Agentic transformation: the EA and planning tooling landscape as of 2026

In 2026 AI agents reached the work of running an enterprise's change: architecture, transformation portfolios, planning cycles and the decisions that move people. They arrive from three directions: the EA or planning tool's own AI, general assistants connected to those tools through MCP, and agent harnesses that sit above them. Each answers a different question well, and the feature lists that compare them rarely show where each one breaks. Agentic transformation means choosing, on purpose, which layer does the work and how its answers get checked.

Three layers, one estate

The tool's own AI. The major EA tools now ship agents, assistants and visual workspaces inside their own repositories. They work well on that vendor's data model and stop at its edge. An estate that spans a second repository, a service-management CMDB and a delivery tracker gets answers from one side only.

A general assistant with the tool's MCP server. Any MCP-capable assistant can now read an EA repository, and many teams have set one up. It handles one question about one record in one turn. It struggles when the question spans systems, the estate is large, or the answer must finish a long run and cite its figures.

An agent harness above the tools. This layer reads several systems in bulk, joins them, keeps long runs on track, and ends in an artifact a forum can decide from. It adds one more thing to adopt, and it depends on each system's MCP server or export. It is the layer that most needs independent checking.

What the feature lists leave out

Analysts now score AI as a product feature, and vendors answer with feature pages. A feature tells you what a product has. It doesn't tell you whether the answers are right on your estate. Three questions close that gap:

• Does it find the same application under three names? Estates carry acronyms, environment suffixes and old names. A tool that matches only on exact names will report false gaps with confidence.
• Does it show the record behind each figure? An answer that goes to a steering committee needs an audit trail, and a refusal when the evidence isn't there.
• Does it finish? A question that needs a thousand records and a dozen joins can run out of steps, pages or context. A clean partial answer is better than a complete-looking wrong one.

Vendor ROI figures describe the vendor's own customers and estates. They are a reason to test, not a result to adopt.

Beyond the architecture team

The same shift reaches the work around architecture. A transformation office asks which initiatives fund changes to applications marked for retirement. A planning cycle asks what a 2027 roadmap costs against the capacity that exists. A change lead asks which teams and systems a reorganisation touches. Each question needs the same thing EA needs: current records, the links between them, and an answer that shows its source.

Agents make those questions cheap to ask, and cheap to get wrong. Three things matter more as they do:

• Planning on current facts. A roadmap built from last quarter's portfolio inherits its gaps. Agents can refresh the inputs, but someone must own the cut-off date.
• Change that people can follow. An agent that reroutes work or changes an owner changes how people work. Announce the change, name who approved it, and keep the old state recoverable.
• Skills that move with the tools. Architects and planners who check an agent's output learn faster than those who only prompt it. Budget for that learning as part of the programme, not after it.

What agentic EA changes in practice

Agents are applications with a model inside. They have owners, data access, capabilities they serve and lifecycles, so they belong in the architecture next to everything else. That makes three things routine work:

1. An inventory read from each agent platform, linked to the application, owner and data each agent touches.
2. Sweeps for the gaps: agents touching personal data that no record shows, duplicates across platforms, and high-risk use without human oversight.
3. Answers that cite their sources, so a decision forum can check a figure against the record it came from.

Questions to put to any vendor or team

• Which systems does it read, and through which interface: a live server, an export, or a copy?
• Can it name the record behind each figure, and what happens when it can't?
• What does a run cost, and who pays the model bill?
• Has it been run on an estate with planted answers, and were its losses published?
• How does it handle a question that needs more records than one request returns?
• Where do agent platforms appear in the repository, and who owns each one?

Measure the layer, not only the model

The choice between layers is easier with evidence than with feature lists. EA-Bench asks AI systems 17 questions EA teams get, on a generated 1,200-application estate with the answers planted, and publishes the questions and the scoring rules so anyone can rerun them. The question kit turns the same questions into a test you can run on your own estate.

The teams that do well with agents in 2026 will not be the ones with the most AI features. They will be the ones that decided where the work runs, kept their systems of record clean enough to be read at scale, and asked for proof before an answer reached a decision forum.