If you can’t trust your people, you won’t trust your agents – RTO mandates mean you’re not ready for AI

Over the last couple of years I have noticed a weird paradox.

On the one hand many enterprises are desperately pursuing the implementation of AI — and in particular dream of agents that are given broad scope to transform productivity. They yearn to give AI agency.

On other hand many of those same enterprises are forcing their employees to return to the office using the excuse of productivity… despite some large‑scale studies that suggest otherwise. They yearn to strip humans of agency.

Is it only me that sees that as weird?

No company is suggesting that they are going to set up desks and put embodied agents onto them — like a crazy enterprise version of the current fad for running OpenClaw agents on Mac Minis.

And so by definition agents will run anywhere and everywhere. They will run in the cloud. They will run on customer devices. They will run on partner infrastructures. By definition they will be distributed. And if enterprises are happy for agents to work anywhere, why do they insist that humans must work somewhere?

Digital transformation?

But this paradox masks something much deeper — the last rearguard action of industrial-age enterprises who were built for an era of top-down control. Enterprises who will never get the full benefits of agents because they cannot abide the idea of agency. Enterprises whose whole perspective on how to embed agents rests on a fallacy about what agency means.

Because when previous waves of digital change collapsed coordination costs, they also created profound changes in the structure of work and the potential distribution of agency.

Despite being co-opted by dubious product framings and consultant gobbledygook, ‘digital transformation’ was never really a technology event. Instead it was a recognition that a new environment had emerged for organizing and connecting information across boundaries.

And so digital-native companies — along with a small number of industrial-age visionaries — built for this new environment, scaling to previously inconceivable complexity by architecting around small, decentralized, self-managing teams. Amazon’s two pizza teams. Haier’s RenDanHeYi structures. Spotify’s squads and tribes. Organizations that embraced digital as a path to new decentralized management models.

And in doing so they distributed agency, information, and judgement into small, manageable chunks — irrespective of whether people were co-located or remote.

Many industrial-age companies did the opposite, however. Rather than see digital as a new operating environment they saw it as a new set of technologies.

So they didn’t invent, they intensified.

They doubled down on control. They automated existing processes. They layered digital channels over their hierarchical operating models to put lipstick on the pig.

But they didn’t redesign for decentralization, for agency, and so co-location frequently became a crutch — a way of compensating for an inability to tolerate distributed judgement.

And they survived. But they survived by protecting hierarchy and control. By making a rational decision to protect present practice at the expense of future possibility.

Operating model debt

The trade-off wasn’t just more technical debt. It was operating model debt. A further calcification of existing practices driven by a structural reluctance to redesign how authority, judgement and coordination were shared.

Which brings us to the part nobody wants to say out loud.

Agents don’t enter enterprises in a vacuum. They are not neutral tools that can work with whatever already exists. They are tools of delegated judgement. Of distributed work. Of bounded agency.

Kind of like self-managing teams perhaps.

And this matters, because the structural advantages enjoyed by decentralized companies are likely to compound when it comes to implementing agents.

Because AI is not a separate trend to digital — it is its next phase. Where digital lowered the transaction costs of information exchange, agents promise to use this environment to collapse the coordination costs of cognition.

But to do this safely, effectively, productively, they need to be granted agency.

Decentralized companies are already built this way. Their teams are small, cohesive, and agile — with bounded context, clear information, and manageable decisions that can be supported by agents.

But industrial-age companies are not.

Instead they have operating model debt — hierarchy, layers of management, and controls meant to limit agency to limit variance.

Which makes them structurally incompatible with agents because they have not re-engineered their management model for distributed agency.

Which is why — for many of these organizations — the default lens for agents is frequently automation, not agency. Of eliminating people to reduce cost. Of deterministic workflows. Of eliminating rather than enabling judgement. Of detailed operations rather than delegated outcomes.

Of demanding employees sit where they can see them.

Imbecile automatons

While executives might justify RTO mandates with talk of mentoring, culture, or the benefits of creativity that come from making people travel hours every week on the off-chance they’ll bump into some rando who gives them the secret of eternal life, none of this really stacks up. Because they are also exhorting them to work with AI models that can provide all of that immediately, continuously, and with a greater scope of knowledge than Bill from accounts has ever bothered to absorb.

And so for industrial-age organizations, RTO mandates become a sign of something more fundamental.

An inability to trust. An inability to distribute. An inability to surrender agency.

Or perhaps more honestly, an insecurity about what happens to authority when you can’t physically stand over someone. But you can’t stand over an agent.

In practice, this incompatibility will not show up as dramatic failure but as neutralization. Agents will be welcomed — but only into tightly defined workflows. They will be required to produce deterministic outputs. They will be wrapped in compliance layers thick enough to smother discretion at birth. And they will be centrally governed, centrally approved, and centrally monitored.

In other words, they won’t be treated as intelligent agents but imbecile automatons.

Leaving these companies still built for a world that digital already disrupted — and which AI is now amplifying.

Because organizations that have already distributed human judgement will be able to distribute machine judgement just as easily. They can drop it into an existing management framework that absorbs it without friction — enabling them to rapidly scale their existing advantages based on focused teams, clear data scope and bounded agency.

But the ones that doubled down on centralized judgement and existing hierarchies during the digital wave won’t be able to absorb agents in this way — and so will reinterpret them into something that fits to their world view.

Compounding their debts and capping their future potential.

Compulsive oversight

So let’s return to the topic of RTO mandates and the seeming paradox encapsulated in the expectation that people should work in an office while agents work in the cloud.

Proximity isn’t really the core issue — because even empowered teams can choose to sit together or sit apart. The core issue is whether colocation is being used to better coordinate distributed judgement — or to compensate for a lack of it.

And the return-to-office mandates being handed down by industrial-age companies are frequently about compensation, not productivity. They are about the compulsive need for oversight. For visible control. For keeping workers under observation.

Which is the same instinct shaping their approach to agents.

In reality they are not relaxed about distributed agents because they like them better than humans. They are relaxed because their operating model requires agents to function as deterministic drones — inherently controllable, auditable and observable — irrespective of location — in a way that people simply aren’t.

In effect they view them as workers with deeper control planes — not bearers of agency.

And so it’s inevitable that organizations that centralize control over their meat will do it to their machines too. They will simply do it with digital walls made of rules and guardrails instead of physical walls made of bricks and plaster.

Which means I was wrong to think of it as a paradox. Or even frame it as RTO. Agents are not getting better conditions than people — their agency is just being limited in a different way.

And that’s the real lesson here. That unreconstituted industrial-age companies will struggle to leverage agents effectively because their operating model doesn’t allow for it. Which means we’re about to see the same pattern play out with AI transformation as we did with digital transformation.

They will announce it.

They will embrace the language.

And then they will engineer it so that nothing fundamental has to change.

And then they’ll wonder why their AI transformation feels so very small.

Plus ça change, plus c’est la même chose.