Build vs. buy: build your own AI agent or pick a platform

It's the same question as ERP, and you know how that went

Twenty years ago many companies built their own ERP. In 2026 nobody does, or it's a deliberate exception with a large dev team. The same discussion is now happening around AI agents. Building it yourself is attractive because your domain feels so specific that a platform seems unlikely to fit. Until you realise how many non-functional requirements you're overlooking.

This isn't an ad for platforms. There are organisations that rightly build themselves, because they have the scale and team, or their process is so idiosyncratic that a platform really doesn't fit. But for 80 percent of cases the honest comparison is different from what teams initially assume.

What you underestimate when building yourself

The prompt and model usage is a small slice of the work. The rest is what a professional security and operations engineer would call a platform: identity, RBAC, audit, observability, secret management, rate limits, retry logic, graceful degradation, model version control, an evaluation harness, change management on prompts and tools. And you have to keep all of that current as models and best practices change.

Non-functional requirements that only become clear in production:

• Audit trail per agent action at user level that an auditor can verify.
• Idempotent tool calls so retries don't create duplicate postings or emails.
• Observability that shows per agent run which tools were called when, with which parameters and which outcome.
• Security tests for prompt injection and credential exfiltration in tool calls.
• An evaluation harness with test cases that re-validates every prompt change.
• A rollback path for model updates that perform just a bit differently.

What a platform brings, and what it costs

An agentic platform brings those building blocks: identity integrations, guardrails, observability, audit trail, model routing. The cost is shared across all customers of the platform, so per organisation it's a fraction of building yourself. The price is in vendor dependence: you accept that the platform sets the pace you go along with.

When building yourself does add up

Three situations where building yourself defends itself: you have an R&D team of more than ten engineers specifically set up for this, your domain has a unique data structure or rule set platforms demonstrably don't support, or agents are an end product for you rather than an internal improvement. In those three cases the R&D investment is justified by the value.

How to have the conversation without ideology

Start with the non-functional requirements, not the prompts. Who's responsible for security, observability, audit and rollback shapes what's realistic. Then estimate per route: time to production, ongoing cost, and how you switch model or vendor in two years. Plan a Quick Scan if you'd like to have that conversation with someone who has built both sides.