The 07:00 bus from Mohakhali drops the morning shift at a lab attached to a research hospital on the south end of the campus, and by 07:14 the technician on bench M is already on her stool with the stereo microscope warm and the kettle on. The bench has six petri dishes lined up in stands, lids ajar, agar the colour of weak honey, with colonies that bloomed overnight in the incubator. Her job, on paper, is to look at each plate through the microscope, count the colony forming units, and write a number into the lab notebook. She has done some version of that job for the better part of a decade.
In March a small instrument arrived on the bench. An automated colony counter, made by a vendor whose representative came in person twice to install it, with a camera over the plates and a model trained on a few hundred thousand reference images. The instrument reads each plate in about twelve seconds, prints a count, and flags the plates it is not confident about. The technician still comes in at 07:00. She still sits on the same stool. The work she does on the stool is almost completely different from the work she did in February, and nobody upstairs has updated her job title.
This is the version of the AI story that the newspapers in Dhaka are not telling. The whole anxiety in the air, the one my friend Rifat in Bashundhara R/A keeps texting me about at one in the morning, is whether the model will replace the engineer. The boring real answer is that the model already replaced something one floor down, around eighteen months ago, and the people in the room mostly kept their jobs because the work was easier to merge than to delete. QA was the first role that happened to. Eval design merged with the generation of test cases, and regression suite maintenance merged with the discipline of rotating prompts, until the test engineer had quietly become a prompt engineer without anyone bothering to change the title on the door.
Cem Kaner and James Bach have been writing about testing driven by context since the late nineties. The whole argument of that movement was that good testing is a skilled human activity that depends on understanding the product, the customer, and the risk model, not on a checklist. That argument aged into the eval era almost unchanged. If you read a Hamel Husain post on evals replacing QA and a Kaner chapter on exploratory testing one right after the other, the prose styles are different and the underlying claim is the same. The model has to be tested against the actual situation it will land in, with someone who knows the situation doing the testing.
What this looks like on the bench, in a real lab in Mohakhali, is that the technician now spends about three quarters of her morning looking at the plates the autoreader flagged. The instrument is good at the straightforward plates, the ones with twenty or thirty distinct colonies on clean agar. The plates it is bad at are the interesting ones: two colonies that grew into each other and now look like one large irregular blob, or a piece of dust on the lid that the camera might or might not be reading as a colony. Then there is the faint translucent colony the model has never quite settled on. She looks at those, makes a call, writes the number in the notebook, and signs her initials next to it. The instrument logs the count it was unsure about and the technician's correction lands as a row in a pipeline of training data the vendor will pull from the next quarter.
This is the part of the story that the LinkedIn posts about AI productivity keep missing. The model takes the volume work and leaves the edge case work, which sounds like a productivity gain until you ask who will be left who can handle the edge case in five years. In Lisa Crispin's Agile Testing she has a line that I keep coming back to, about how testers earn their credibility on the cases where the system surprised them and they were the ones who caught it. Take the volume away from the tester and the surprises become a smaller share of a smaller body of work, and the muscle for catching them slackens the way any underused muscle does. The technician who only ever sees the flagged plates is, in the long run, a better technician than the one who counted three hundred plates a day, but only if the lab keeps feeding her flagged plates and pays her to think about them carefully.
Watch the same shape play out on the engineering floor. The test engineer who used to write assertions against an API endpoint is now writing rubrics for a model. The thing she actually does is sit with a domain expert, work out what "good" looks like for this customer journey, and codify it. That is the same activity she was doing in 2019, with a different syntax around it. What actually made someone good at this was knowing the difference between a customer outcome and an internal proxy, rather than fluency in any particular assertion library or eval framework. That is the one thing everything else in the job rests on, and the one thing that transferred. The boilerplate for unit tests did not transfer because Copilot writes most of it now, and most of the test engineers I know say good riddance, because they did not get into the work for the boilerplate.
The next role this happens to, by my own quiet bet, is technical writers and spec authors. The model is already drafting the boilerplate of an API reference and the boilerplate of a product spec. What survives, and what gets harder, is the part that depends on knowing the customer the spec is for and the politics of the room the spec has to land in. That is the same shape as the bench. The volume work goes to the instrument, and the edge case, the disagreement, the call you have to make with your initials next to it, stays with the human who has done the work long enough to know what an unusual plate looks like.
The technician on bench M still signs out at 14:30, and today's sheet has one red pencil mark on it, next to plate M-04, the plate she counted again by hand because the first number had looked wrong to her, initialled underneath.
That mark is the whole job now.
The counting was automated eighteen months ago and nobody in the building misses it, least of all her. What did not get automated is the judgement that plate M-04 deserved a second look. There is no field on the form for it, no metric that rewards it, and no record anywhere of the four other plates she looked at twice and let stand before the next shift trusted the notebook completely. As far as I can tell, nobody has ever audited it.