You Are Not Deploying Agents. You Are Redesigning Work

The fastest route to scalable autonomy is designing for humans.

Matthew Skelton recently put it well: the firms that have learned to design for humans are the ones best placed to unlock agentic AI.

Right now, a lot of firms are stuck in a strange kind of half-knowledge.

They know something significant is happening. They can feel the pressure. They can see scattered wins inside their own walls. But they cannot see the terrain. They are spending in hope as much as in certainty. They do not know what peers are actually doing, what is working in production, what is still demo-grade, what has already been pulled back, where the real cost is landing, or when any of it becomes durable ROI.

I do not blame them for this. The information environment is asymmetric. Most of what circulates is marketing, not operating evidence. Nobody publishes the reversal, the throttles, the exception load, or the internal redesign work required to make early wins hold. So leaders make decisions under uncertainty and scrutiny at the same time. In that kind of fog, organisations reach for the moves they know.

They start with technology. That makes sense, because this is a technology.

But that is the mistake.

The right starting point is the organisation the technology is about to enter.

Because the moment AI touches live work, it stops being “just software.” It becomes behaviour on the organisation’s behalf, and that creates obligations most leadership teams have not priced. Operations become your problem. Reputation becomes your problem. Legal and regulatory reality becomes your problem.

So they buy the tools, encourage experimentation, ask for a roadmap, build a centre of excellence, measure adoption, and hire someone senior to “own AI.”

And then the same questions keep returning, in different words, in different rooms.

What are we actually deploying here? Are we seeing what we expected?

Because “agent deployment” is a misleading phrase. It makes the work sound like an installation, as if the main risk is technical and the main job is rollout discipline.

In practice, what is being introduced is not software.

It is delegated behaviour.

Once you delegate behaviour, you are no longer improving a tool. You are changing how work moves, how decisions get made, where judgement lives, how exceptions are handled, and who absorbs the consequences when the workflow is wrong.

This is why so many teams feel busier, not lighter, after the first wave. Not because the agent failed. Because the work was never redesigned to carry it.

The pivot most firms are about to make is simple. The question is no longer “can the model do the task.”

The question becomes: can the organisation carry autonomy without turning speed into constant checking.

That failure mode deserves a name, because it hides inside progress.

I call it AI Translation Debt.

AI Translation Debt is the unpriced work required to make fragmented organisations function as one system. It is the meaning that has to be repaired as work crosses boundaries. It is the informal interpretation that keeps handoffs from breaking. It is the judgement people add when the process is underspecified, the data is incomplete, and reality refuses to match the template.

For years, this debt was survivable because humans absorbed it. They carried context. They made calls. They repaired meaning. They knew who to ask. They knew when to stop.

AI does not remove AI Translation Debt.

It collects it, amplifies it, and presents it back to you as “review.”

It also changes the shape of work in ways that feel productive at first and expensive later. Even when AI use is not mandated, people tend to work faster, widen the scope of what they take on, and extend work into more hours because starting is suddenly frictionless. The organisation looks more capable, and simultaneously becomes less certain about what it can trust.

That is why autonomy so often produces strain before it produces relief. Local throughput rises. Downstream confidence does not. Work reaches the next boundary faster than the next team can safely interpret. Exceptions stack up. Oversight becomes reactive. Senior staff become the checksum for a workflow that is moving too quickly to be socially held together.

It looks like momentum.

It feels like an organisation paying interest.

This is where you get a more useful definition of redesigning work. Redesign is not just process mapping. It is the act of making meaning legible at the seams: where work crosses teams, systems, roles, approvals, vendors. It is deciding what must be true before the workflow continues, where judgement is mandatory, and what happens when reality does not match the template.

The organisation has always had seams like this. It survived by tolerating vagueness and relying on experienced people to repair meaning as work moved.

Agents do not remove that dependence.

They remove the time you used to have to hide it.

And this is where the Team Topologies bridge becomes genuinely useful, in a new way. Not as something for software teams, but as a lens on whether your organisation can support bounded autonomy at all. Team Topologies is a commitment to survivable autonomy: clear ownership, stable interfaces, deliberate interaction modes, and fewer handoffs left to goodwill. Conway’s Law still holds. Whatever your organisation cannot coordinate cleanly, your agents will not coordinate cleanly either.

