On the first day of UNLEASH in Paris, I asked a seven-figure AI orchestration platform how many people worked for the company.

Two minutes and 36 seconds later, it told me I wasn’t authorised to know.

In the intervening 156 seconds, this marvel of enterprise technology had rummaged through Workday, interrogated a knowledge base, consulted assorted other systems and burned through enough tokens to make a small woodland nervous. All to establish that I lacked the necessary permissions to access a number published in the company’s annual report.

I was watching a machine capable of coordinating multiple autonomous agents, navigating complex workflows and integrating with virtually every enterprise system known to humankind. Yet ask it how many people work here (something I could Google in seconds) and it behaves like a particularly officious receptionist at a Soviet-era government ministry. Your request has been considered and, alas, you do not have the appropriate clearance to know how many of you there are.

I laughed (obviously). But the more I thought about it, the less funny it became.

Because the agent hadn’t simply got an answer wrong. It had exposed something fundamental about how we’re designing workplace AI.

It understood the individual. It could identify me, retrieve my records, check my permissions and help me complete a task. It understood processes, too: the workflows, approvals and transactions that keep an organisation functioning.

What it seemingly couldn’t recognise was a fact about the organisation itself.

It had a model of me. It had a model of the machinery. It had no model of us.

After several days wandering around UNLEASH, watching vendors demonstrate increasingly sophisticated ways to optimise individual employees and automate the processes connecting them, I began to wonder whether this wasn’t the defining blind spot of the industry.

We’re building extremely clever systems for managing the parts. But who’s making sure the whole still works?

Two layers, both sensible

Walk the exhibition floor at an HR technology conference in the Year of our Lord 2026 and you’ll find an industry smearing AI all over two parts of organisational life.

The first is the individual. Copilots that draft emails, agents that book leave, skills engines that plot your next career move, platforms that personalise your working experience down to your role, location, circadian rhythm and presumably your Myers-Briggs type, if somebody can find a way to charge for it.

The promise is productivity. Save everyone a few hours a week, multiply by salary and number of seats, and, BINGO, you’ve got an ROI calculation that bangs in PowerPoint. Whether anyone actually gets to go home earlier is another question.

The second is the system. Orchestration, workflow automation, agents handing tasks to other agents, the slow extinction of the ticket queue. Nobody will mourn filling in a form to request permission to fill in another form (except possibly the person who designed the forms, who’s now Head of AI Governance).

The promise here is efficiency: faster resolution, fewer handoffs, lower costs. Ker-ching.

I’m not dismissing either. Some of what I saw in Paris was seriously impressive, and a depressing proportion of working life consists of exactly the administrative nonsense these tools can eliminate.

But both layers have a buyer, a budget and a business case. And somewhere between the productivity gains and the efficiency savings, we’ve mislaid the organisation.

The layer nobody owns

Between the individual and the system sits a third layer: the shared one. It’s what allows thousands of people to behave as an organisation rather than as independent contractors with matching lanyards.

Part of this is factual. How many people work here? What do we actually do? What do we sell, and to whom? How do we do it? Where do I find the current policy? What are we trying to achieve this year? These are questions any reasonably curious new starter might ask in their first week. And questions any AI claiming to understand the organisation ought to be able to answer.

These aren’t facts about me, or instructions for completing a transaction. They’re facts about us.

The other part is identity. What are we trying to achieve together? Why does this organisation exist, beyond providing a convenient administrative arrangement for distributing salaries? Facts establish a common picture of reality. Identity gives people a reason to care about it.

Consider my unfortunate orchestration agent. It had access to an impressive range of systems, but apparently no understanding of what constituted common organisational knowledge. It treated a published headcount figure as though I’d asked for somebody’s confidential personnel file.

That was a design failure. But it points to a wider organisational problem: in many companies, nobody has been given the job of deciding what the organisation knows about itself, where that knowledge lives and which version is authoritative.

Headcount sits in an HR system behind permissions designed for personal records. Strategy exists as a presentation, a town hall recording and fourteen departmental interpretations, each progressively more convenient for the department concerned. Decisions disappear into email threads. Policies survive in three versions, two technically obsolete but still enjoying a devoted following.

Gallagher’s 2026 Employee Communications Report offers a useful indication of the problem. For five of the seven foundational communication assets it examined, including strategic narratives and employee value propositions, more than half of responding teams lacked an agreed, documented standard. The asset either didn’t exist or lived in someone’s head. Just 15% reported an EVP that was active and socialised.

