Hire an AI Employee · set by the scope

You got a job to do, but not enough time.

Software assigned a bounded job inside your operation: the front desk, the follow-up, the work between systems. The agent gets the job. A person keeps the responsibility, and that person is who defines the job and makes the agent fit to do it. Hire one without the other and you have a login.

Already know you want an agent built and wired rather than a seat filled? That is Service 02, AI Agents.

We need something to call you.

That doesn’t look like a working address, and it’s our only way back to you.

Your answer comes back to this address. No mailing list. We check in as needed.

Spam and bot-written pitches eat up a lot of our time. You’re a real person with a real project, and a link to your LinkedIn or company site is the quickest way to show it. Optional, but it goes a long way, and it gets you a faster reply.

* Nothing sensitive, please: no passwords, no customer data. The shape of the job is enough at this step.

Have your own NDA? Send it. We sign reasonable ones.

A job you haven’t filled.
Software that can hold it.

An AI employee is software assigned bounded work, with a named human owner.

The work is the kind you keep meaning to hire for. The inbox that gets answered at nine the next morning. The second follow-up nobody sends. The copying between systems that someone does all day without ever having been hired to do it. An agent can hold work like that: it reads your systems, uses your tools, carries the task from asked to done, and writes down what it did.

Bounded is the word that matters. “Handle our operations” is not a job description, and neither is a person’s whole job. A seat is the part of one that can be written down: what comes in, what goes out, what done looks like, and what a wrong answer costs. Somebody has to write that down before an agent can hold it, and that is where the human comes in.

We use a lot of agents here, to do work, and every one has a job written down and a person who answers for it. The people who run ours are the people who would set yours up.
The human agents

Every AI employee
comes with a human one.

The human does more than approve the output. They make the job executable.

A business need arrives as a sentence: the quotes are slow, the leads go cold, nobody reads the filings. Turning that into a job an agent can perform is work, and it is the work. Somebody has to decide what the seat is for, which systems it may touch, what it may do alone, what it brings to a person, what done means, and what a wrong answer costs. Then the seat gets trained on how your business actually runs, which is never how the org chart says. Then somebody reads what it did. A staffing agency does this for people. Here it is done for software, and it does not get skipped because the employee is one.

On our side the human agents are Phil Cowan, Martin Naithani and Jill Bradbury. Martin manages product and Jill manages marketing, and both are on the AI delivery work alongside Phil.

Defines the job before the build

What the seat is for, which systems it may touch, what it may do alone, what waits for sign-off, what done means, what a wrong answer costs. On paper before anything is built.

Makes it fit trained on how you work

The tone of a reply, which customer gets the exception, where the real record lives. None of it is written down anywhere, and all of it goes into the seat.

Reads the log every day

What it did, what it skipped, what it handed over. Drift shows up in the log before it shows up with a customer, and somebody is reading the log.

Answers for it and can stop it

When it is wrong, a person you can call fixes it and owns the fix. The person who enables the agent has the authority to stop it, and uses it.

What we deliver.
What stays yours.

The seat has an owner on both sides, and each side has a job. Ours ends with a working agent and a person who answers for the build. Yours starts on day one and does not end.

InTheory delivers

The job description, in writing. The agent, built inside your tools and trained on how you work. The rules for what it may do alone. A runbook your team can actually run on. The log, and a person reading it. And when the model or a vendor changes underneath it, the plan for that, already written.

Your people stay responsible for

The call on what the job is worth. The named person in your company whose name is on the seat, with the authority to stop it. The decisions the scope reserves for a person: the exception, the refund, the customer who gets a phone call instead of an email. And knowing what a wrong answer costs, because that number is yours and nobody else can set it.

Rule 03 says it for everything we ship, and it holds here: everything ships with an owner, in writing. We publish the rules we work by. Your agent needs rules too.

How the hire starts

A working demo has not
earned access to payroll.

Access is earned the way a new hire earns it: on evidence from real work. A seat starts on drafts, with a person sending. It gets wider access when the log shows it deserves it, and never wider than the job needs. Least-privilege, in your accounts, under your control, so the keys were yours the whole time.