If you have never designed the organisation so humans can steward work end to end with real clarity and autonomy, it is fantasy to believe autonomy in software will behave better than autonomy in people.

It will behave exactly like the organisation.

Just faster.

So there is an almost paradoxical truth here: designing for humans is now the fastest route to scalable autonomy in machines. Not as a moral stance. As an operating stance. The organisations that get value first will not necessarily be the ones with the most ambitious agent demos. They will be the ones with the cleanest boundaries, the clearest ownership, the least Translation Debt, and the highest ability to intervene without politics.

This is what makes the private equity angle sharper too.

Portfolio AI is not primarily a model problem. It is a repeatability problem. It is the difference between scattered capability and a transferable operating pattern.

Translation Debt is the portfolio killer because it scales invisibly. Each company looks busy with AI. Each company generates local wins. But the portfolio accumulates new checking load, new exception pathways, new reconciliation work, new manager time, and new operational friction that never appears in the business case.

You do not get leverage.

You get a distributed tax.

And most portfolios are currently structured in a way that makes this predictable: one group funds capability, another group absorbs exceptions, another group explains the numbers, and nobody has authority over the redesign that would make the system cohere.

That is not transformation.

That is cost relocation with better branding.

If you want this to work at portfolio scale, you do not need another AI owner in the abstract. You need an owner of redesign. Someone with the standing to force boundary clarity, collapse unowned handoffs, decide where judgement must remain human, and make the gain show up in sponsor-legible outcomes.

This is not restraint. It is speed. It is the only way to avoid paying for ambiguity later, in management attention, Translation Debt, escalation, and reversals.

So what does redesign look like when you are serious?

Start with one workflow where delegation could plausibly change unit economics. Not an isolated task. A workflow with handoffs, decisions, exceptions, and measurable outcomes.

Then run a redesign sequence that forces the same questions every time.

Begin with the seams. Where does work cross teams, systems, roles, vendors, approvals? What must be true at each seam for the next step to be safe? What is currently being supplied by human memory, judgement, or goodwill?

Then name the translation work. Who is reconciling conflicting signals? Who is adding missing context? Who is deciding what the process really means when reality does not fit the template? If you do not name this work, you will not price it. If you do not price it, you will pretend it does not exist.

It does, and it will show itself later with force.

Then reduce the handoffs you can reduce. Every transfer point is a place where meaning can blur, responsibility can thin out, and machine speed turns into human checking. This is where Translation Debt either gets paid down or is allowed to compound.

The redesign job is not just to make the agent perform.

It is to stop the organisation converting machine speed into downstream checking.

Then locate decision rights and judgement. Who can pause? Who can override? Who can stop? Where is human judgement mandatory? What conditions must be true before the workflow is allowed to continue? Delegation without an owner is not autonomy. It is unmanaged risk.

Then design the exception path as part of the workflow, not as an afterthought. Who intervenes when the system is uncertain, contradictory, out of scope, or wrong? How quickly? With what authority? What returns the workflow to a known safe state?

Weak business cases model the median path and leave the tail to absorb itself. In real organisations the tail becomes supervision, escalation, recovery, stitching, quality reversals, management attention, forecast distortion, margin leakage.

The gain does not vanish. It reappears elsewhere in the system.

Do this repeatedly and something changes. AI stops being a sequence of bespoke experiments and starts becoming organisational muscle. Boundaries become explicit. Judgement becomes visible. Exception handling becomes designed rather than improvised. Reuse becomes possible because the operating conditions stop being ad hoc.

That is what scale actually requires.

This is also the more hopeful story, if you let it be one. If you redesign work properly, autonomy does not create chaos. It creates relief. It reduces Translation Debt. It reduces managerial stitching. It allows judgement to be used where it matters, instead of being consumed by constant reconciliation.

And it becomes possible to approve more delegation, not because risk disappeared, but because the organisation has made responsibility legible.

That is the real upgrade.

Not the agent.

The work.

And if you are not willing to redesign the work, what exactly do you think you are deploying?

--

Stuart Winter-Tear

Independent advisor | Author of UNHYPED

Advisory: Executive Calibration