The wrong question is "can AI do this job?"
The right question is "what are the fourteen things this person does in a week, and which of them would they pay me to take away?"
Those two questions produce completely different companies.
What a loan officer actually does
Take a mortgage shop with eleven people. Ask the owner what a loan officer does and he'll say: originates loans. Ask a loan officer what she did last Tuesday and you get something else entirely.
She responded to nine borrower emails, six of which asked a question she has answered four hundred times. She chased a bank statement from a self-employed borrower for the third time. She re-keyed income figures from a set of pay stubs into the LOS. She compared the appraisal to the purchase contract to make sure nothing moved. She called the listing agent to update him. She sat on hold with a servicer for twenty-two minutes to get a payoff. She built the same rate comparison she builds every day. She had one genuinely hard conversation with a borrower who was about to walk, and talked him back into the deal.
That last one is the job. That's the fourteen thousand dollars of value produced on Tuesday. Everything else is throughput.
If you ask "can AI replace a loan officer," the answer is obviously no, and you'll conclude there's nothing to do here. If you ask "which of those fourteen tasks are mechanical," the answer is most of them, and you have a project.
The decomposition exercise
It takes an hour and it works for any role in your company.
Sit with the person. List every distinct task they perform in a normal week. Push for granularity — "handles borrower communication" is not a task, it's a category, and categories are where the money hides. Get to the atomic level: answers the "what documents do you still need from me" question.
For each task, ask two things. Is it rules-based — could you write down, on one page, how to do it correctly every time? And does it require judgment, relationship, or the ability to read a person?
Sort into three buckets.
Mechanical. Document collection reminders, data extraction from pay stubs and W-2s, status update emails, payoff requests, standard borrower FAQ, appraisal-vs-contract reconciliation. All of it is rules. All of it is automatable today, well, with an audit trail. This is usually forty to sixty percent of the week.
Judgment. The hard borrower conversation. The structuring decision on a complicated file. The call to the underwriter to advocate for an exception. The relationship with the referring agent. This is the reason the person is employed, and it is untouchable, and it should stay that way.
Contested. The interesting middle. Drafting the initial communication to a new borrower — a machine can draft it, a human should send it. Reviewing a file for red flags — a machine can flag, a human decides. This bucket is where the actual design work lives, and the rule we use is: the machine prepares, the human decides. Never the reverse.
Why this framing beats "AI agents"
The market right now wants to sell you an agent that does a job. It's a compelling story and it maps neatly onto how owners think about payroll, which is exactly why it sells.
But jobs, in real companies, are bundles. They're bundles of tasks that got grouped together historically, often for reasons that no longer apply — because one person happened to be sitting near the fax machine, because that's how the last system worked, because someone quit in 2018 and their duties got distributed.
When you try to automate a bundle, you fail, because ten percent of that bundle is irreducibly human and that ten percent is load-bearing. So the agent works for the easy files and breaks on the hard ones, which are the ones that matter, and everyone loses faith.
When you automate tasks, you succeed, because tasks are small enough to be specified, tested, and verified. And you succeed without ever having the conversation nobody wants to have about whose job it was.
The number
Back to the loan officer. Say the mechanical bucket is 45% of her week. She's at a loaded cost of $95,000, call it $47 an hour. Forty-five percent of 2,000 hours is 900 hours a year, or roughly $42,000 of her cost spent on throughput.
Multiply across four loan officers and you're at $168,000 a year in mechanical work, sitting inside people you'd never dream of letting go.
Now the honest part: you're not going to recover all of that, and anyone who tells you otherwise is selling. Realistically you get sixty or seventy percent of it, some of the reclaimed time refills with other work, and there's a real ongoing cost to run the system. Model it that way and it still clears 2x comfortably in year one. If it didn't, we'd tell you not to build it.
What actually happens
Here's what the owner cares about, once the dust settles.
His loan officers now close more files with the same headcount, because throughput stopped being the constraint. The borrower experience got better — updates go out the same day instead of when someone gets to it. The referral partners notice, because agents talk, and agents send deals to the person who calls them back.
And nobody got fired. The team got faster, and then the team got bigger, because the pipeline could support it.
That's the whole play. Don't automate the job. Automate the parts of the job that are beneath the person doing them, and then find out what they're capable of when they have the room.