Chapter 11
On Track!
The Utopia of the Agentic Enterprise, Part II. The Agents Who Take Your Job
In a large organisation, the gap between what a project reports upward and what is true on the ground is a stable feature of how projects run. It’s widest where there’s no comparable project to check the report against, and AI is still too new to offer many. Most people who have sat through a quarterly review will know the slide: a portfolio of 15 initiatives, 11 on track and 4 addressing initial quality issues. In many cases, those 4 are not producing the outputs the business case projected, and their delivery teams have rotated through a second or third lead within the year. Which, to be fair, is more candour than a slide where all 15 are green.
Nobody has to lie for this to happen. Everyone in the chain smooths the rough edges of their own area, competently and mostly in good faith, and by the time the quarterly review summarises the portfolio for the steering committee, the edges have been sanded into something you can present. If the system reliably rewards smoothing, the difference between smoothing and deliberate concealment matters less than executives would like to think. So the organisation has decided, through the normal operation of its people and processes, not to know what its own data would tell it.
Every step in that chain does a flunky’s work in Graeber’s sense, making the person above feel that the project is in hand. Agents now draft the summaries at each step, and a model trained on its users’ approval smooths without being asked. The steering committee pays later, when it funds the next phase on the strength of the slide.
The technicians submit a technical summary to their leads. The leads soften it for the PMO. The PMO rolls it into a portfolio view for the transformation lead, who selects what the steering committee gets to see. At every step, something gets smoothed. A note that leaves the technicians saying “reliability concerns” reaches the committee as “initial quality issues being addressed through iterative improvement”, which takes four times the words to say rather less.
Of course, the smoothing also runs the other way. The executive sponsor’s appetite for difficulty gets softened on the way down. What the sponsor says in a pre-brief, “we need this to work in six months”, reaches the technical team as “the business really wants this by Q3, push to make it happen.” Somewhere between the two, working turned into a date. The technicians hear the request through the filter of the people who depend on the sponsor’s goodwill.
Chris Argyris spent 30 years studying this and called it an organisational defensive routine: a practice that protects people from embarrassment or threat, and in doing so prevents the organisation from finding out why it’s failing.1 He found that the routines are skilled. People are good at them and can’t see themselves doing them, which is why a workshop on candour changes nothing, even if everyone rates it highly in the feedback survey afterwards.
Meanwhile, each position in the chain has its own plausible deniability.
The technicians reported their concerns in the right template at the right time. Covered. The executive sponsor relied on the information the team provided, as sponsors do. Covered.
The result is an outcome nobody is accountable for. All it requires is that each position protects itself against being the one named when the outcome needs to be explained.
Nils Brunsson, studying Swedish public organisations in the 1980s, went further and argued that the gap is functional. An organisation facing contradictory demands, be seen to transform and don’t disrupt operations, satisfies one with talk and decisions and the other with action, and keeps the three loosely connected on purpose. He called it organised hypocrisy and meant it neutrally.2 The steering committee’s portfolio slide is the talk. The delivery teams rotating through their third lead are the action. The project survives as long as nobody puts the two in the same room, which, given the state of most calendars, takes no effort at all.
Reporting Meets Sanction
The technicians report to a lead, and that lead also signs their performance reviews. So “the model is not production-ready” goes on a status report addressed to the person who can staff its author off the project or cut their bonus. Which concentrates the mind when choosing adjectives.
The report softens, and nobody decides to soften it. Ten years of working life have taught the technician that the reports which trigger career consequences are the ones that flag problems loudly.
The fix, at least on paper, is to separate reporting authority from sanction authority. It is rarely implemented, because it means giving a quality function or an independent review body the authority to hear from the technicians without career consequences for them. That authority would come out of another function’s budget, run by a head with their own standing among peers to protect. So the fix costs something, and the cost is paid by whoever has to stop being the only interpreter of their own work. In many cases, that is also the person who would have to sign off the change.
Agile as Theatre
Agile was designed to produce fast learning in environments where requirements are unstable, which in software is arguably all of them. Sprint reviews exist to show what did not work. Retrospectives exist to decide what the team wants to change.
In the corporate adoption, the sprint review demonstrates progress to stakeholders, preferably with a slide. The retrospective produces action items that get logged and folded into the next sprint’s capacity, where they wait patiently without being addressed.
The ceremonies keep running, long after the learning has been extracted from them.
What the organisation ends up with is a weekly rhythm that looks like intense feedback. The team reports velocity points, because that is what the PMO wants, and whatever it learned about the product doesn’t reach the PMO in a form that would change the plan. Velocity tells you how fast the team is going, and nothing about whether it is going anywhere useful.
