AI can cost a company its experts: what we saw when parsers freed up 60 hours a week

Several pieces this week arrived at the same idea: AI takes routine work off employees, and along with the routine, experience leaves the company. The person who spent years retyping insurance estimates knew which ones only looked clean. If a parser reads the documents now and that person has quit, the knowledge leaves with them. For a business owner the question is practical: two years from now, who will explain to a new hire why this particular request has to be checked by eye?

What came out

On October 26, Camunda published a breakdown of its survey of insurance companies. 43% of employees who work with AI or automation had to redo the result by hand because nobody adjusted the process for it. 73% were never asked how AI should support their work, and four in ten would consider changing jobs over poorly implemented AI. According to Sedgwick data cited by Camunda, almost a quarter of claims adjusters will retire by the end of 2027. The recipe from an orchestration vendor is what you'd expect: move the experts' rules into the process model, and the knowledge stays with the company.

On October 23, FreightWaves reviewed a report by ASCM and DSCI on AI in supply chains, released on October 14. 83% of respondents consider their staff not ready to work with AI. Among the 17% who are ready, 71% of projects succeed; among the rest, 28%. The words closest to me come from Abe Eshkenazi, CEO of ASCM: a specialist used to build judgment while putting reports together and working through spreadsheets himself, and now AI takes that work away and people are asked to make decisions without that foundation.

On October 19, the same outlet ran a partner piece from Evos about dispatcher knowledge that isn't written down anywhere. Evos launched on October 14 and promises to take whole areas of work off a team, from dispatching to billing. It collects a company's logic during onboarding with questions like “why did this load go on that truck and not the new one?” This promotional text has no numbers on results. But it frames the problem correctly: the rules a dispatcher decides by live in his head.

Directum wrote about the same thing on Habr on October 26: a company can have hundreds of regulations and instructions, and the most popular standard is still the spoken one, “show me how this is done.” I only saw the announcement, so I won't judge the article's conclusions.

Pew Research Center, on October 22, ran the numbers across 480 occupations: for each point of AI exposure, employment of men grew by 6.3%, while for women it barely moved, 0.2%. Pew does not claim a direct causal link.

On October 19, SaaStr summarized Anthropic's report on its own workforce. The share of juniors among new technical hires fell from 27% to 9%. By August, agents were working 6 times more hours than the people managing them, and about 80% of the time they were waiting for a human to respond. The line that matters most to me is about juniors: stop hiring beginners, and in a few years there will be nobody left to become the experts.

Example: a factory and its process engineer

Take a manufacturer where the product structure is described in Excel specifications. The process engineer knows that parts going to paint have to be sent in one batch, or the shade will come out different. While he calculates an order by hand, a junior design engineer sits next to him and picks this up. If a program calculates the order in a minute, the junior engineer gets a finished answer and never learns why it looks that way.

So automation should take over moving data and leave the rules in plain sight. A rule like “paint in one batch” should be written down in words where the next person will read it, in the code and in the documentation.

What happened at our company when the parsers freed up 60 hours

Since 2021 I have been the technical partner of a PDR company in the US. In 2023 I started building parsers for insurance estimates, reports and VIN photos. Back then about half of the documents went through without manual corrections; in 2024 it was about 75%, and in 2025 roughly 95%. In 2026 I added AI validators and shipped the release version: about 99.9% of documents now go through without corrections. This frees up about 60 hours of manual work a week.

So where is the knowledge of the employees who used to retype estimates? Part of it is in my code. Most of the work is done by our own logic, which knows how these documents are built: which keywords to look for values by, when a line belongs to a different topic than its wording suggests, how items add up into groups. Part of it stayed with people: an employee does the final check, and documents the parser wasn't sure about have a path back to a person.

Over this time the company grew from 20 people to more than 30, lead volume multiplies from year to year, and the sales team grew only a little. The employees who used to retype documents into a spreadsheet now work on more important tasks.

What I would do

Before automating, write the rules down together with the person who does the work. A list of exceptions is enough: “if the estimate has such-and-such a line, check it by eye.”

Keep a path back to a person and track how often it gets used. If there are a lot of uncertain documents, improve the rules, but keep the check.

Keep a task that new people learn on. Let a new employee spend their first week checking the parser's output by hand; that's how they get to know the documents.

I have a separate review on how I read the numbers in vendor case studies. Earlier I looked at where quickly prepared work gets stuck. My approach to checking results is described in the article AI in data processing and on the solution pages AI in a process where the result can be checked and process automation. Projects: document parsers and technical partnership with a PDR company. If the know-how at your company lives in two people's heads, tell me which tasks they handle. Discuss your task.