In brief. Speeding up the production of work can multiply what another department has to check. The value shows at delivery, when the wait has already been paid for. At Anthropic the code changes merged each day went from around 2,500 to over 30,000 between May and September 2026.
During a company visit, I found a team getting ready to work in shifts to rename about four thousand images. The new CRM required a precise coding. Each photo had to be opened, the product recognized, the necessary information retrieved and the file given the right name.
We had AI carry out that step. I have described it in several talks because it was a very concrete example of work that can change quickly: the system looked at the images and applied the coding. The decisive part, though, had already been clarified. Someone knew what the files had to be called for the next step to accept them.
That condition deserves more attention than the time saved. The automation had a recognizable destination. Had we produced thousands of names for a colleague to approve one by one, we could have emptied one desk and filled another.
The code arrives before the infrastructure
At Anthropic this tension showed up on a very different scale. In the account published on October 13, 2026, the company describes going from around 2,500 code changes merged each day, the previous May, to over 30,000 in September. To sustain that volume it built Scion, an internal version control system, and completed the migration in five months. Anthropic describes its own workload as exceptional. Anthropic Engineering’s original account
The comparison is not meant to say every company will have to rebuild its IT systems. It serves to observe a consequence: when one phase speeds up a lot, the phases built around its old pace can become inadequate. Writing code faster does not eliminate the work needed to integrate it and make it function with the rest.
In a commercial company it can happen with quotes. The assistant prepares more offers; someone still has to check price, production capacity, terms and timing. The customer gets a benefit when a valid proposal arrives. Until that moment, part of the productivity is waiting on a desk.
The same offer can take a few minutes to write and several days before it can go out. Measuring only the first interval rewards a speed the customer does not see. Worse: it can push people to generate even more proposals and lengthen the queue of whoever has to check them.
Productivity is also measured from the point of view of whoever is waiting. Total time includes actual work, waits and rework. A department can improve its own figure while overall delivery gets worse.
Abundance can be useful
In creative work, research or design, producing more alternatives can be exactly the result you want. A larger number of attempts increases the chances of exploring a direction that at first seemed uninteresting. If we ask the tools to generate only what we are already ready to approve, we risk keeping the previous limit.
The useful distinction concerns when an alternative starts to take up other people’s time. A salesperson can explore many hypotheses before bringing one for review. A design group can compare variants without turning them all into formal requests to the technical office. Exploration deserves room; the operational queue needs entry criteria.
This separation has to be made explicit. A draft generated to think with should not follow the same path as an offer promised to the customer. If both land in the same inbox with the same urgency, the recipient has to work out every time what they were really asked to do.
In the catalogue case, the coding required by the CRM defined the useful result. With quotes, the equivalent can be a complete request, with customer data, quantities, terms and the point to approve. Checking starts better when whoever receives the work does not first have to guess what is missing.
Saved hours have a destination
To see where to intervene, it is worth following a small group of files from start to delivery. Note when they are produced, when someone opens them, why they come back. The manager can then tell a shortage of capacity from a shortage of information. Hiring a reviewer helps little if every file arrives incomplete; improving the form helps little if the necessary expertise belongs to one person who is already saturated.
The next automation should be chosen starting from that path. It may help to prepare materials better for whoever checks, to delegate the less risky exceptions or to reduce the number of decisions that require going up the hierarchy. Sometimes the solution is to give up earlier on offers with no reasonable chance of being accepted.
The new balance has to be checked too: speeding up a review can move the wait even further downstream. The process therefore requires a continuous measure of delivery, especially when the volume of requests changes.
The image team had a visible amount of work in front of it: four thousand files. The queues created elsewhere are often harder to count, because they are called emails, drafts, urgent requests. They still take up days.
Before celebrating the hours saved in one department, it is worth following the work to the next person. Their calendar says whether those hours were freed or only transferred.
Read next The first assignment is part of the training · The test comes before the code
Image: Anthropic Engineering