All services Every engagement has a fixed, published price. Book a strategy call
AI is collapsing the cost of building systems while doing nothing to the cost of coordinating them. What that means for the future of enterprise architecture: the work that gets automated, the decisions that remain, AI agents as new citizens of the estate, and a 12 to 24 month plan.
Most writing about the future of enterprise architecture in the AI era says the same three things: AI will change everything, architects must become more strategic, and here is a framework. This article makes a different argument, from a firm that runs architecture functions for a living: AI makes enterprise architecture more necessary rather than less, and it changes which parts of the job are worth paying for. The reasoning, the evidence and the plan are below.
The short version. AI is collapsing the cost of building systems while doing nothing to reduce the cost of coordinating them. When anyone can generate an application in an afternoon, estates sprawl faster, and the scarce skill stops being production and becomes judgement: deciding what should exist, what connects to what, and what gets retired. Documentation, discovery and impact analysis become substantially automated. Decision rights, accountability and governance of AI agents become the core of the job. Enterprise architects who define themselves by diagrams are in trouble. Those who own decisions are about to become more valuable than they have ever been.
Every era of computing has lowered the cost of building and raised the cost of coordinating. Cloud made infrastructure cheap and estates sprawled. SaaS made applications cheap and portfolios doubled. AI is the same move at a larger scale: code generation, low-code agents and AI features inside every product are collapsing the marginal cost of creating new systems towards zero.
What AI does not collapse is the cost of coordination. A hundred generated applications still need to agree on what a customer is. Data still has to live somewhere, lawfully. Integrations still multiply as the square of the systems they connect. Security still depends on knowing what exists. The economics are blunt: when production gets cheap, the bottleneck moves to judgement, and judgement about technology estates has a name. It is enterprise architecture.
That is why the future of enterprise architecture is not obsolescence, and it is not business as usual either. The parts of the job that were about producing artefacts are being automated. The parts that were about making decisions are becoming the whole job. Five forces drive this, and they are all already visible: generation everywhere, agents entering the estate, data gravity meeting regulation, AI-accelerated sprawl, and boards demanding AI results with governance to match.
It helps to be precise, because AI and enterprise architecture meet at three distinct levels, and each changes different work.
Level one: AI as the architect’s tooling. Current-state discovery, documentation and first-draft analysis, the work that consumed most of a traditional EA team’s hours, is increasingly done by machines reading the estate directly. Level two: AI systems as architecture subjects. Models, AI platforms and AI features inside products are new components of the estate that need placing, connecting, securing and paying for, like everything else only faster-moving. Level three: AI agents as actors. Software that takes actions on its own initiative inside your systems, which is a different kind of citizen altogether, and gets its own section below.
| EA task | How it works today | Where AI takes it | What stays human |
|---|---|---|---|
| Current-state discovery | Interviews, spreadsheets, months of effort, out of date on publication | Continuous, machine-read from the estate itself: networks, identity, cloud bills, code | Deciding what the picture means and what to do about it |
| Documentation and models | Hand-drawn diagrams and repositories that decay | Generated and regenerated on demand from live sources | Choosing which views matter and which decisions they serve |
| Impact analysis | Expert judgement over incomplete maps, slow and cautious | Fast first drafts of what a change touches, with evidence | Owning the risk call, because the model does not carry accountability |
| Standards and pattern checking | Manual review boards catching a fraction of deviations | Automated checks in delivery pipelines, continuous rather than ceremonial | Writing standards worth enforcing, and judging exceptions |
| Roadmapping | Annual documents negotiated in workshops | Living views with options and costs updated as facts change | The trade-offs: what the organisation values and funds |
| Governance | Meetings, templates, gates | Evidence gathering and policy checking automated | Decision rights, exceptions and the authority to say no |
Read the right-hand column top to bottom and the pattern is unmissable: everything that survives is a decision. That is the honest summary of AI-driven enterprise architecture, and it is why the charter-and-decision-rights approach we describe in our enterprise architecture guide for CIOs matters more in the AI era rather than less.
The most consequential change is the least discussed. AI agents, software that pursues goals and takes actions across systems rather than waiting for instructions, are entering ordinary business estates: booking, purchasing, summarising, triaging, writing to systems of record. Architecturally, an agent is a new kind of actor, and AI agents in enterprise architecture raise questions that existing operating models simply do not answer.
