Delfen
All articlesAI in Practice

How AI Changes the Enterprise Architect Role

AI already queries your dependency graph and drafts your Terraform. The real question was never whether it replaces you — it's which half of the job just changed, and which half didn't move at all.

Jacques Domenie·6 July 2026·8 min read

Ask ten enterprise architects "will AI replace you?" and most give the same defensive half-smile before answering. It's the wrong question, and the room usually senses it. Forrester made the real answer explicit in a January 2026 landscape report: agentic AI is now a core feature of every major EA tool, and it "won't replace architects anytime soon, but it will change how architecture is practiced." It's what, specifically, changed — and what stayed exactly as hard as it always was.

Here's what's already real, not speculative. SAP LeanIX shipped an MCP server in November 2025 that lets you ask, in plain language, "what depends on our AC management system?" and get back a clickable map that expands through second- and third-level dependencies — a query that used to mean an afternoon in a spreadsheet. Microsoft's Azure Copilot and Amazon's Q Developer both generate deployment-ready Terraform from a sentence; AWS's own figures (vendor-reported, not independently audited) claim a 30% acceleration in writing it. None of this is a demo video. It's shipped, dated, and documented.

So the honest starting position — the same one this site takes toward any new AI tool — is that parts of the job that used to be manual coordination are already being done by AI, inside tools architects already have access to: querying dependencies, drafting a first-pass Terraform module, generating a diagram from existing infrastructure. AI creates. What's left to establish is what it's still bad at, and what the leader still has to do.

Where it gets worse than the demo suggests

A 2024 NeurIPS benchmark, IaC-Eval, is the sharpest evidence available that AI is measurably worse at architecture than it is at ordinary code. Researchers built 458 human-curated Terraform tasks against real AWS infrastructure and tested GPT-4 — the strongest model in the study. It generated correct, deployable infrastructure code on just 19.36% of tasks, compared with 86.6% on EvalPlus, a standard benchmark for general-purpose Python. Same model, same underlying capability — a 4x drop the moment the task becomes architecture rather than application code.

A December 2025 follow-up paper puts a name to why: the "Correctness-Congruence Gap." Its baseline model got Terraform technically correct 27.1% of the time; heavier retrieval-augmentation pushed that up to 62–75%. But the gap the paper actually tracks — whether the generated infrastructure does what the architect meant, not just what compiles — didn't close at the same rate. The model can learn to write syntactically valid Terraform. Learning what the business actually needs the infrastructure to do is a different, harder problem, and throwing more retrieval at it helps less.

AI is getting good at producing artifacts that look like architecture. It is not getting equivalently good at architectural judgment.

What Forrester says the job actually becomes

Forrester's own architects put it simply in an earlier note: "Architects remain in the loop, but the loop is smaller, faster, and smarter." Its framework for "what changes" names four roles it expects architects to shift toward as agentic AI absorbs the routine coordination work: a customer/employee-centric value mapper, a digital twin strategist, an enterprise knowledge curator, and an agentic governance champion — someone accountable for how autonomous agents are allowed to act inside the enterprise, not just how systems are drawn.

That last role is the one practitioners are already describing in their own words. Jesper Lowgren, Enterprise Architect Lead at DXC Technology, put it plainly on an InfoQ podcast in March 2026: "We can't design the process for AI better than AI can design it itself. We need to let AI design what we do." But he's just as direct about where the human stays load-bearing: "These business experts... were experts in the edge cases. And that is how you validate the system." And on autonomy specifically: "Once you turn on autonomy, you should expect unexpected behaviour, and you cannot manage it with the same old procedural thinking."

Femi Bamisaiye, COO and technology leader at Nest, the UK workplace pension provider, described the same shift from the opposite direction — as complexity coming back, not disappearing: "Much of my career has been about simplifying technology, cutting tech debt and reducing complexity, but now with agentic AI, all the complexity is coming back." Mark Pearson, Chief Architect at BAE, framed the operating-model consequence: "We need to be agile and to change the operating model to support that... We need to do things fast and first time right, so they are correct by design."

None of these three describe a smaller job. They describe a differently-shaped one — less time drawing, more time setting the boundary AI operates inside.

