Start Bottleneck Management Before Agentic Harness
Agile never really left software. Software teams assume the rest of the corporate world adopted it too, because that is who they talk to. Walk into finance, legal, manufacturing, or government, and the model still in charge is waterfall: define requirements, build, review, approve, move to the next stage. No sprints. No retrospectives. A chain of gates, each one requiring sign-off before the work is allowed to continue.
The recent SG GovTech’s restructuring is a useful trigger for this conversation, because it exposes what happens when a waterfall organization tries to respond to AI without asking the right question first. GovTech retrenched 93 workers in the first phase of a 2-year exercise expected to affect around 300 roles, or 7 to 9% of its workforce. The agency describes it as changing shape rather than shrinking, and says it expects to employ more people once the transformation completes.
GovTech Chose Vertical, Not Horizontal
A waterfall organization moves through a chain of gates, each one requiring sign-off before work continues. AI compresses the work sitting in front of those gates without compressing the gates themselves, since a checkpoint is sized to a reviewer’s calendar and a committee’s cadence, not to how fast the work arrives. Andrew Ng described this in software: when building gets 10 to 100 times faster, almost everything downstream becomes a bottleneck. The same logic holds outside software. Any checkpoint sitting downstream of faster work becomes a constraint, regardless of what industry the checkpoint belongs to. That mismatch is the real disruption, and GovTech’s restructuring is a useful place to see how an organization responds to it.
GovTech’s chairman said the transformation predates the AI wave, and there is no reason to doubt that account. The restructuring was framed around moving from vendor-managed delivery to in-house ownership, not around AI adoption. That distinction matters and should be taken at face value.
Read the GovTech transformation again with that in mind. The move from vendor-managed delivery to in-house build, operate, and secure is vertical integration. It collapses the handoff between GovTech and its vendors, pulls ownership inward, and reduces the number of external checkpoints a project has to clear. That is a real fix, and it solves a real problem.
What it does not do is address the horizontal question, and this is where the AI decision actually sits for any organization: does an AI investment make every existing team and function more valuable, or does it just concentrate the same amount of value into fewer people? Vertical integration answers who owns the work. It says nothing about whether the checkpoints inside that ownership, security review, compliance, testing, still match the speed of the work now arriving at them.
The same choice repeats everywhere organizations respond to AI. Using AI means pointing a tool at a stage of the waterfall and making that one stage faster. Applying AI means asking where the new bottleneck actually sits after that stage speeds up, and redesigning the checkpoint around it, not just the headcount around it.
Most restructurings right now do the first thing and call it the second. Consolidate roles, bring capability in-house, reduce vendor layers, announce the transformation. All of it is vertical. None of it touches the gate that will now stall the faster work sitting in front of it. The organization gets leaner without getting less bottlenecked.
What Actually Needs Reassessing
The honest diagnostic is not whether your organization runs Agile or waterfall. It is whether you can name, stage by stage, which checkpoints were sized for a throughput AI has already made obsolete. A legal review built around one contract a week does not survive a tool that can draft ten. A security clearance process built around one system launch a quarter does not survive a team delivering features daily.
The claim is not that AI caused GovTech’s layoffs. It is that any organization trimming operations today is doing so in an environment where AI has quietly become one of the considerations, whether or not it appears in the official explanation. The checkpoint problem described above is where that consideration will actually get tested, months after the restructuring is announced, when the faster work starts arriving at gates that were never rebuilt for it.
Horizontal improvement is the actual AI at Work strategy question, and it is the one vertical integration skips. It is not about consolidating ownership into fewer hands. It is about whether the same AI investment lifts every existing function, finance, legal, operations, delivery, so each one produces more value without anyone being pulled off the org chart to make the math work. That is value augmentation in practice, and it is a harder result to show in a board update than a headcount reduction, which is exactly why fewer organizations bother pursuing it.
None of this starts with the technology. An agentic harness is not a new capability dropped into an organization. It is an artificial replica of the working loop that already runs between people, the data that moves from one desk to another, the approvals, the informal checks, the conversations that fill the gaps no policy document ever wrote down. Deploy a harness before that loop is mapped and the agent is not automating a workflow. It is guessing at one. Bottleneck management, not agent deployment, is the actual next task sitting in front of most organizations, and it starts with writing the workflow down honestly enough that a machine, or a person, could follow it without you standing in the room.
 | A guest post by
|