Who is this agent, in identity terms, and what may it access? Under what authority does it act, and within what limits? What data can it see, and can it move data between systems that were deliberately kept apart? What does it log, and who reviews the log? What happens when it misbehaves, and who can turn it off, quickly, everywhere? An estate with fifty agents and no answers to those questions has recreated shadow IT with initiative.
This is where enterprise architecture earns its place in the AI conversation. Someone must decide the platforms agents run on, the identity model they use, the data boundaries they respect and the audit they leave behind, and those are architecture decisions, made once, deliberately, rather than fifty times by accident. Treat agents like a new class of employee: vetted before arrival, given the least access needed, and dismissible in one action. The organisations that write those rules now will adopt agents faster than the ones that will spend 2027 doing forensics.
The role of the enterprise architect in the AI era shifts along one axis: away from producing representations of the estate, towards owning decisions about it. The uncomfortable version: if your value is that you know Visio and remember the integration map, a model now does that. The confident version: if your value is that you can decide, explain and be accountable, you are about to have better information and fewer excuses than any architect in history.
| What shrinks | What grows |
|---|---|
| Drawing and maintaining documentation by hand | Owning the decision list: portfolio, platforms, integration, data and AI placement |
| Manual estate discovery and audits | Designing the guardrails that let generated systems and agents move fast safely |
| Review board ceremony and gatekeeping | Judging exceptions, risks and trade-offs, with the authority to say no and the habit of saying yes-with-conditions |
| Framework administration for its own sake | Translating between board intent, regulation and technical reality, in both directions |
| Being the person who knows where things are | Being the person who decides what should exist, and what gets retired |
The seniority profile changes with the work. Teams built on juniors producing artefacts have a problem, because the artefacts are being automated. Small teams of senior people who own decisions, backed by AI tooling for the mechanical work, are the shape that wins, which incidentally is the same conclusion the economics of EA teams pointed to before AI arrived.
The TOGAF and AI question comes up in every conversation about the future of enterprise architecture, and the honest answer is: the vocabulary survives, the ceremony does not. TOGAF’s lasting contribution is a shared language: capabilities, building blocks, target and transition states, the distinction between business, data, application and technology architecture. That language is how architects, vendors and auditors understand each other, and AI does not change it. What AI hollows out is the liturgy: the phase-by-phase document production that consumed architect-years, because the documents are exactly what machines now draft. Treat the ADM as a checklist of questions rather than a sequence of deliverables, keep the vocabulary, and spend the recovered time on decisions. Our guide to enterprise architecture frameworks covers this in full.
A future-of piece owes you the other side of the ledger, so here is what AI does not change. Decision rights: someone accountable must still decide what the organisation builds on, and a model cannot carry that accountability. Data gravity and law: data is physical and regulated, it lives in jurisdictions, and no amount of generation changes where it may lawfully sit. Vendor economics: lock-in, licensing and exit costs behave exactly as before, except signed faster. Integration arithmetic: connections still grow faster than systems. And organisational politics: the hardest architecture problems were never technical, and the AI era does not make the sales director want to share their data any more than the cloud era did. Anyone selling you a future where these disappear is selling something other than architecture.
Preparing an EA function for the AI era is not a transformation programme. It is five capabilities, built in roughly this order, each useful on its own.
| Capability | What it is | The first step |
|---|---|---|
| 1. Machine-readable estate | An application, data and integration inventory that lives in systems rather than slides, so both humans and AI tools can reason over it | One inventory with owners and costs, however rough, refreshed automatically from at least one live source |
| 2. AI placement decisions | Named, approved platforms for AI workloads, and rules for where each class of data may meet a model | A one page decision: which AI platforms are approved, which data classes they may touch, who signs exceptions |
| 3. Agent governance | Identity, least privilege, audit and a kill switch for non-human actors, before they multiply | An agent register, even if it has three entries, and an identity pattern agreed with security |
| 4. Guardrails in the pipeline | Architecture standards enforced automatically where systems are built and bought rather than in a monthly meeting | Pick the three standards that matter most and automate checking for just those |
| 5. Decision velocity | A forum and decision record habit fast enough that AI-speed delivery does not route around it | The weekly forty five minute design forum and one page decision records, exactly as they worked before AI, now non-negotiable |
Sequenced sensibly, the move to an AI-era EA function fits in two years without heroics, and pays for itself as it goes.
