The biggest change program of the decade is running at your company right now, unowned

Somewhere in your organization today, an employee gave an AI agent a piece of work that a colleague used to do. A team decided, in a channel you will never read, what their agents are allowed to handle without asking a human. A manager quietly changed what she expects from a first draft. A junior employee delegated away a task he was supposed to learn from. A workflow that touches your customers acquired a non-human participant.

No steering committee approved any of it. No communications plan announced it, no training calendar accompanied it, and no risk register anywhere carries a line for it. It does not appear in your transformation portfolio, which is currently tracking, at a guess, an ERP consolidation, a pricing change, and an operating model refresh, each with a named executive sponsor, a budget, and a risk register. Meanwhile the deepest change in how work gets done since email, the arrival of AI agents as working participants in teams, is being rolled out at your company by the company itself, informally, unevenly, and entirely without adult supervision.

This essay is about two mid-sized companies, composites assembled from real ones, that took opposite positions on that fact around the same time. Call them Meridian and Halcrow. Eighteen months later, the difference between them is the most instructive thing I know about the current moment. Neither company will surprise you at any individual step. The compound interest will. And to be clear about what this comparison is not: it is not adopters versus refusers. Both companies adopted, at comparable spend, with comparable enthusiasm at the top. The variable was ownership, and only ownership.

Meridian: the change that managed itself

Meridian's executive team was not negligent, which is what makes it the useful case. They did what most competent leadership teams are doing: they approved the tools, funded licenses, issued a responsible-use policy drafted by legal, and encouraged experimentation. "We don't want to over-bureaucratize this," the COO said, in a sentence that would cost more than any line item that year. Adoption was treated as a technology deployment with a training webinar, and the org was trusted to find its way.

The org found many ways. That was the problem.

By month six, Meridian had what an internal skeptic called "eleven companies in a trenchcoat." The GTM team ran agents aggressively: agents drafting outreach, handing research to each other, updating the CRM, pinging account owners when deals stalled. Finance banned agents from anything touching numbers after one early error, and meant it. Customer support built an elaborate agent triage layer that management learned about when a customer mentioned it. Engineering had three teams with three incompatible philosophies about what agents could approve. HR was not involved, because nobody thought to involve them, and so the questions accumulating in manager one-on-ones, is it fair that Priya's output doubled because her team uses agents and mine bans them, how do I performance-review agent-assisted work, what do I do about the analyst who no longer does any analysis, had no owner and no answers.

The costs arrived the way coordination costs always arrive: sideways. A deal review meeting where two teams' numbers disagreed because one team's agent-maintained pipeline was days fresher than the other's hand-updated one, and neither trusted the other's. A quiet exodus of two strong operations people who concluded their skills were depreciating faster in the ban-happy departments and transferred, or left. An embarrassing near-miss when a GTM agent, operating under rules that team had improvised, contacted a customer with pricing language legal had never seen. And, most expensively, the invisible cost: departments that had frozen, waiting for guidance that was not coming, while their industry did not wait.

The subtler damage was epistemic. Because nothing was written down, Meridian's leadership could not actually answer basic questions about its own operation. When the board asked what share of customer-facing work involved agents, the honest answer was "we don't know," delivered as a confident estimate. When a well-run pocket produced a genuinely excellent pattern, the support team's triage design was, by any standard, better than what most vendors sell, there was no mechanism by which the rest of the company could even learn it existed. The organization was simultaneously ahead of its market in three places and behind it in six, and could not have told you which places were which.

None of this appeared in any dashboard as "change program failing," because there was no change program to fail. There was only the slow accumulation of incompatible local decisions, each locally reasonable, collectively producing an organization that worked differently in every corner and trusted itself less each quarter.

Halcrow: the change with a name on it

Halcrow's insight was not technological. It was categorical. Their COO, a former change practitioner, looked at the same phenomenon and made a single reframing move: this is not a tool deployment, this is the largest change program we will run this decade, and it will be run like one, with the discipline the label implies and, just as important, the modifications the label needs, because this change never ends and the standard playbook assumes changes end.

The named owner came first. Not a committee; a person, reporting to the COO, with a small team drawn deliberately from change management, HR, and operations rather than IT, because the judgment call Halcrow made early, the one Meridian never consciously made at all, was that this is a change in how humans work together, in which technology happens to be the trigger.

