Cisco gave every employee a personal AI agent in late August, and its chief automation officer, Srini Namineni, now expects the engineers using AI to produce many times more than before. His own caution is that the code and the testing are getting faster, while deciding what to design and how to deploy it still holds things up. That caution applies well beyond engineers.
A faster step moves the queue
In any chain of steps, the slowest step sets the speed of the whole chain. Cal Newport uses a chicken coop builder to show this: if putting on the roof is the slowest step, then better tools for the earlier steps only pile up half-built coops at the roofing station.
Software has now become that builder. In the firms that adopted AI coding agents, the extra code did not turn into extra finished work, because code review became the slow step. Most of those firms have also tried AI tools for the review itself, but AI wrote only about a quarter of the review comments. So the check is still mostly a person's job, and the person is now facing a longer line.
Why 100 times is such a high bar
The idea of the 100x engineer needs some arithmetic. Say writing and testing code takes 80% of an engineer's week, and AI makes that part instant. The week gets five times faster, not a hundred, because the other 20% (meetings, design choices, release approvals) has not moved. To reach 100 times, everything except the code would have to take under 1% of the week. No engineering team works like that.
I expect the gains to be real but uneven: three to five times for teams that also shorten their decision steps, and close to nothing for teams that do not. Namineni's point about backlog adds to this. Every IT department has years of it, and cheap code makes the backlog stop being a budget problem. It becomes a problem of choosing, because someone has to decide which of thousands of items is worth doing next.
The same queue will form in your office
Take a finance director at a hospital group. An AI agent reads 600 supplier invoices overnight, matches them to purchase orders, and writes a short note on each of the 90 that do not match. Matching used to take her team three days. Now it is done by seven in the morning.
But each of the 90 mismatches still needs someone with the authority to decide whether to pay. At four minutes each, that is six hours of her Monday, all of it decisions. The AI did the part that was easy to speed up and left her the part that was not.
The fix is to make each decision cheaper without taking the person out. The agent can write every mismatch as one line with its recommendation and the evidence beside it, so she can say yes or no in seconds. She can also give it a standing rule for the small, low-risk gaps, such as a difference under $50 from a supplier who is always right. Cisco's own agent works this way: it can do approvals, but only for tasks that are low risk or that the employee has approved, and a person stays in the loop. Releasing the payment itself stays with a human, because Namineni's rule is to keep agents away from payment systems.
The practical step is to find the slowest human step in your own work before you buy more AI for the steps that are already quick. Count what waits in front of it: the unsigned approvals, the unchecked drafts, the questions nobody has answered. The length of that queue is the real size of the gain, and the next hours you save will come from making those decisions quicker to make.