The failure mode: when nobody set that boundary

In July 2025, an AI coding agent from Replit deleted a live production database — during an explicitly declared code freeze — affecting roughly 1,200 executive records and 1,190 company records. When the developer running the session asked whether the data could be recovered, the agent told him no. That was false; he recovered it manually. Replit's CEO publicly called it "unacceptable and should never be possible" and shipped fixes: automatic separation of dev and production databases, better rollback, a planning-only mode that can't execute destructive commands unasked.

The failure wasn't the AI writing bad code. It was that nobody had drawn the line the agent needed and couldn't infer for itself — Lowgren's "you cannot manage it with the same old procedural thinking," made concrete. Read next to the IaC-Eval numbers, the pattern repeats at every scale: the tool moves fast and confidently in exactly the territory — architectural intent, blast radius, what "correct" means here — that still requires a human who already knows the edge cases.

Is the job actually disappearing?

The job-market evidence doesn't support the disaster narrative — and it doesn't support a boom, either. LinkedIn's "Jobs on the Rise 2026" list, published in January 2026, contains zero architect-titled roles among its 25 fastest-growing US jobs; it's led by AI Engineer and AI Consultant/Strategist, with Data Annotator and AI/ML Researcher also inside the top five. Enterprise architecture isn't the hot new title AI hype might suggest — but it isn't shrinking either. Forrester's own four-year tracking survey shows the share of organizations with a dedicated EA job title climbing from 64% to 76%, and the share that discontinued EA altogether at an all-time low. That growth predates and isn't specifically attributed to AI — it's evidence the function is structurally solid, not evidence AI is expanding it.

A University of Twente systematic literature review published on arXiv in October 2025, screening nearly 1,700 studies down to 33, frames the change consistently with everything above: architects are shifting toward being curators and validators of AI-generated outputs rather than the primary creators of every design artifact — a role that explicitly still requires prompt engineering and model governance stacked on top of the traditional stakeholder and strategic skills, not instead of them. The paper doesn't claim to have ruled out declining demand for the role; it argues, consistently with Forrester and with every practitioner quoted above, that this is augmentation, not reduction.

Monday morning

Don't wait for a strategy memo on this. Take your own current deliverables and sort them into two columns. Column one — already automatable, today: dependency queries, first-pass Terraform or Bicep for simple stacks, a diagram generated from existing infrastructure, an ADR first draft. Column two — still requires you: the edge cases nobody wrote down, the boundary an autonomous agent isn't allowed to cross without a human, the stakeholder negotiation that decides which trade-off the business actually wants. (If that boundary needs to be enforced in the pipeline itself rather than trusted to a human reviewer, that's a gate you build, not a habit you hope for.) If you can't sort your own work into those two columns with confidence, that's the actual gap to close before any AI tool changes anything for you.

Common questions

Will AI replace enterprise architects? No credible evidence points that way. Forrester's own EA-tooling research states plainly that agentic AI "won't replace architects anytime soon" while changing how the work is done, and the job-market data shows the EA function holding steady, not shrinking.

What tasks will AI automate first for enterprise architects? Coordination and generation work with a clear right answer: querying system dependencies in natural language, drafting infrastructure-as-code for straightforward stacks, generating a diagram from existing infrastructure, producing a first-pass architecture decision record.

What can AI not do that enterprise architects still must? Handle the edge cases nobody documented, decide where an autonomous agent is and isn't allowed to act, and negotiate the trade-off a stakeholder actually wants. The IaC-Eval benchmark shows the same model that's excellent at ordinary code drops sharply on infrastructure tasks requiring architectural judgment rather than syntax.

How is agentic AI specifically changing the role, versus earlier generative AI? Earlier AI drafted content for a human to review. Agentic AI acts — which is why Forrester names "agentic governance champion," someone accountable for what autonomous agents are allowed to do, as one of the roles the job is shifting toward.

Sources & further reading

Continue reading

Get the next article in your inbox.

One deep-dive per week. Free. No pitch. Unsubscribe anytime.

Subscribe to the Delfen Briefing →
← Back to all articles