AI Employee or New Hire? How to Split a Role Before You Fill the Seat

Somewhere in your company there is a job posting open for a coordinator. Operations coordinator, account coordinator, recruiting coordinator, the title varies, but the shape is the same: somebody to chase statuses, keep systems updated, move information from one place to another, and make sure nothing falls through. The posting is up because the work is real and it is drowning somebody who should be doing something else.

Loaded with payroll taxes, benefits, equipment, and management time, that seat will cost you somewhere around $50,000 to $60,000 a year. Every year. Before you spend it, it is worth running one exercise that almost nobody runs, because the framing everyone reaches for is wrong. The question is not "AI or people." The question is: which 60 to 70 percent of this seat is repeatable?

This article is the method. How to split a role into its repeatable share and its judgment share, what that split looks like on three real coordinator-class roles, the honest math on both sides, and, just as honestly, the cases where you should close this tab and hire the human.

What an AI employee actually is

The market has settled on "AI employee" as the name for an AI agent that takes over a meaningful share of one role's workload. The name oversells and undersells at the same time. It oversells because the thing is software, not a colleague: it does not own relationships, it does not exercise novel judgment, and it does not carry accountability. It undersells because what the software does, it does continuously, without backlog, and without ever deciding that trafficking status updates is beneath it.

There are two very different things sold under the name, and the difference matters more than any feature list. One is a subscription bot: built for the average company, it does one job its way, and you adapt your process to the product. The other is a custom-built AI employee: an agent designed around how the role actually runs in your business, wired into the systems you already use, with explicit rules about what it may do alone, what it drafts for human approval, and what it must never touch. The first is an app. The second is closer to a hire, which is exactly why the hiring question is the right place to evaluate it.

Everything that follows assumes the second kind, because the split method only works when the software can be shaped to the role rather than the other way around.

The split: take the role apart before you fill it

Open the job description you were about to post. Ignore the requirements section, which is aspiration, and go to the responsibilities, which are the actual work. Better still, sit with whoever is doing this work today, informally and badly, and list what the role really consists of, task by task. A coordinator seat usually decomposes into fifteen or twenty concrete tasks.

Now test each task against one question: is it repeatable? A task is repeatable when

• the inputs arrive in roughly the same form every time,
• the rules for handling them can be written down, including the ugly exceptions, and
• there is a clear check for whether it was done right.

Chasing a vendor for a status update is repeatable. Compiling the weekly report from three systems is repeatable. Scheduling interviews across four calendars is repeatable. Drafting the first version of a routine client note is repeatable, because a human reviews it before it goes anywhere.

What fails the test is just as easy to name: relationships, where the value is a person knowing a person. Taste, where "done right" cannot be written down. Negotiation, where the other side is also improvising. And genuine exceptions, the situations that are new every time and need someone to actually think. That work is human, and no honest vendor will tell you otherwise.

Run this on a coordinator-class role and the repeatable share typically lands at 60 to 70 percent of the hours. That number is the whole decision. The repeatable share is buildable as an AI employee. The judgment share still needs a person, and here is the part that changes the hiring plan: a person who is only doing judgment work often fits in part-time hours, or inside a role you already have. The AI employee does not replace anyone. It replaces a hire you no longer need to make.

Three roles, split

Abstractions hide the decision, so here is what the split looks like on real seats.

An account coordinator at a small agency. The repeatable share: chasing statuses across clients and vendors, trafficking assets and approvals, assembling the weekly client report from the project tool and the ad platforms, and drafting routine client notes for an account manager to review and send. The human share: the client relationships themselves, reading the room on a tense call, and knowing when "fine" from a client means not fine. Split verdict: roughly two thirds buildable, and the judgment third belongs with the account manager you already employ, who gets their week back.

A recruiting coordinator at a staffing firm. The repeatable share: sourcing passes against the spec, screening summaries drafted from résumés and intake notes, interview scheduling across candidates and clients, and keeping the ATS honest. The human share: the candidate conversations, the gut read on fit, and the close, which is a negotiation with a nervous human on the other end and will stay that way. Split verdict: again around two thirds, and the recruiter spends their recovered hours on the only part that fills roles, which is talking to people.

A data entry seat. The honest near-total case. Documents arrive, values get read, validated, and keyed into a system, and exceptions get flagged. That is 80 to 90 percent buildable, with a human on the exception queue instead of the keyboard. If you are about to hire for pure data entry in 2026, run the split before you post anything, because this is the one role where the answer is nearly always the same.

