The firm is not being automated. Coordination is.
That is the single fact underneath Salim Ismail’s Organizational Singularity, the HBR critique of AI layoffs, IBM’s work on human systems, and MIT Sloan’s theory of task chaining. Read as one argument, they describe the same inversion at three scales: the economics of the firm, the design of a job, and the culture that decides whether anyone will challenge the machine.
The inversion
Ronald Coase’s firm existed because it was cheaper to coordinate inside a hierarchy than on the market. Agentic systems drive the cost of execution — drafting, routing, reconciling, scheduling, first-pass analysis — toward zero.
The expensive remainder is no longer labor. It is the handoff: the review, the status meeting, the approval that exists because the last person cannot see the next person’s work.
Once that inverts, two mistakes become available at once.
The first is to treat the inversion as a headcount event. Cut the coordinators, keep the chart, bolt a model onto the old sequence, and call the smaller pyramid a transformation. HBR’s evidence is already in: more than half of firms that laid people off “for AI” later say the decision was wrong; almost none of the HR leaders would repeat it; customer-service cuts are already being walked back; engagement falls among the people who stayed.
There is no correlation between workforce reduction for “autonomous capabilities” and return. Smaller is not smarter.
The second is to treat the inversion as a tool event. Automate the email, the summary, the ticket reply, and leave the sequence that joins those tasks untouched. MIT’s result is that value doesn’t live in the isolated task. It lives in how tasks are ordered, clustered, and handed off.
A chain is only as automatable as its hardest link. Occupations with the same task list automate differently depending on adjacency: lecture-based teaching can be prepared and chained; continuous tutoring cannot. System-level efficiency can beat task-level perfection. An agent slightly worse than a specialist at two steps still wins if it removes the queue between them.
Both mistakes share a starting point. They begin with the org chart — roles to delete, tools to insert — instead of with the work.
The thesis
AI pays off when you redesign the workflow so that compatible tasks sit together, agents run the chain, and humans sit above the gates that still require judgment — then change the incentives, the managers, and the data plane so that the system can be trusted, challenged, and improved.
Headcount is an output of that design. It is not the design.
Three claims follow, and they have to be held together.
1. Work is a sequence, not a roster
A role is a historical bundle assembled for yesterday’s coordination costs. The thing AI actually touches is a task inside a sequence. Until you decompose the role, you do not know what can be chained, what must remain a human gate, or what exists only because someone used to carry information from desk to desk.
That is why Citigroup starts from processes rather than a number, why Walmart and JPMorgan treat stable or shifting headcount as a consequence of strategy rather than a goal, and why Ismail’s REWRITE begins by mapping the prescriptive workflow – including the tacit steps nobody wrote down – before anyone talks about a twin.
Design the sequence first:
- Cluster AI-compatible tasks so the chain does not break on an unstructured exception in the middle.
- Put a named human on every material gate: purpose, liability, taste, ethics, the call that has no playbook.
- Cut the relays. If a step exists only to pass state, it is a candidate for the stack, not a job to protect or a job to delete in advance.
Do this inside the existing hierarchy, and the hierarchy will reinsert the handoffs. That is the immune system Ismail is describing, and it is also why HBR insists on reversibility: you are discovering the design, not announcing it.
2. The firm becomes a thinking loop with a fiduciary shell
If the unit of work is the workflow, the unit of the firm is no longer the function. It is a loop: sense, interpret, decide, orchestrate, learn – wrapped in a control plane that is never off.
The intelligence stack is that loop made explicit. Purpose becomes a protocol, not a poster. Agents carry passports. Actions are logged. Rollback is granular. A human review queue sits on the anomaly. The ERP is demoted from owner of the work to a transactional consumer. The governed data layer – not the model – is where most enterprise AI actually fails, so it moves to the core.
The company does not dissolve. It thins into what Ismail and Shelton call a container for trusted agency: IP, brand, liability, purpose, and the people who bind agents to account. Coordination leaves the middle of the chart. Judgment does not.
This is the destination HBR will not put in a layoff memo and should not. A firm that runs on a fraction of today’s coordinators is a design target after the workflows have been rewritten and measured. Used as a calendar, it produces the surveys in that article. Used as architecture, it tells you where humans go: above the loop, as validators and exception handlers, not as relays inside it.
Build that loop at the edge, not as a decoration on the core. Fork the data. Rebuild one or two high-throughput workflows as a stack. Run them in parallel. Deprecate the old path when the twin wins on cycle time, quality, cost, and risk. Then take the next workflow. Under fifty people, you can rewrite in place. Over fifty, an internal program will be metabolized by the chart it is trying to replace.
3. The control plane is human, or it is not a control plane
A stack that cannot be questioned is not autonomous. It is unaccountable.
IBM’s gap is the operational form of GOVERN/ASSURE. Executives report role change at twice the rate employees feel it. They believe people are rewarded for AI skills while nearly half of employees say they were not trained.
A large share of both sides report that it is unsafe, or simply unused, to challenge a model output. Performance systems cannot see judgment. Managers still direct tasks rather than coach decisions.
The organizations that show the revenue and margin advantage are the ones that closed that gap: clear ownership of outputs, norms for when to trust and when to override, incentives that pay for learning and challenge rather than for speed of acceptance, sandboxes before scorecards.
That is SHAPE in Ismail’s language and change capability in IBM’s. Without it, task chaining just accelerates unexamined error.
The scarce human work is therefore not “the tasks AI cannot do yet.” It is the willingness and permission to say the chain is wrong, and a manager whose job is to develop that judgment rather than be the next relay.
The sequence that makes the thesis usable
Hold the destination and refuse the shortcut.
1. Pick two high-throughput workflows, not a politically safe pilot. Decompose them into tasks. Cluster, chain, gate.
2. Stand the new path up as a governed twin: data layer first, agents native to the work, logs as the record, passports and rollback on every agent, a human yes/no on every material decision. Run it beside the old path. Let evidence retire the incumbent.
3. Change the human system on the same clock: how performance is seen, how managers coach, how challenge is rewarded, how the people whose work was coordination are bridged into exception, design, and oversight.
Skip (1), and you are decorating a bottleneck. Skip (2), and the chart eats the design. Skip (3), and you get a faster system nobody will correct. Start with headcount, and you get all three failures at once.
The picture is not to replace the people. The picture is this: keep the judgment, keep the purpose, keep the learning that is yours, and stop asking a structure built for yesterday’s coordination to administer tomorrow’s speed.
Rewrite the work. Start an evolution before the revolution occurs. The firm that remains will be the one that can still be held to account.