How Compressing Training With AI Fills Seats Faster While True Expertise Still Requires Real-World Experience In The Role

A process runs slowly, so the workflow gets redesigned and new software gets bought on a promise. Sometimes it drops instead, the backlog starts building, and the conclusion is that people need more time with the new system. The training plan already says how much time, and that number usually goes unexamined. That number sets headcount and decides how much experience a posting demands before anyone reads a resume.

Jason Bell is a Business Analyst and Transformation Lead in operations at Providence Health Plan, where he has worked for 17 years, long enough to remember suites full of metal filing cabinets and colleagues whose entire eight-hour shift was pulling and filing documents. He describes himself as a systems person by instinct, someone who looks for the flaw in the process before looking at the people running it. His thinking changed after a software rollout meant to speed up claims processing did the opposite.

"Production went down and kept going down until the backlog started building," Bell says. He had expected a dip while people learned the tools, but what he got was an inventory problem. "Someone sold something and someone bought something based on a promise that it was going to be better, and it's not." He tried three or four fixes before deciding he was compounding the damage, a sequence familiar to any team that bought the tool first and asked what it improved afterward.

The four-hour test

Rather than sit on it for a year, Bell decided to find out what the work demanded by doing it. "What if I was a claims processor and I lied on my resume and said I knew how to do all of this, but I don't? I'm gonna fake it till I make it," he says. He worked from the same documentation available to any new hire and processed claims himself.

By his own account, he reached the hourly output of a strong performer within a single sitting. "It took me all of a good three, maybe four hours to where I could produce enough claims per hour that I was on average with someone who is a top producer." The comparison he ran was one of volume, and he doesn't claim to have measured his accuracy against the team's, or to have handled the cases that make the job hard. "Obviously we're going to run into all kinds of edge case scenarios and things where I'm not going to have the experience to deal with them," he says.

The same caution applies to Bell's estimate that roughly 90 percent of the workflow is suitable for automation. The figure applies to the claims process he examined rather than to work generally, and it covers the tasks he judged routine enough to hand to a machine. "It's reading one thing here, putting it somewhere else, deciding I need to do this or I need to do that. There's no open analysis here."

Thirty days with people who knew nothing about it

Trainers and managers pushed back when Bell brought the numbers, so he ran the test again with people who had no stake in the answer. "I hired a vendor and asked, 'What if we could do it in 30 days?'" he says. By his tracking, it worked. Six weeks after training ended, he says the vendor team was matching what his best in-house processors had been producing. Six months in, he puts their output at 150 percent of what the process had been handling before.

Bell attributes none of that to the work being trivial, explaining that the team carrying it was overqualified for what they were handed. "Most of the vendors I have doing this process have master's degrees and are well over qualified to be doing this type of work," he says, noting that they came in without the native language, the culture, or any history with the company.

The routine compresses faster than the craft

Bell's finding has a limit, and the research points to where. Researchers from Stanford and Harvard Business School ran a writing task at a fintech company, splitting employees into three groups by how much relevant experience each one had. Every group did the task with AI and without it. On coming up with ideas, all three did about equally well once AI was involved. Writing the piece was different. The group with some relevant experience nearly matched the specialists, while the group with the least barely improved. Experience is what tells people whether the AI's draft is any good.

Bell had already drawn that line for himself, separating the work he sped through from the work he couldn't. "Where you actually have to learn how to do something well, that is a craft that takes time," he says, separating skilled work from the procedural volume he tested. What his experiment moved was the speed of reaching competence on a defined process. Judging whether the output is right still rests on something the four hours did not supply, which is the same split engineers describe when they say the writing of code got cheap while judgment, design, and evals did not.

Where the next expert comes from

The routine work Bell compressed is also the layer where people historically learned the job. If AI strips out that layer, the ramp gets faster while the path to senior judgment could get thinner, a trade that engineering leaders are already watching as AI absorbs the struggle that builds experts. Bell's own test shows both sides of it. The edge cases he couldn't handle after four hours are the ones people learn to handle over years of doing the work.

Bell traces the old timelines back to an era when getting information took real time, and nobody went back to question them once it didn't. "People think it takes weeks to learn something that it takes minutes to read," he says.

What he asks managers to test is simple: Are the requirements actually based on the work? "If you believe something about them but you don't do any of their work, then you probably got a bias or assumption that is working against you," he says. Those assumptions can surface in requirements written for roles the budget can't actually buy, where the years of experience were never checked against what the job actually requires.

"We are giving them access to tools. Knowledge is a tool now, it's not a badge," Bell says.