Foundations
AI employee FAQ
Tobiloba Odejinmi · 4 Sept 2026 · 7 min · 1,585 words
Direct answer
An AI employee is a named job that runs in your existing tools, with an owner, logs, and a path to a person when the model is wrong. It does not replace a team. It takes a pile a team is already tired of. If you cannot write the steps, name the owner, and accept review on the risky cases, I will tell you not to build yet.
- Start with one process, not a digital workforce.
- A person owns outcomes. The vendor does not.
- Wrong output becomes an escalated state, not a disappearing ping.
- Hiring and money make this a legal design problem, not a demo.
What people mean by AI employee
People search “what is an AI employee” and land on two different products. One is a chat tab with a name. The other is a job that runs while the team sleeps. I only sell the second. If you want a copilot in a sidebar, say so. The build, the owner, and the risk are not the same.
The definition I use: a named workflow, in your tools, with a stop button and a person on the exceptions. Support that drafts and routes. A screen that returns a shortlist with reasons. A follow-up that writes the first note and flags the call. Internal copy-paste between two systems. That list is the work. “Digital teammate” is not.
- Has a job description you could hand a new manager.
- Runs without a prompt on every item.
- Can be paused by the owner, not only by an engineer.
- Is measured on the pile: time, misses, aging escalations.
Ownership and mistakes
Search queries about responsibility are the right ones. Who owns it. Who tells the customer. What happens when it is wrong. If those answers live in a vendor’s status page, you bought a dependency. I will not hand over a workflow whose only expert is me, and I will not let “ops” be the name on the runbook.
When the model is wrong, the case becomes escalated. That is a state: reason, owner, payload, clock. Anthropic’s 2026 agents work keeps stressing the path back to a person. I agree. A fluent apology in Slack is not a path. A list you can query on Monday is.
People, jobs, and training
Teams ask if this replaces them. I tell them what the system will take: the first pass, the copy-paste, the midnight draft. I tell them what it will not take: the refund above the line, the candidate who does not fit the rubric, the clinic that cannot wait on a vibe. Vague reassurance creates two failures at once — fear, and rubber-stamping to look loyal.
Training is a week of real cases, not a town hall about the future of work. Teach review. Teach send-back. Pin three examples. If review time collapses to a few seconds, you did not get adoption. You got a stamp. BCG’s 2026 point still holds: value shows up when the operating model changes. The operating model here is “a person still looks.”
Building one and going live
How long: seven days for one process when the inputs exist. Day one we pick the work. Days two and three we map steps, tools, and the cases a person still sees. Days four to six I build in your stack. Day seven it is live, with docs and a call. That is the week I sell. It is not a mystery.
What I need from you: steps on a page, a counted pile, an owner, and doors into the tools. What I will not do: add a new login so the demo looks modern, or go live without a stop condition. IBM’s frontier share is still about 9 percent. Pilots that never get an owner explain a lot of the rest.
Failure, review, and escalation
There are three common kinds of wrong. Missing fact: retrieval or access. Broken shape: the system wrote an essay instead of fields you can check. Bad judgment: it ranked, promised, or prioritized in a way a person will not defend. Retrying the same prompt treats all three as weather. They are not.
Review is not optional on money, health, or someone’s job. Structured output makes review possible. Escalation-as-state makes it unavoidable for the cases that fail the shape. If you cannot explain a miss in one sentence, pause. I have sat on systems that moved serious money. A row that is wrong is not an interesting research topic.
Law: hiring, audits, and the EU AI Act
If the AI employee ranks, scores, or filters people for a job, you are not in a sandbox. New York City Local Law 144 requires an independent bias audit for automated employment decision tools that substantially assist hiring or promotion, plus notice to candidates. Ignoring that because “it’s only a shortlist” is how you find out what “substantially assist” means the hard way.
The EU AI Act treats many employment and worker-management systems as high-risk. That means risk management, data governance, logs, transparency, and human oversight — not a footer that says a model was involved. I will not build a hiring screen without a reviewer and a trail. If you need the legal memo, get counsel. If you need the workflow not to pretend it is counsel, that is the design.
Slack, email, languages, and tools
Yes, it can live in Slack and email. It should, if that is where the pile already sits. Do not add a dashboard people will open twice. Slack is a queue. Email is still the system of record in more companies than admit it. Write permissions down. A workspace bot that can read every channel is not “helpful.” It is a leak with a friendly name.
Multilingual is not a dropdown. Tone, legal language, and who reviews in each language are part of the job description. A model can draft in more languages than your team can defend. The reviewer has to exist in the language you will send. If they do not, you do not have a multilingual employee. You have a translation lottery.
What I will not automate
I will not automate a process nobody can write down. I will not put an unattended system on money, medical decisions, or a hire. I will not leave you with a workflow only I can explain. I will not treat a pilot that never got an owner as a success.
On the call I ask three things. What should we take off the plate. What would that save. What would it take to ship. If the honest answer is “you need a process first,” I will say so. Thirty minutes is enough to find that out. It is cheaper than a week spent decorating fog.
Questions people ask
What is an AI employee?
A production workflow that does a defined job in the tools you already use. It has a job description, a named owner, monitoring, and a way to escalate to a person. It is not a chatbot you visit when you remember, and it is not a copilot that waits for a prompt on every case.
How is an AI employee different from a chatbot?
A chatbot answers when someone types. An AI employee runs a queue: tickets, CVs, follow-ups, inbound calls. It takes steps in other systems. It can be paused. You measure it on the pile, not on how clever the chat feels.
Who is responsible when an AI employee makes a mistake?
The named owner of the process. Engineering owns a broken connection. A vendor owns their outage. A person still tells the customer. If you cannot point at that name, do not go live. “The model” is not a responsible party.
Do AI employees replace human workers?
They replace a first pass, not a function. At SmartComply, reviewers stayed. The pile got smaller. If your plan is to delete the people who understand the exceptions, you will learn why those exceptions existed. I will not help you staff a silent miss.
How long does it take to put an AI employee into production?
One process can go live in a week if the steps are written, the tools have doors, and an owner is ready to take exceptions. The week includes mapping, building, testing, and handover. If those inputs are missing, the honest timeline is “after we write the process,” not “after we pick a model.”
Is it legal to use an AI employee for hiring?
Sometimes, with a reviewer, notice, and an audit trail. In New York City, Local Law 144 requires a bias audit and candidate notice when an automated employment decision tool substantially assists hiring or promotion. The EU AI Act treats many employment systems as high-risk: oversight, logs, documentation. A ranking demo is not a compliance program.
Can an AI employee work in Slack and email?
Yes, and it should meet people there if that is where the work already lives. Slack is a queue, not a personality. Email still runs most companies. Neither place is where you should quietly move money or close a hire. Permissions matter. A bot in a workspace can see more than the job needs.
What should I do when an AI employee is unsure or wrong?
Move the case to an escalated state. Keep the payload. Give it to a person. Fix the case, then decide whether retrieval, the schema, or the scope was the cause. Do not only rerun the model and hope. If you cannot explain the miss, pause the workflow.
Written by
Tobiloba Odejinmi
Head of Engineering at 10mg Health. I have run engineering at Zeeh Africa and sold Insurpass and Shopl. I still write the code. If you have one process that still runs on people copying things, we can look at it in thirty minutes.
