1,100 Employees, 128 Manual Onboarding Steps: What Claude Found

Every HR leader has sat through some version of this meeting. Someone from finance says the quiet part out loud. The onboarding team makes too many mistakes. New hires are showing up without the right position code, without payroll set up, without whatever it is this week. Fix it.

‍

The usual response is to promise more care. Build a checklist. Add a second reviewer. Retrain the team.

‍

Greg Doss did something different. He asked one of his onboarders what was frustrating him, and got back a nine-page document.

‍

The number that changed the conversation

‍

Doss started at Community Medical Services in December overseeing recruiting. By February the company had handed him onboarding too. CMS provides medically assisted treatment for opioid use disorder across clinics in multiple states, employs about 1,100 people, and is backed by private equity. The financial controller's complaint about onboarding errors was not new, and it was not entirely wrong.

‍

The nine-page document listed every manual process the onboarding team ran. There were 128 of them.

‍

"I went up to him, I go, we have 128 manual processes," Doss said. "At a 90 clip, that's 13 mistakes."

‍

That one line reframed the conversation. A dozen errors per cycle was not evidence of a careless team. It was the arithmetic of running 128 manual steps at a realistic human accuracy rate. The process was producing the errors. The people were doing exactly what the process asked of them.

‍

What came back from the list

‍

Doss dropped the document into Claude with a straightforward request: get 128 down to 50.

‍

What came back was more honest than the ask. Claude could get him to roughly 70, and it documented which processes were candidates and in what order to work through them. That second part mattered more than the number.

‍

"If you handed me this hundred and twenty-eight, okay, I can get you down to one twenty-seven, one twenty," Doss said. "But this was like, here's the order in which you should go, here's how you should attack it, here's how you need to move forward."

‍

His team took the prioritized list to Paylocity, the HRIS and ATS that runs CMS. The vendor's technical lead read it and started crossing things off. Most of this you do not even need to be filling out. That one is already caught over here.

‍

"A lot of them weren't just manual processes," Doss said. "They were redundant processes."

‍

They are at about 70 now, and the remaining work is happening alongside Paylocity rather than around it.

‍

The meeting that stopped being about onboarding

‍

The clearest evidence that something had changed never showed up in a dashboard.

‍

Before the work started, the relationship between operations and onboarding had eroded into open distrust. CMS's chief operating officer set up a standing weekly meeting for exactly one purpose: let it all out. For the first three or four weeks, more than half the regional directors showed up and spent the hour listing what was broken.

‍

Within about three weeks, attendance dropped to two or three directors and the meeting shrank to 30 minutes. By the Monday before this recording, it had stopped being about onboarding entirely and moved on to a few open issues in recruiting.

‍

"We saw the problem. We solved for the problem. We're going to continue to manage the problem," Doss said.

‍

"How do I know this is right?"

‍

The question every HR leader gets before an AI-assisted number reaches a board packet came from Nick Stavros, CMS's CEO, who by Doss's account pushes hard on technology and expects HR to push with him. His version was short. He loves AI. The data has to be correct.

‍

Doss's answer separates two things people tend to collapse together.

‍

"I don't think AI is gonna take somebody's job," he said. "I do think AI is gonna replace people who don't utilize AI."

‍

The human check is the part he refuses to skip. A report that takes three minutes instead of eight hours creates an obvious temptation to forward it immediately. Doss reads it first, hunting for the number that does not make sense.

‍

That habit turned into a discipline. His team now treats every report as a checks and balances layer against the HRIS. Headcount climbed on one report and did not match what Doss expected. He cross-checked against Paylocity and found a gap of 35 to 40 employees, traced back to a small coding change inside the HRIS whose downstream effect nobody had followed. The same check has surfaced a single clinic entered under two different names, and nursing position types that were not flowing correctly between systems.

‍

Which leads to a small heresy about the oldest rule in reporting.

‍

"Bad data in, good data out as a result of seeing the bad data," Doss said. The old version of that rule held when nobody could see the bad data. Once two systems disagree in front of you, the wrong number stops being invisible and becomes a work item. His framing for the alternative is blunt: how many times have you looked at a report and realized you have had the wrong information for six months?

‍

Security is a policy problem before it is a tool problem

‍

Healthcare makes the AI security conversation concrete in a way most industries do not.

‍

About six months before this recording, IT and HR at CMS sat down together and wrote guidelines. What AI can be used for, what it cannot, what should never be put on your own machine, which prompts are in bounds. Roughly 50 people, Doss among them, were given Claude for Healthcare. Try to upload a spreadsheet containing patient information and it gets rejected.

‍

The line Doss draws is between groups and individuals. Analyzing the 25 to 30 year old employee base, or comparing male and female populations, is the work. Pushing anything down to the individual level is not. Wherever information can stay inside Paylocity, it stays inside Paylocity.

‍

"In the industry that we're in, people want to remain anonymous," he said. "We want to respect that as well."

‍

What the reports are actually for

‍

It would be easy to read all of this as an efficiency story. Doss keeps pulling it back toward something else.

‍

CMS treats people with opioid use disorder. Outcomes track with continuity of care, and continuity of care tracks with the tenure of the counselors and nurses in the clinics. "The longer tenured counselors, the longer tenured nurses that we have, the better," Doss said. Turnover in Ohio compared against clinics in Colorado is not a reporting exercise. It is a proxy for whether a patient stays with the program.

‍

The same logic ran through the 60-day onboarding survey. The first finding was not about paperwork. It said new hires were not being treated well by the people onboarding them. Doss's instruction to his team was one word, repeated three times: hospitality. The scores moved from a weakness to a strength, and the frontline feedback shifted to a new ask, more hands-on training. CMS rebuilt nursing and counselor training into a structured 90-day program on the strength of that signal.

‍

The advice

‍

Doss is not an early adopter by temperament. "I didn't get an iPad one, I got an iPad two," he said. He describes his own AI knowledge as having gone from not knowing what he was doing to being just good enough to get in trouble with it.

‍

That makes his closing advice more credible, not less.

‍

"A year ago is the absolute best time to do it. The second best time is today. So if you're on the fence about it, this is the time to get in there. You're gonna make mistakes."

‍

The mistakes are not the real exposure. The exposure is the HR leader who waits for AI to be finished, inside a company whose private equity sponsor has already asked what they are doing with it. Twenty years ago the consensus in recruiting was that LinkedIn would not last. Doss remembers.