The CTO Augmentation Line™
It is the function where AI adoption is furthest along, where the evidence is clearest, and where everyone else in the business is watching to see whether it works. It is also the function where the value is most visibly stuck, and not for the reason most boards assume.
Code is no longer the bottleneck, the human-speed steps around it are. Anthropic, The AI-Native SDLC Playbook
The model is not what is holding you up. What is holding you up is a lifecycle built for a world where writing the code was the expensive part: approval queues sized for weekly output, review capacity sized for what a person could type, and security checkpoints that assume a human is waiting at both ends. Those policies were correct when they were written. They are now the constraint.
Every engineering function has a line between what a bounded agent runs and what a named human signs. It is not one line, it sits in a different place at each of the six stages, and the honest answer differs at each.
Intent captured in the originator's own words, in version control, before anyone opens a ticket. If intent lives only in a meeting, everyone downstream is guessing, agents included.
Requirements and design collapse into one working session, bounded by the organisation's own standards, which means those standards have to exist in a form a machine can read.
The plan is reviewed before the code is written. That single inversion, approve the approach, not just the diff, is what makes the volume reviewable at all.
Work verifies itself before a person is asked to look. Your test suite stops being a safety net and becomes the thing that makes delegation defensible.
Governance enforced as code rather than as a paragraph in a policy document. A gate that depends on someone remembering it is not a control, it is a hope with a process diagram.
Monitoring that does not merely alert but writes the next problem statement, which closes the loop back to Plan and makes the whole thing a cycle rather than a pipeline.
In most functions it is none of the six stages. It is the policy layer wrapped around them, written for output at human speed and never revisited.
Weekly or monthly gates become intractable when the build finishes in an afternoon. The queue, not the work, sets the delivery date.
Reviewers sized for what people could type cannot absorb what agents produce. Quality does not fall, throughput does, and then quality follows.
Ceremonies that exist to forecast human effort keep running after the effort has moved. They now cost more than the work they estimate.
Correctly non-negotiable, wrongly implemented as a queue. The control is right; its position in the flow is what is expensive.
Conventions no one wrote down cap how consistent anything can be, human or machine. This is the cheapest one to fix and the most often skipped.
Rules written for a previous generation of tools, enforced by people who cannot say what today's risk actually is.
The failure to avoid is symmetrical. Removing a gate because it is slow, without rebuilding the control it carried, is not modernisation, it is deregulation with extra steps. Every gate in that list was written for a real risk. The work is to rebuild each control at the speed the work now moves, not to delete it and hope.
Four moves, in this order. The order is the method.
Thirty questions across the six stages plus the policy layer. You get your band, your gaps and the shape of a 90-day sequence on the spot, free, and yours whether or not we ever speak.
Measure where the time actually goes between an idea and production. In most functions the answer surprises the people who own the process, which is the point of measuring rather than asking.
Take the single heaviest gate. Establish what risk it was written for, then rebuild that control so it runs at delivery speed. One, done properly, is worth more than a programme.
Agree stage by stage what an agent runs and what a named human signs, written down, enforced where it can be, and reviewed on a date that is in the diary.
Not lines of code, and not seats deployed. The measures that matter are about elapsed time and what survives contact with production.
The CTO line is one altitude of the same method. The others are already published, and the pattern is deliberately identical, score the function, find the drag, rebuild one control, draw the line.
Finance, where there is no tolerance for a hallucinated number.
Procurement, where the evidence has to survive a contested process.
Assurance, where the standard is what you can show, not what you did.
What happens when an agent takes its authorisation from a message. Worth ten minutes of any engineering leader's time.