Having a strategy and having a shared, operational understanding of it are two very different achievements.

And you can’t ground an AI agent in a story nobody has written down.

Nobody buys shared

Why is the shared layer so neglected? Not because anyone is actively opposed to shared understanding. You’d struggle to find a CEO willing to announce that organisational coherence was being discontinued as a cost-saving measure.

It’s because the economics favour everything else.

Software vendors sell seats, transactions and measurable efficiency. The individual productivity story is easy to demonstrate. The process automation story is easy to price. Both produce numbers that can be entered into a spreadsheet and presented to a finance director without requiring an extended philosophical discussion about the nature of collective endeavour.

Shared understanding is harder to package. There’s no obvious licence for everyone knowing the same thing. No straightforward metric for the value of employees understanding why the organisation has changed direction, or recognising that another team’s priorities are connected to their own.

That doesn’t mean these things have no economic value. Their value is distributed across the organisation, while purchasing decisions tend to be made within functions.

There’s a principal–agent problem lurking here: vendors are rewarded for satisfying the people who sign the contracts, whose priorities don’t necessarily match those of the employees using the technology, still less the organisation as a whole.

But workplace AI adds another twist. Even when the buyer is happy and the users are happy, the organisation as a whole may be worse off.

HR buys employee experience. IT buys orchestration. Operations buys efficiency. Communications buys channels. Everyone gets a dashboard showing their investment is working. Nobody gets a dashboard showing whether the organisation still makes sense.

We pour millions into optimising the parts and leave the whole to chance.

And personalisation introduces another problem: it makes organisational inconsistency harder to spot.

Imagine two employees asking their AI agents what the company’s most important priority is. One receives an authoritative answer about reducing costs. The other is told that growth is paramount.

Both answers draw on genuine company documents. Both are plausible. Both are fluent, confident and appropriately personalised.

Neither employee knows the other has been told something different.

In the old world, they might have read the same badly written intranet page, disagreed about what it meant and had a row. That row was useful. It exposed the ambiguity, giving somebody an opportunity to resolve it.

Twenty thousand private conversations with twenty thousand AI agents don’t necessarily produce that moment of discovery.

Every employee gets a perfectly coherent version of reality. Unfortunately, they’re not all living in the same one, and nobody’s checking.

For decades, organisations have tried to break down silos. Now we’re building a bespoke, AI-powered silo around every employee, and congratulating ourselves on the user experience.

The cost of being individually right

As anyone who’s ever had the misfortune to disagree with me will attest, I fucking love being right. But even I can see that twenty thousand people being individually right is no guarantee of an organisation getting anything right.

We’ve known for decades that collective performance depends on more than individual competence. In a 2010 meta-analysis of 65 studies, Leslie DeChurch and Jessica Mesmer-Magnus found that shared understanding within teams contributed to performance over and above how people behaved and how motivated they were. Being good at your job is one thing. Knowing how your job fits with everyone else’s is another.

Jody Hoffer Gittell found something similar in her research on airline operations. At American Airlines, employees understood their own tasks but had limited awareness of the wider departure process. They could do their individual jobs without necessarily understanding how those jobs fitted together. Her subsequent work on relational coordination found that shared goals, shared knowledge and mutual respect help people coordinate work across organisational boundaries.

Which is worth remembering the next time someone demonstrates an AI agent that can complete an employee’s entire workflow without requiring them to speak to another human being. Removing unnecessary handoffs is excellent; removing the understanding of how your work connects to everyone else’s is something else entirely.

You need shared understanding to coordinate. Without it, one team is cutting costs, another is investing in growth, and a third is explaining to the board that both are happening simultaneously. Everyone hits their targets and nobody can explain why the organisation is going backwards.

You need it to disagree productively, too. There’s a massive difference between arguing about what to do and spending three meetings establishing what was actually decided in the first place. The former is how organisations make progress – the latter is how you end up with a steering committee.

And there’s another dimension that productivity metrics don’t capture: belonging.

People need to understand what they’re part of, how their work connects to their colleagues’ and why any of it matters. A personalised AI agent might make my working day easier. It might even make me happier. But it doesn’t necessarily help me understand what we’re doing together.

Nobody raises a service ticket saying, ‘We appear to have lost our collective sense of purpose.’ And if they did, they’d receive an automated response confirming that their sense of purpose was important to us, followed by a link to the employee assistance programme. The individual productivity dashboard can be glowing glorious green while the organisation becomes progressively less capable of acting as one.