The math, without the hype

A US coordinator-class salary runs roughly $45,000 to $55,000. Add the standard 25 percent or so for payroll taxes, benefits, and the rest of the load and the seat costs $56,000 to $69,000 a year, recurring, before a manager spends a single hour managing it. Recruiting the seat costs more still, and coordinator roles turn over, so you may pay that recruiting cost more than once.

A bounded AI employee build, one role's highest-payback workflows wired into one or two systems, starts at $3,500 as a fixed fee and is live in weeks. Running costs, meaning model usage and hosting, are modest: typically hundreds of dollars a month, not thousands. Suppose you assume the worst reasonable case, a build several times the floor because the role is thorny, plus ongoing management on a retainer. The repeatable 60 to 70 percent of the seat still costs an order of magnitude less to cover with software than with a salary, and the build cost is paid once while the salary is paid every year.

To keep the comparison honest, note what the math does not say. It does not say the AI employee replaces the whole seat; the judgment share still needs human hours, and those hours cost money too. It does not price the review time your team spends approving what the AI drafts, which is real, though it shrinks as trust calibrates. What it says is narrower and more useful: for the repeatable share of a coordinator-class role, paying a salary is the most expensive way to get the work done, by a wide margin, and the gap is not close enough for clever assumptions to flip it.

When you should just hire the human

The split is a method, not a sales funnel, and it regularly returns the answer "hire." Four cases come up again and again.

The role is mostly judgment. If the task inventory comes back 70 percent relationships, taste, and negotiation, you are looking at a real job for a real person. Buy them good tools and move on.

The volume is too low to matter. Automation has fixed costs. If the repeatable work adds up to a few hours a week, the build will not pay for itself, and the honest advice is a part-time hire or nobody at all.

The process is too unstable to spec. An AI employee executes a process. If the process changes every month because the business is still finding its shape, you would be automating sand. Hire a human, who is flexible by design, and revisit the split when the process settles.

Nobody internal will own the review. Every AI employee worth building runs with a human on the exceptions and the sign-offs. If no one in the company will own that queue, the system will decay quietly until it fails loudly. No owner, no build. That rule has no exceptions.

Why these projects fail, and how not to

The failure modes are documented, not folklore. Upwork's 2026 survey of SMB leaders found the two biggest barriers to AI adoption were data security and compliance concerns, named by 27 percent, and uncertainty about ROI, named by 24 percent. Both are rational, and both are answerable at design time rather than after the incident.

Security first, because an AI employee is an insider. It works inside your email, your CRM, your books, and an insider with broad access and no boundaries is a breach waiting for a prompt. The answer is the same discipline you would apply to a human hire, engineered in: scoped credentials with the least access that does the job, hard boundaries on what the agent can read, write, send, and spend, human sign-off queues on anything consequential, and a full audit trail so every action is explainable afterwards. This is why we think AI employees belong with people who also run a security practice: the cheapest time to design out the failure modes is before the thing is built.

ROI uncertainty has a structural answer too: bound the bet. A fixed-fee build on the role's single highest-payback workflow means the payback math happens on paper, before money moves, and the first live workflow proves or disproves it within weeks. If the numbers do not survive contact with reality, you have lost the price of a conference trip, not a transformation budget. What creates ROI uncertainty is the opposite shape: open-ended engagements billed by the hour, where the meter decides the cost and nobody wrote down what success checks against.

For what the discipline looks like when it holds, I can offer eight years of evidence. I have been fractional CTO of a New York promotions agency since 2017, where I built and still run an OCR receipt-promotions platform that has read, validated, and adjudicated roughly 100,000 receipts for brands like Unilever and Fandango, with humans on the exceptions and real money on every decision. That system is an AI employee in all but name, and the reason it has run for eight years is everything around the model: the guardrails, the exception queue, and the owner. The full case study is here.

Run the split on your posting this week

You do not need a consultant to start. Take the job description that is open right now and give it an hour.

• List the real tasks, fifteen or twenty of them, from the responsibilities section and from whoever is covering the work today.
• Mark each one repeatable or judgment: same inputs, same rules, clear success check on one side; relationships, taste, negotiation, genuine exceptions on the other.
• Estimate hours per week against each task, and total the two columns. The repeatable share is your build candidate; the judgment share is the role you actually need to fill, which is often part-time.
• Name the owner: the person who will review the AI's exceptions and sign off on its drafts. If there is no name, stop here and hire.

If the repeatable column comes back heavy and owned, you have a build decision, and the scope is already half-written. A single workflow is a fine place to start; that is our AI automation practice, and it is the narrower door. A whole role is an AI employee engagement: the role map, the build, the guardrails, and the team that runs it. Either way, bring the job posting and your split to a free 30-minute call and we will tell you what is buildable, what it would cost, and what should stay human. Straight answer either way, including "post the job."