The problem nobody names: when AI optimises one step and the case still stalls

Ask most operations leaders whether their organisation has adopted AI, and the answer is usually yes. Ask whether a cross-department approval, an exception case, or an audit request has become faster as a result, and the answer is often much less certain.

That gap is not primarily a technology problem. It is a coordination problem.

AI can make one step faster while leaving the overall case stuck between teams, systems, approvals, and exceptions. The tools have improved rapidly. The connections between them have not always kept pace.

After 25 years in process and transformation roles, I have seen this pattern recur across sectors and technology cycles. What has changed is the range of tools available and how easily individual teams can adopt them. That increases the need for orchestration: coordinating the data, people, systems, agents, decisions, and evidence that make up the complete process.

Strip away the industry vocabulary — banking, logistics, manufacturing, healthcare — and almost every operational process reduces to the same underlying shape: data exists somewhere, someone or something needs it to make a decision or take an action, and a rule determines what that decision is allowed to be.

Most “AI in operations” initiatives optimise one piece of that chain. One system gets smarter. One process step gets automated. One department gets a chatbot or a coding assistant. What is usually missing is the coordination layer around it: making sure the right data reaches the right person at the right time, that AI agents and automated steps hand off to each other cleanly, and that the full decision trail is traceable by default rather than bolted on afterwards when a regulator asks for it.

This shows up in very ordinary places, not exotic ones: an approval that needs sign-off from two departments and still happens over email. A supplier request that gets rekeyed manually between three systems because they don't share data. An exception case that a junior staff member escalates by walking over to someone's desk, because there's no structured path for it. An auditor asking, “Who decided this, and on what basis”, with the honest answer requiring someone to reconstruct it from memory and email threads.

None of these are AI problems on their own. But they are exactly where AI initiatives quietly stall.

A question that comes up fairly often is whether we are building an industry-specific product, such as a banking or logistics version. The answer is no. This is a considered position, not a gap, and it is worth explaining why.

The same patterns — approvals that cross departments, exceptions that don't fit a standard workflow, AI agents that need to hand off to each other, decisions that need to be reconstructable months later — recur in essentially the same shape across every sector, in every market that career has touched. Building a separate “industry solution” for each one means re-solving the same underlying problem five times, in five different vocabularies, instead of solving it once and letting each client's own domain expertise sit on top of it.

In practice, this means the client's team remains the domain expert on their process, their regulations, and their exceptions. Our role is to bring the structure that connects their systems and orchestrates their workflows, rather than a prepackaged template that assumes it understands their business better than they do.

A pattern that keeps showing up: an organisation doesn't have an empty AI landscape to design from scratch. It already has a mix — a chatbot here, a coding assistant there, an automation tool somewhere else, often from different vendors, rarely built to talk to each other. Individually useful. Collectively disconnected. Some call this “agent sprawl,” and it tends to arrive well before anyone planned for it.

This is the part that's genuinely different from earlier in that career. Adopting a new system or automation tool used to require budget approval, an IT project, a vendor relationship. Today, almost anyone on a team can spin up an agent or subscribe to a tool on their own. Access has gotten radically easier — which is good news for speed, but it also means the number of moving pieces an organisation has to coordinate has multiplied, usually faster than its governance has caught up.

So, the first move is rarely to bring in something new. It's to map what's already running: which tools, which agents, which automations exist across which teams, and where they currently stop talking to each other. Almost every organisation, when it looks honestly, finds it already owns more of the pieces than it realised — the gap isn't a missing tool, it's the missing layer that connects the ones already there. Only once that's mapped does it make sense to ask whether anything genuinely needs to be added, and that should close a real gap, not just add another island.

By mid-2026, both the US and China have demonstrated that access to leading AI models can be restricted — temporarily and selectively, but with real effect on any organisation that built its agents around one specific model or provider. That's no longer a hypothetical risk to plan around eventually; it's a live design consideration.

Multi-model capability is very difficult to retrofit after the fact. It must be designed in from the start, through abstraction layers that orchestrate agents independently of any single model. As delivery cycles shorten, with agents drafting designs, tests, and documentation in days rather than weeks, the need for human evaluation does not shrink. It grows. The ability to genuinely assess a result, rather than simply approving what an agent proposes, is the real bottleneck now.

For a COO or CTO, the question boils down to this: if access to the AI model or agent platform you rely on changed overnight, could you switch, or would the process stop entirely?

On 5 August 2026, Singapore's Monetary Authority confirmed that its forthcoming Guidelines on AI Risk Management will apply to agentic AI, not only conventional models — the first major financial regulator to state this explicitly. For financial institutions, that confirmation sharpens the points above into supervisory expectations: a comprehensive AI inventory that extends beyond in-house models to the AI embedded in SaaS platforms such as CRMs and personalisation engines; controls placed where actions actually execute, since models and prompts change quickly but execution is where accountability lives; and audit trails generated as a direct by-product of execution, not stitched together after the fact across fragmented systems. What was already good practice is becoming a supervisory expectation.

A first conversation on this topic doesn't need to open with “do you want AI orchestration.” In practice, it opens with recognisable, everyday symptoms — the kind that get raised as complaints rather than technical problems:

• When a request needs sign-off from more than one department, how does that happen today: through a system, or by email and follow-up calls?
• If a regulator, auditor, or customer asked why a specific decision was made six months ago, how long would it take to reconstruct the answer — and would it hold up?
• Have different teams brought in their own AI tools or agents that don't currently talk to each other?
• When something doesn't fit the standard process, what happens to it? Is there a defined path, or does it depend on who happens to pick it up?

If any of these produce a long, frustrated answer rather than a quick confident one, that's usually the real signal — not a request for a product demo.

The organisations that benefit most from AI will not necessarily be those with the most agents or the fastest individual tools. They will be the ones that can coordinate those tools across the full operational case. That means mapping what already exists, designing reliable handoffs, managing exceptions, preserving human accountability, and recording the evidence needed to explain each decision. AI can optimise a step. Durable operational value comes from making the whole process work.

If any of the questions above sound familiar, that's the point where a first conversation is worth having.