Perspectives on technology strategy, architecture, and long-term risk.
Do you sometimes wonder why the money you spend on IT feels more like a never-ending and ever-increasing tax on the business, rather than a strategic investment? [1]
Why is so much of the IT budget consumed by keeping the lights on, while comparatively little is directed toward meaningful change? [1]
To understand this, it helps to look at how organisations have approached technology over the past several decades.
Since the 1960s, organisations have invested in technology through a continuous stream of initiatives — each intended to improve efficiency, expand capability, reduce costs, or reach new markets [1].
This includes everything from data exploitation and analytics to automation, customer channels, and security capabilities [1].
Individually, these initiatives make sense. Collectively, they create increasingly complex and interdependent systems.
Over time, complexity accumulates. Systems become harder to change. Dependencies multiply. A larger proportion of the IT budget is required simply to maintain what already exists.
At that point, IT begins to feel less like a strategic investment and more like a necessary cost of doing business — a tax that must be paid.
The issue is not that organisations made poor decisions — it is that they made many reasonable decisions without managing their long-term consequences as a system.
↑ Back to topIf AI can increasingly write the code, test it, deploy it and even run it, what is left for people to do? And who makes sure all that automated activity is actually taking the organisation where it needs to go?
A growing number of organisations are finding the same answer: they don’t need more coders, they need more architects. As AI takes on more of the systems development lifecycle, the disciplines that decide what should be built, how it fits together and how it stays safe become more important, not less.
For decades, much of the cost and effort of change sat in execution: writing, testing, integrating and running software. Architecture was often seen as a step to get through on the way to delivery, and sometimes as something that slowed delivery down.
AI is turning that around. When execution becomes fast and cheap, the cost of a poor decision isn’t a slow project. It’s a fast one heading in the wrong direction. Badly designed systems can now be built, connected and scaled in weeks rather than years.
That raises the stakes for every architectural discipline:
As I described in an earlier article, IT starts to feel like a tax when many reasonable decisions are made without anyone managing their long-term consequences as a system. AI speeds up both the decisions and the consequences. Enterprise architecture is the discipline whose job is to manage the system.
To be blunt, many EA functions are not ready for this.
Too often, Enterprise Architecture has been treated, and has treated itself, as a technical function. It produces standards, reference models and diagrams, reviews designs and approves technology choices. That work still matters. But it happens downstream of the conversations where the future is actually decided.
Enterprise Architects need to move upstream, into the rooms where:
That means understanding the business model as well as the technology stack. It means understanding how change actually happens in organisations: how decisions get made, how sponsors are won and lost, and why sound initiatives still fail.
Staying current with technology remains essential. Arguably it’s harder than ever, given that models and tooling now change weekly. But technical knowledge is the foundation, not the job. The architects who matter most in the next decade will be the ones who can turn business intent into a coherent, safe and adaptable path to get there, and who are in the room when that intent is formed.
As AI takes over more execution, fewer people are involved in day-to-day operations. Work that used to pass through dozens of hands, each applying judgement and context, may now pass through none.
That judgement doesn’t disappear. It has to be designed in.
This is where Enterprise Architecture becomes critical. It is the discipline that keeps autonomous systems aligned with business intent, safe to operate and grounded in how the organisation really works. In practice, that means four shifts in how architecture is done.
When AI agents carry out workflows at machine speed, an architecture review board that meets monthly is no longer a control. It’s a bottleneck that people will work around. If it isn’t worked around, it is reviewing designs that have already changed.
The answer is not to drop governance but to change its form. Architects need to move from approving individual designs to defining the boundaries that all designs, and all agents, must operate within. Increasingly, those boundaries can be expressed as policy-as-code: rules written in a form that machines can read and enforce automatically, at run time, every time.
Guardrails let agents act autonomously inside safe boundaries. Architects stop being gatekeepers and become the people who design the boundaries.
Not every decision needs a human, and not every decision should be left to a machine. Reviewing everything doesn’t scale. Reviewing nothing invites disaster.
Instead, architects need to design tiered intervention models that match the level of oversight to the level of risk:
The design work lies in defining the tiers, the criteria for each, and the route between them. It also lies in reviewing those criteria as agents, models and business conditions change.
In many organisations, autonomous agents will soon outnumber human employees. Yet most identity and access models were designed for people: joiners, movers and leavers, roles, and annual access reviews.
Enterprise Architecture must treat AI agents as first-class identities in their own right. That means:
If you can’t say which agent did something, who owns it and what it was allowed to do, you don’t have control of your estate.
Every architectural decision about AI is also a decision about dependency. Which models do your agents rely on? Whose infrastructure do they run on? Whose identity services authenticate them? Which jurisdiction’s laws apply to the data they process?
Left to default, the answers will mostly point to a small number of American providers. That may be acceptable, but it should be a deliberate choice, made with an understanding of the consequences, not something the organisation drifts into one proof of concept at a time.