We know how to measure whether AI makes an employee faster. We’re much less interested in whether it makes twenty thousand employees better at being an organisation.

Building for the whole

Before we spend another seven figures teaching AI agents to think, perhaps we could establish whether the organisation has managed to think anything through itself.

This doesn’t require a new category of enterprise software (which I know comes as devastating news to several hundred people currently preparing pitch decks). It requires someone to take responsibility for what the organisation knows, what it’s decided and which version of reality employees are expected to inhabit.

Not every document, conversation and half-remembered decision. That way lies an enterprise knowledge management transformation programme, three reorganisations, a consultancy-led maturity assessment, an entirely preventable exodus of competent people, and a steering committee that eventually decides to buy SharePoint.

I mean a relatively small set of facts, decisions and priorities that ought to be consistent wherever an employee encounters them, with clear authority over who can change them.

Those facts need authoritative sources, named owners and a process for keeping them current. And if we’re expecting AI agents to answer questions about the organisation, they need to know which sources to trust. Giving an agent access to seventeen systems isn’t the same as giving it the ability to distinguish an approved policy from a PowerPoint somebody abandoned in 2023.

We also need to distinguish between information about an individual and information that belongs to the organisation.

My salary, performance review and medical records are personal. They should be protected accordingly. The company’s published headcount, strategic priorities and approved policies are shared knowledge. They need appropriate access controls too, but they shouldn’t be treated as though every question is a request to rummage through someone’s personnel file.

This is partly a technical problem of classification, permissions and retrieval. But it’s also an editorial one. Someone has to decide which version of a fact is authoritative, what it means, when it changed and how employees should understand it. AI can’t resolve a disagreement the organisation hasn’t bothered to settle.

And finally, we need to stop treating personalisation as an unqualified good.

Of course a warehouse worker and a finance director need different information. Of course a colleague in Austria needs different employment policies from someone in Australia. Nobody is arguing for the return of the homepage carousel, where the CEO’s latest thoughts on quarterly earnings compete for attention with the lunch menu.

But some things need to be shared. A strategy isn’t much use if every employee receives a bespoke interpretation that happens to flatter their department’s existing priorities. Nor is a policy if the answer changes depending on which agent you ask.

Personalisation should change how people receive the truth, not give everyone their own version of it.

This ought to be familiar territory for internal communication professionals. Establishing a common understanding of priorities, decisions and organisational identity is – or should be – a big part of the job.

Let’s not pretend that spending two decades publishing leadership announcements nobody asked for has somehow qualified us to become the guardians of organisational knowledge. Earning that role requires editorial judgement, content governance, the ability to challenge contradictory leadership messages and the willingness to tell someone that their lovingly crafted departmental strategy is incompatible with the one the organisation supposedly agreed.

It’s hard to sell a product that tells a senior executive their strategy contradicts another senior executive’s strategy. Much easier to sell them both an AI agent that will explain each strategy with total, glassy-eyed, Elizabeth Holmes conviction.

The organisation-shaped hole

We may be heading towards a world where employees rarely visit an intranet, navigate a page tree or even know which system holds the information they’re using. They’ll ask a question, an agent will find an answer, and the machinery behind it will become invisible.

Fine by me. I’ve spent twenty years watching people struggle to find things on intranets. Contrary to popular belief, I’m not emotionally attached to the experience.

But somebody still has to maintain the common knowledge that makes those answers trustworthy: resolving contradictions, keeping decisions current and making sure there’s an authoritative version of what the organisation knows.

That task doesn’t disappear because we’ve replaced the interface. If anything, it becomes more important when employees stop visiting a shared destination and start receiving answers through thousands of private conversations with AI agents.

Otherwise, we’ve just built a more sophisticated system for distributing organisational nonsense. Faster, more personalised nonsense, delivered with the confidence of a machine that has never once wondered whether it might be talking bollocks.

An organisation exists because people working together can achieve things they couldn’t achieve separately. Right now, every meat puppet is being optimised for productivity and the machinery connecting them for efficiency. Both are worthwhile.

But the value of an organisation isn’t simply the sum of its employees’ completed tasks. It’s what those people can understand, decide and achieve together.

We risk building workplaces where every employee has a brilliant AI assistant, every workflow runs beautifully, and the organisation itself has absolutely no idea what it’s doing.