Communications, Vendors
Corporate Communications has a narrow lane where it is legitimately expert: the voice the organisation uses in public, and reputational risk during incidents.
The organisation uses the function outside the lane when it needs a written version of its own behaviour that is more orderly than the behaviour was. The AI transformation that did not meet its stated objectives becomes, in the year-end narrative, the AI transformation that laid the foundations for future gains. Foundations are a wonderful thing to have laid, because nobody expects to see anything standing on them yet.
A coherent story in public is worth something to the organisation, which is why the habit persists. But an organisation that routinely has Corporate Communications rewrite its own history loses the capacity to learn from that history, because the record the function produced no longer matches the events it describes.
Vendors and consultancies are part of the chain too. Commissioning an assessment from a major consulting firm buys an executive, among other things, the firm’s endorsement of a direction they have already chosen. The assessment comes back in the consulting firm’s house format, with the partner’s signature, and establishes that the direction has been externally validated. The invoice helps, since an expensive assessment is harder to argue with.
In many cases, executives select the firm precisely because it is predictable in this regard. Over time, the market selects for firms that find what the commissioning executive wants.
The coverage, then, extends outside the organisation. The external assessment and the analyst report that names the organisation a leader come out of the same mechanism that produces the smoothed status slide in the steering committee, only with a nicer cover page. What the organisation has chosen not to know now comes with an external certificate, which makes it much harder to dislodge.
Inside many organisations, the first return on AI is theatrical rather than operational. The company needs a line in the annual report, and a reason for leadership to look as though it has not spent three years being ambushed by the technology in quarterly increments.
An operational implementation takes years of integration work and the retirement of several internal fictions. A reportable implementation takes a pilot and a chatbot with a confident name.
And the reportable version is worth money, because investors and boards reward AI a company can show over AI it is still building, and that makes the visible version necessary well before the operational version is ready.
Inside, incentives follow the money. Executive equity grants increasingly include AI-transformation milestones as vesting conditions, and the milestones get defined by what an external committee can verify: a launched pilot or a published maturity score. A milestone that said “the thing works” would be harder to verify, and would probably vest less often. Sell-side analysts cover AI maturity as a valuation signal, and the signal is whatever the company chooses to disclose. Every one of these mechanisms rewards producing the reportable version, whether or not there is an operational version behind it.
So the pilot can work as an alibi. Once it exists, the transformation can be declared, and once declared, much of the political and narrative value has already been collected. The retrofit work begins after that, when the attention and the budget have moved on. Meanwhile, measurement drifts towards the presence of AI rather than the value of it: seats licensed and logins. And an organisation that counts logins has stopped measuring value and is now checking whether the congregation still turns up on Sundays.
AI is particularly susceptible to this pattern, because its outputs are plausible without having to be correct. The steering committee that wants a supplier-consolidation review can now generate one. The assessment that used to take four weeks now takes four days, comes in the format the committee expects, and reaches the conclusions the sponsor prefers. Which is a genuine productivity gain, as long as nobody asks what the review was for.
Many such artefacts exist for legitimacy coverage. Nobody expects a consolidation review prepared for a steering committee to change the decision to consolidate. It is expected to exist, as the record that a review occurred. The AI-drafted assessment does the same work the consultant-commissioned assessment did, at a lower cost per artefact, and with the added benefit that no external party knows what was commissioned.
The behaviour of the people hasn’t changed. The artefacts got cheaper, so more of them get commissioned, and every one of them still needs a slot on the steering committee’s agenda.
Adding review steps doesn’t fix this. If the reviewers are part of the chain, the review produces coverage of itself. The gatekeeper reports to the transformation lead and passes the gate, because the lead’s budget depends on the project continuing. The facilitator works for the function under examination and produces a post-mortem that function is happy to sign.
The outcome changes when the reporting path bypasses the person who hands out sanctions, and the reviewer’s career doesn’t depend on the system looking like it works. Each is expensive to build, because each removes something the chain currently produces at no visible cost, which is cover for everyone in it.
Large organisations rarely confess to systemic failure. Executives depart and failed initiatives are renamed, so that what failed becomes foundational. The cycle runs every few years.
The turnover keeps learning from turning into accountability. Tracing who claimed what, and who got promoted on the back of the theatre, would be more than the organisation can take in one sitting, so the next leadership team prefers to start fresh, much as the previous one did.
And the timing helps. When the first generation of AI capital commitments comes due, leadership churn and budget cycles will offer plenty of chances to start over. The habits will survive the reset, and the audit trail, in many cases, will not.