| Phase | Focus | What exists at the end |
|---|---|---|
| Months 1 to 3 | See the estate and take the two urgent decisions: approved AI platforms, and the agent identity pattern | A live inventory, an AI placement decision, an agent register |
| Months 4 to 9 | Wire guardrails into delivery and procurement. Retire the first duplicated systems the inventory exposed | Automated checks on the standards that matter, first retirement savings banked |
| Months 10 to 18 | Shift the team: automate documentation and discovery, move architect time onto decisions and agent governance | An EA function whose hours go to judgement, with tooling doing the mechanical work |
| Months 18 to 24 | Institutionalise: charter refreshed for AI-era decisions, metrics reported in money, agents governed as routine | An operating model that treats AI systems and agents as normal, governed citizens of the estate |
Predictions age badly, so treat this as a direction rather than a promise. The EA function that thrives looks like this: small and senior, with architects who own explicit decision rights written into a charter. A live, machine-read model of the estate that nobody draws by hand. Standards enforced in pipelines, exceptions decided in days, and every significant decision recorded on one page. A register of AI systems and agents governed with the same normality as user accounts are today. And a quarterly report to the board written in money: systems retired, duplication removed, change shipped without rework, incidents avoided on known risks. Nothing in that picture is exotic. Every element exists somewhere today. The future is unevenly distributed, and the distribution favours organisations that start early.
We build and run architecture functions, at published fixed prices, and the AI era work maps onto services that already exist. EA for Data and AI Platforms makes the placement decisions: platforms, data boundaries and the architecture AI runs on. The AI Governance QuickScan and AI governance services cover the control side, including agents. The EA operating model service installs the charter, forum and decision cadence, and the Virtual EA Department runs the function as a service while you build your own. Every price is on the page, which tells you this article was written to be useful rather than to open a sales conversation.
AI is replacing the artefact production side of the job: documentation, discovery and first-draft analysis. It is not replacing the decision side, because decisions about what an organisation builds, connects and retires carry accountability that a model cannot hold. Architects who own decisions become more valuable. Architects who only draw diagrams do not.
A smaller, more senior discipline focused on decision rights: which systems should exist, where data and AI workloads run, how agents are governed, and what gets retired. The mechanical work of mapping and documenting the estate becomes automated and continuous.
Discovery and documentation shift from manual effort to reviewing machine-generated views. Time moves to decisions, exceptions, agent governance and translating between board intent and technical reality. Standards checking moves into delivery pipelines and runs continuously.
Software actors that pursue goals and take actions across systems on their own initiative. Architecturally they need identity, least-privilege access, data boundaries, audit trails and a kill switch, decided once as patterns rather than improvised per agent.
The vocabulary is still relevant: capabilities, building blocks and the business, data, application and technology layers remain the shared language of the discipline. The heavyweight phase-by-phase document production is what AI makes obsolete, so treat the ADM as a checklist of questions rather than a sequence of deliverables.
It makes them easier to see and faster to grow. Discovery and documentation improve dramatically, but the ease of generating new systems accelerates sprawl, which makes deliberate portfolio decisions more important rather than less.
Three moves: get the estate inventory live and machine-readable, decide which AI platforms are approved and which data classes they may touch, and start an agent register with an identity pattern agreed with security. Everything else builds on those.
Like a new class of employee: vetted before arrival, given the least access needed to do the job, individually identifiable, fully audited, and dismissible in one action. The rules should be architecture patterns decided centrally rather than per-team improvisation.
Written by Emmanuel Olatunji, Chief Enterprise Architecture Partner at Intelance. TOGAF 9 and 10 certified. Last reviewed 8 July 2026.
Emmanuel Olatunji
Chief Enterprise Architecture Partner
Emmanuel founded Intelance in 2017 and has led architecture, AI governance, cyber and deal engagements across the UK, Europe, the Middle East and Africa. Every Intelance engagement runs under his direction, and every guide he publishes comes from that work.
See his record on LinkedIn About IntelanceThirty minutes with the architect who wrote this guide. A clear recommendation and a fixed fee in writing, within one UK business day.
Talk to a senior adviser
Enterprise Architecture
