have no per-user licensing — the cost is infrastructure and LLM API calls, which scale with usage, not headcount.

You need orchestration across non-Microsoft, non-Salesforce systems. ERPNext, legacy ERP, bespoke supplier portals, industry-specific software — COTS platforms have limited connectors here. Custom agents can integrate with any system that has a REST API or supports file-based data exchange.

IP ownership. Custom-built agents are your intellectual property. The workflow logic, integration patterns, and prompt engineering Techseria develops for your AP automation are owned by you — not licensed from a vendor who can change pricing or deprecate features.

Data Ownership: The Non-Financial Cost

This is the factor most cost analyses omit.

When you use a COTS AI platform, your business process data — invoice contents, customer records, approval decisions, supplier terms — is processed by a third-party's AI infrastructure. Even with appropriate DPAs in place, this creates:

• Training data concerns: COTS vendor policies on whether your data improves their models vary and change. Custom agents with your own Azure OpenAI deployment have explicit guarantees that your data is not used for training.
• Competitive intelligence exposure: For M&A-sensitive businesses, the contents of processed documents passing through vendor infrastructure is a risk.
• Regulatory audit complexity: Under UK FCA, ISO 27001, or SOC 2 frameworks, demonstrating controls over AI processing is simpler when the AI runs in your own environment.

The Customisation Ceiling

Both Copilot Studio and Agentforce hit a ceiling when your requirements diverge from their intended use cases.

Copilot Studio limitations:

• Multi-agent orchestration (multiple specialised agents coordinating on complex tasks) is possible but requires Power Platform architecture that adds significant cost and complexity
• Custom reasoning patterns beyond the platform's built-in flow types require workarounds
• Long-running processes (hours to days) with state persistence are not a strength of the platform

Agentforce limitations:

• Tightly coupled to Salesforce data model — orchestrating against external systems requires extensive connector work
• Pricing per conversation makes high-volume automation scenarios expensive
• Workflow customisation outside the Einstein framework requires Apex development

A custom LangGraph.js agent has no such ceiling. The graph architecture supports arbitrary complexity: multi-agent orchestration, long-running state persistence across days, conditional logic of any depth, and integration with any API.

When to Revisit This Decision

The build vs buy calculus shifts when:

• Your process volume grows significantly — what was affordable at 50 users becomes expensive at 200
• COTS vendor pricing changes — both Microsoft and Salesforce have increased AI licensing costs 20–40% in the last 24 months
• Your compliance requirements tighten — new regulations may change the data residency calculus
• Your processes become more complex — what started as a simple chatbot evolves into a multi-step orchestration that the platform can't handle well

We recommend mid-market businesses revisit the build vs buy decision annually, particularly as LLM infrastructure costs continue to fall (GPT-4o costs have dropped 60% since launch) making custom development economics increasingly favourable.

Making the Decision

The right answer depends on your specific process complexity, user count, compliance requirements, and existing technology investments. Neither option is universally superior.

What is universally true: the decision made with a vendor's sales presentation and without a structured TCO analysis over three years is almost always the more expensive decision.

[Get a Fixed-Fee Quote from Techseria](/contact) — we'll provide a detailed build vs buy analysis for your specific automation requirements, including a three-year TCO comparison, within 5 business days. No obligation.