The AI-native EA playbook for the mid-market

01Reality

You are the integration

Run a small test tomorrow morning. Ask your organisation which systems you have, what they cost, who owns them, and which ones support your most important customer process. In most mid-market companies the answer does not live in a system. It lives in a person.

That person assembles the answer from the finance system, the SSO overview, three spreadsheets, an old diagram and their own memory. They are the integration between the company's tools, and every time leadership has a question, the company pays that person to be human middleware.

It gets worse for two reasons. The portfolio keeps growing as each department buys its own software, and a meaningful share of licenses is never really used. And the agents have moved in: Gartner expects 40 percent of enterprise applications to carry task-specific AI agents by the end of 2026, up from under 5 percent in 2025. Agents read, write and act in your systems. If no single place knows which agents exist and what they may do, you have a new kind of shadow IT, now with hands.

The problem is not that your people are too slow to maintain the overview. The problem is that anyone has to do it by hand at all.

02The loop

The loop that makes it all hang together

AI-native architecture is not an EA tool with a chatbot bolted on. It is three moves that reinforce each other. Consolidate: strategy, capabilities, systems, integrations, data, processes, risks and AI use-cases become typed entities with relationships, not slides and spreadsheets. When 'system supports capability' and 'agent has access to data' are relationships in a graph, you can ask across them: what breaks if we retire this system, which capabilities have no digital support, where do three tools overlap on the same job.

Connect: data flows in from the sources where decisions are actually made, and the model serves context back out through the Model Context Protocol, the standard that has passed 97 million installations and already runs in production at 78 percent of enterprise AI teams. Your own assistants ask the model for the organisation's truth instead of guessing.

Agents: AI writes the drafts nobody wants to, enriching the system map, spotting duplicates, flagging stale records, proposing classifications. People approve. The moves compound: a better model gives agents better context, agents keep the model fresh, and a fresh model makes every new connector and every new question worth more. It is compound interest for organisational knowledge.

In Atlas the loop is not a diagram in a whitepaper. It is the product's architecture: a typed graph underneath, connectors and MCP in both directions, agents with Verify-review as the default.

03Start scrappy

Start scrappy. Five fields are enough.

Classic EA often fails on ambition: six months of metamodel design before the first insight. The playbook answer is the opposite. Start messy, but start in something that can grow. Five fields per system are enough to begin: name, owner, what it is used for, criticality, and which capability it supports.

That is a manageable import from the spreadsheet you already have, and it is enough for the agents to start working: finding duplicates, proposing relationships, flagging gaps. The rest of the fields do not come from a project. They come from operation: every time an agent enriches a record and a person confirms, the model gets a little deeper. After three months you have the metamodel you would otherwise have designed in half a year, with the difference that it reflects reality.

Never model in advance what an agent can infer and a person can confirm later.

04Agents

The anatomy of an agent, and why review is not optional

An agent in production is three things. Instructions: what it should do and what it must not, written so a person can read and correct it. Connections: which sources and modules it can reach, an explicit list, not 'everything'. Triggers: when it runs, on a schedule, on an event, or on request.

In Atlas that is the anatomy of Architecture Agents: named AI teammates such as Portfolio Librarian, Freshness Warden, Duplicate Hunter, Compliance Scout and Review Clerk. You can see what each agent did, when, and on what basis. And here is the most important design choice in the whole playbook: agents never write directly into the model. They write proposals that land in the Verify inbox, where a person confirms, corrects or rejects.

That sounds like a modest UI choice. It is not. It is the difference between a system you can hold to account and a system you have to hope about. Rules of thumb for your own agent practice, whatever the tool: one agent, one responsibility; all writes through review until an agent has earned trust on a narrow, measurable area; every agent has a named owner and a purpose in writing; log everything, because the log is your compliance evidence the day someone asks.

05Connect

Connect where the decisions are made

The order of integrations decides whether the model gets used or admired. Do not connect by what is technically easy. Connect by decision volume. Identity and access first, the SSO overview, gives the real system inventory and the first usage signals, including shadow-AI candidates. Then the financial signals, contracts, renewal dates and license counts, make the model interesting to the CFO from week one. 'Three tools on the same capability renew within 60 days' is a sentence that pays for the whole exercise.

Then the work systems, repositories, ticketing and documentation, feed the agents the context that makes their drafts precise. And finally the AI tools themselves through MCP: when your assistants and agents pull organisational context from the model, the model becomes infrastructure instead of reporting. Note the shift in economics: now that SaaS vendors ship their own MCP servers (Forrester expects 30 percent in 2026), you no longer build N integrations. You need one client and a place to gather the context. The question moves from 'what can we integrate' to 'where should the truth live'.

06Governance

Governance from day one, without it feeling like it

The mid-market got a gift that is oddly little known. The Digital Omnibus package, approved by the European Parliament on 16 June 2026, deferred the AI Act's high-risk obligations to 2 December 2027 and widened the relief regime to companies with up to 750 employees or 150 million euro in revenue: simplified documentation, reduced fines, access to regulatory sandboxes.

Translated: you have sixteen months of runway and a regime designed for your size. Panic is no longer a strategy, and neither is delay. The sensible plan is boring and doable. Inventory: register your AI systems and agents in one place, with purpose and owner. Classification: decide per use-case whether it is in scope and in which category, and document the rationale, not just the conclusion. Cadence: set a review rhythm so classifications and ownership do not rot.

In Atlas compliance becomes a by-product of running the portfolio well, not an annual project with external hours.

07Three days

Three days to first insight

Plans that need a quarter never happen in the mid-market. Day one: import the mess. Take the system list as it is, the spreadsheet, the SSO export, the finance list. Five fields per system, gaps allowed. Let the agents run the first enrichment and duplicate scan. By end of day you have a system map that is roughly 80 percent true, which is about 80 points more than yesterday.

Day two: connect the first two sources, SSO and financial signals. The Verify inbox starts filling with proposals: missing owners, likely duplicates, systems without a capability. An hour of coffee and confirmation clicks, and the model has relationships, not just rows. Day three: put the first agent to work and open the boardroom views, Capability Map with system support, Value Streams with support gaps, Portfolio with criticality. Take it to the next leadership meeting and ask one question: which of these gaps hurts most.

The architect now handles in a morning what used to be a quarter-long project. And when the board asked about the AI portfolio, pulling the report took ten minutes, not three weeks.