← Back to Blogs
November 16, 2026By Cosmoneural Insights

The Enterprise Architect Is Now a Supply Chain Manager for Intelligence

Enterprise ArchitectureAI StrategyVendor Risk
The Enterprise Architect Is Now a Supply Chain Manager for Intelligence

Photo by Conny Schneider on Unsplash

Most descriptions of the enterprise architect's role in an AI-first operating model still assume the architect is designing systems. Increasingly, the architect is procuring behaviour. The intelligence in your workflows is bought, rented, and re-rented on someone else's release schedule, and the capability you depend on today may be deprecated, repriced, or quietly regressed before your next planning cycle closes. That makes the EA less of a designer and more of a supply chain manager: responsible for sourcing, substitution, concentration risk, and the cost of changing suppliers. It is a less glamorous framing than "AI visionary," and a far more useful one.

The build/buy line moved into someone else's release notes

Supplier churn at this layer is unusually fast. Menlo Ventures' 2025 State of Generative AI report put enterprise spend on generative AI at roughly $37B and showed how quickly share moves between labs — Anthropic climbing to around 40% of enterprise LLM API usage while OpenAI's share fell to about 27%. Those are not slow-moving market dynamics; they are quarterly ones. Meanwhile a Stack Overflow survey of enterprise teams found that single-provider estates have become rare, with only a small minority of organisations using one model vendor. The architectural implication is blunt: you are already running a multi-supplier estate, whether or not anyone designed it that way.

Substitution cost is the metric, not model choice

Which model is "best" is a benchmark question with a shelf life of weeks. Which model you can leave is an architecture question with a shelf life of years. The same Stack Overflow research found that a large majority of organisations have delayed, scaled back, or restructured a migration because switching costs looked too high — the classic signature of supplier lock-in, now applied to intelligence. And the lock-in rarely lives in the API call. It lives in everything wrapped around it:

  • Prompt and tool scaffolding tuned to one provider's quirks, with no abstraction between orchestration logic and provider SDK.
  • Retrieval and context plumbing — embeddings, chunking, permission filters — rebuilt per vendor rather than owned as a shared asset.
  • Evaluation history that exists only as screenshots and anecdotes, so nobody can prove a replacement is equivalent.
  • Human process — review queues, escalation thresholds, operator habits — calibrated to one model's failure modes.
  • Contractual gravity: committed spend, co-engineering arrangements, and vendor-delivered agent code your own teams cannot evolve.

That last one is now a named risk. Gartner has predicted that 70% of enterprises will abandon agentic AI built through vendor forward-deployed engineering by 2028, trapped by cost and unable to evolve the systems themselves. Fast delivery that your teams cannot maintain is a liability disguised as velocity.

Evals are the only acceptance test that travels

A supply chain manager does not take a supplier's word on quality; they hold an incoming inspection standard. The EA equivalent is a task-level evaluation suite owned by the enterprise, not the vendor: a frozen set of representative cases, expected outputs, tolerances, and cost-per-task ceilings for each workflow that uses a model. If you have that, swapping suppliers becomes a test run. If you don't, every swap becomes a re-implementation project with political risk attached, which is precisely why most teams stay put. This is the same discipline we apply in architecture and AI engagements: make the acceptance criteria portable before you make the dependency deep.

Second-sourcing without doubling the stack

Second-sourcing everything is waste. Classify instead. For high-volume, low-variance tasks — classification, extraction, summarisation inside a bounded domain — maintain a tested alternate supplier and a documented cutover. For frontier-dependent work where only one provider currently clears the bar, accept the single source deliberately, record it as a concentration risk with an owner, and keep the surrounding scaffolding neutral. In regulated and SAP-heavy estates this matters twice over, because the orchestration sits beside core process systems that outlive any model generation by a decade.

What this changes in the operating model

Three practical shifts. Architecture review should ask "what does it cost to replace this supplier?" alongside "does it meet the requirement." Vendor management should inherit eval results as a standing input, not a one-off bake-off. And the portfolio view should track supplier concentration per capability, the way you already track single points of failure in infrastructure.

The labs will keep leapfrogging each other, and pricing will keep moving. Architects who treat intelligence as a commodity input with managed suppliers will absorb that churn as a routine refresh. Those who treat their current model as a platform decision will spend 2027 explaining why a migration everyone agreed was necessary never happened.

Frequently Asked Questions

How do we start building a portable evaluation suite without a large investment?

Begin with your two or three highest-volume AI workflows and capture 50-100 real cases with agreed correct outputs and acceptable tolerances. Store them as versioned test data owned by the business function, not the vendor, and run them on every model or prompt change. That alone converts most model swaps from a project into a test cycle.

Doesn't a provider-neutral abstraction layer cost performance and feature access?

Yes, a thin abstraction will lag provider-specific features by weeks or months, and for frontier-dependent use cases that trade-off may not be worth it. The practical compromise is to abstract orchestration, retrieval, and logging while allowing deliberate provider-specific code in a small number of clearly flagged places. Document those places so the exit cost stays visible.

Where should responsibility for AI supplier concentration risk actually sit?

It works best as a shared accountability: enterprise architecture owns the technical substitution cost and the eval standard, while vendor management owns commercial terms and exit clauses. One named person should hold the consolidated view per capability so concentration risk appears in portfolio reviews rather than surfacing during a pricing change or deprecation notice.

Comments (0)

Leave a Comment

No comments yet. Be the first to comment!