Enterprise Architects are best placed to make sovereignty a design principle rather than an afterthought. That means keeping a layer between business logic and any single model or platform, designing for portability, and making sure the organisation keeps the ability to change course if a provider’s terms, pricing, behaviour or legal position changes.
If your Enterprise Architecture function is mainly a design review service, it is at risk of becoming irrelevant, or worse, a bottleneck that AI-enabled teams simply route around.
Organisations that get this right will:
I have worked as an Enterprise Architect for more than 30 years. I have helped several large financial services organisations build and improve their EA capabilities, using TOGAF, Zachman, or my own framework, QBAM. I can help you:
The future of Enterprise Architecture is not about drawing better diagrams of technology. It is about making sure that as more of the organisation runs itself, it still runs in the right direction.
If you are rethinking the role of architecture in your organisation, I’d welcome a conversation.
↑ Back to topDo AI agents represent an opportunity to streamline your business processes, cut cost and complexity, and reduce risk? Or will they add a new layer of dependency, much of it on a handful of hyperscalers, that you will be paying for long after the excitement has faded?
For most organisations, the honest answer is that it could go either way. Which way it goes depends on decisions being made now, often quickly and by default, using governance and decision-making frameworks designed for a slower, less chaotic technological and geopolitical world.
Two things set agentic AI apart from earlier waves of technology.
Earlier waves of AI mostly advised. They summarised, predicted and recommended, and a person decided what to do.
Agentic AI acts. Agents can call systems, move data, raise orders, change settings and start workflows, often without a person checking each step. That is where the value comes from. It is also where the risk comes from.
New and more powerful models are now released almost every week. Alongside them, AI providers and third parties are releasing new harnesses at the same pace. Harnesses are the frameworks, tools and orchestration layers that let models plan, use tools and work together.
Some of these advances are significant. Agent swarming, for example, lets many agents split up a task, work on it in parallel and coordinate with each other. They can deliver results far faster than a single agent, or a single person, could. The gains in speed and productivity are real.
So are the risks. When many agents act at once:
The pace is a risk in its own right. The model you tested last month may already be superseded, withdrawn or behaving differently. The harness your teams adopted this quarter may be overtaken by a new one next quarter. Most governance, architecture review and procurement processes run on quarterly or annual cycles. A technology that changes weekly can move on several times before a single decision is made.
Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value and weak risk controls. It also expects that by 2029, at least 70% of organisations running agentic AI in production infrastructure will have a significant service, security or cost incident linked to insufficient runtime controls.
None of this means agentic AI is a mistake. It means the usual pattern is repeating, only faster.
In an earlier article I described how decades of sensible individual technology decisions have added up to estates that are complex, tightly interlinked and expensive to keep running and to change. That is how IT ends up feeling more like a tax than a strategic investment.
Agentic AI risks squeezing that cycle from decades into a few years, or even months. Every team trying out its own agents, with its own tools, models, harnesses and access rights, is adding to tomorrow’s complexity. With new generations of tools arriving every few weeks, estates will soon contain agents built on several of them at once, each with its own quirks, dependencies and vulnerabilities. The difference is that this complexity can act on its own.
Every agent needs an identity, and every identity needs permissions. Gartner reports that 94% of organisations are seeing more machine identities because of AI agent deployments. Swarming multiplies this: one task can now involve dozens of agents, each with its own identity and access. Agents nobody manages become hidden attack surfaces, open to prompt injection, hijacked automation and insider-style data leaks.
Treating every agent the same will fail. Gartner predicts that by 2027, 40% of enterprises will demote or decommission autonomous agents because they did not separate an agent’s ability to act from its scope of access. An agent that drafts emails and an agent that can move money need very different controls.
Approving an agent once is no longer enough, either. If the model or harness underneath an agent changes, its behaviour can change too, even when nobody on your side has touched it.
Up to $234 billion of enterprise application software spending is at risk from what Gartner calls “agentic arbitrage” by 2030, as agents bypass traditional user interfaces and undermine per-seat licensing. That is a real opportunity to rethink what you pay for. It is also a warning that your suppliers’ pricing models are about to change, and they will look for new ways to charge you.
This is the risk that gets the least attention. Most of today’s leading AI models, platforms and agent frameworks come from a small number of American companies. If your agents rely on one provider’s model, tools, identity services and hosting, you are not just buying a capability. You are giving that provider control over how your business operates.
That raises practical questions:
Digital sovereignty doesn’t mean avoiding American technology. It means making these choices deliberately, with your eyes open, and keeping the ability to change.
The organisations that get value from agentic AI, and avoid ending up in the cancellation statistics, will:
I have spent more than 30 years helping organisations adopt new technology without losing control of their estate, their costs or their options. For agentic AI, that means:
Agentic AI may well be the most significant technology change in a generation. The aim is to make sure it becomes a strategic investment, not a new and more autonomous tax, nor a dangerous and uncontrolled risk domain, and that you keep the freedom to choose where your business depends on it.
If you are planning, piloting or struggling with agentic AI, I’d welcome a conversation.
↑ Back to top