Then, instead of a policy, they shipped working agreements, team by team, starting where energy already existed. Each team, with the change team facilitating, wrote down its own answers to a standard set of questions: what agents on this team handle autonomously, what they draft for human sign-off, what they must never touch, who owns each agent, and how the team renegotiates the line as capability moves, because it will move. The agreements were allowed to differ across teams, sales and finance should not have identical rules, but they had to exist, be written, and be reviewable, which meant the eleven-companies problem could not develop, because divergence was visible and deliberate instead of invisible and accidental.

Sequencing was chosen, not emergent. Halcrow rolled the pattern deliberately through functions in an order picked for compounding: operations first, because ops work is legible and the wins are quick, then GTM, then the sensitive functions, finance, HR, legal, last, with the strictest human-review rules and the benefit of everything learned upstream. Each wave was staffed for the Tuesday-after rather than only the launch: agents in the working channels answering how-do-I questions instantly, and the change team receiving a weekly synthesis of where friction was clustering, which teams were inventing workarounds, and which had gone quiet. The change lead spent her time exactly where that sensing pointed, in the human conversations underneath the signal. When the synthesis showed one region's questions stopping abruptly, she flew there, and found not resistance but a beloved local manager telling his people this would blow over. That conversation, manager to manager, executive sponsor eventually involved, is change management as it has always actually worked. The difference is she had it in week three instead of discovering the crater in month eight.

And the people questions got owned before they got sharp. HR, in the room from the start, rewrote what needed rewriting: how agent-assisted work gets evaluated, so that the answer to "is it fair about Priya" was policy rather than improvisation. What the development path looks like for juniors whose grunt work went to agents, so that apprenticeship was redesigned rather than silently deleted. Where the surveillance line sits, agents may assemble context for a person, never evaluate a person without their knowledge, so that trust had something written to stand on. When capability shifted, and it shifted three times in eighteen months in ways that mattered, the working agreements were renegotiated on schedule, because renegotiation was in the design. Refreezing was never the goal. Steering was.

The eighteen-month mark

The scoreboard reads like this. Halcrow's measurable gains, cycle times, response times, throughput per team, are real but honestly not the headline; Meridian's aggressive pockets posted similar local numbers. The headline is the variance. Halcrow's gains are broad, compatible, and compounding, because good patterns transplant across teams that share a common frame. Meridian's are narrow, incompatible, and plateauing, because every team's improvements are trapped in that team's private dialect of working. Halcrow can answer, in one document per team, what its agents are doing and who owns them, which turned out to be worth actual money when a major customer's diligence questionnaire asked exactly that. Meridian answered the same questionnaire in three weeks of panicked archaeology.

The scoreboard has a talent column too, and it may be the one that decides the next five years. Halcrow's job postings now describe, matter-of-factly, what candidates will work alongside: the agents on the team, what they handle, what remains human judgment. Candidates in the functions that matter have started selecting for exactly this, the way engineers once selected for companies with modern tooling, because ambitious people can smell where their skills will compound and where they will calcify. Meridian's recruiters, meanwhile, field a question they have no approved answer to, "how does your company actually work with AI," and watch candidates hear the improvisation in the reply. Attrition tells the same story from the other side: Halcrow's regrettable departures cited the usual reasons at the usual rate, while Meridian's exit interviews began growing a new category, phrased politely but unmistakably, of people leaving because staying felt like falling behind.

And the asymmetry deepens from here, because this change compounds. Every quarter of coherent adoption makes the next quarter's adoption easier, every quarter of fragmentation makes eventual coherence more expensive, and the gap between the two curves is not linear. Meridian is not doomed; it is now running, eighteen months late, the program Halcrow ran on time, except it is running it against installed habits, burned trust, and eleven incompatible local constitutions, which every change practitioner knows is the difference between planting and transplanting.

The strangest part of the whole comparison is how cheap Halcrow's advantage was. No moonshot budget, no consulting army, no proprietary technology Meridian lacked. One reframing, this is a change program, one named owner, and the unglamorous machinery of working agreements, staffed Tuesdays, and owned people-questions. The kind of thing any competent organization can do, and most will not, because the change arrived without a launch date, and organizations only take seriously the changes that send calendar invites.

Yours didn't send one either. It started anyway. Check whose name is on it.