The same goes for the number. Two seats are never the same size, and the review and correction a seat needs are part of what it costs to run, not a line you find later. A price printed on this page would be a bluff, so the price is set in the scope, in writing, before the seat is built.

01
Free Assessment $0 · two days
Tell us the job the way you would tell a hiring manager. We read what you have and where the hours go, and two days later you have a straight answer: is there a seat in this or not. If not, we say so and you keep the answer.
02
AI Opportunity Audit $5,000 · within 5 business days
The job description, in writing: which seat, which systems it touches, what it may do alone, who it reports to on both sides, what done means, what a wrong answer costs, and what the seat costs to build and to run. Signed. Take it anywhere, including to someone else’s build team. And if the right answer is don’t fill this seat with software, the scope says that, too.
03
The seat, if it’s a go set by the scope
The agent is built inside your tools, trained on how you work, and started with a person on it and a person over it. Fixed scope, fixed window, acceptance criteria agreed before it starts. Access widens on the evidence in the log, not on the demo.
Set by the scope means set before the work: the number exists before the seat does, and you hold it in writing either way. No meter.
Off the shelf

There’s an AI employee on sale now.
Setting it up is the job.

OpenAI calls it a dot. It has its own computer, it reads whatever apps you connect, and it keeps working while you sleep.

OpenAI launched dots on September 29: an always-on agent included in the Pro and Business Premium ChatGPT plans, with its own cloud computer and access to the apps its owner connects. The software is ready the day you buy it. Keeping it safe is left to you, and OpenAI’s own docs are plain about how much that is. The rules you write are “instructions your dot tries to follow, and it can make mistakes”. Pausing it stops the current task and “doesn’t stop every delegated task or cancel future scheduled runs”. And you cannot view, edit or delete its individual memories. The only way to clear what it has learned is to delete it.

That is the work this page has been describing, on a dot somebody already bought. One person decides what it is for, which apps it may reach, what it may do alone, what it asks about first, and how to stop it. We do that work in two sizes.

Dot Setup one dot, done properly

One named owner. The fewest apps the job needs, each connected with the least access it allows. The rules written down. Anything that sends, deletes or shares set to ask first. Model training off. A written stop procedure: revoke the apps, pause it, delete it, in that order. A walkthrough with the owner, and a check-in two weeks in.

Dot Audit the one you already have

The dot you already have, checked on six things: who owns it, what it can reach, what it may do alone, what it asks about first, whether training is off, and whether anyone can actually stop it. The findings in writing, within two days.

Want someone reading its log every month? We do that too.

Let’s find out

Every exception
lands somewhere.

Name the person. Then name the person who changes the agent when the job changes.

What is an AI employee?
An AI employee is software assigned a bounded job inside a business, with a named human owner who defines the job and answers for it. The job is written down before anything is built: what comes in, what goes out, what done looks like, what it may do alone and what a wrong answer costs. Innovation Theory builds the agent inside the client’s own tools, and a person on the client’s side keeps the authority to change it or stop it.
What happens when an AI employee makes a mistake?
An AI employee will make mistakes, and the scope says in advance what it may do alone and what waits for a person, so a wrong move is a draft somebody catches, not an action nobody saw. The log shows what happened. The person whose name is on the seat fixes it, and if the job needs to shrink for a while, they shrink it.
Who updates an AI employee when the job changes?
Jobs change. A new product, a new rate card, a vendor that rewrites its API, a model that behaves differently on Tuesday. The seat has a runbook and a named owner on each side, and the change goes through them, in writing, the way the first version did. An agent nobody is allowed to change is one nobody can trust.
Is an AI employee just a chatbot?
No. A chatbot answers when you type. An AI employee holds a bounded job: it watches, acts inside your tools, and writes down what it did.
Can I turn an AI employee off?
The accounts are yours and the access was least-privilege from the start, so stopping is turning it off. The scope and the runbook stay with you.
Does an AI employee replace my staff?
These are the jobs you haven’t filled and the work nobody wants. A seat is the bounded part of a job, not a person, and every seat still reports to one of yours.
What is a ChatGPT dot?
A ChatGPT dot is an always-on agent from OpenAI that runs on its own cloud computer, connects to the apps its owner allows, and keeps working between conversations. It takes direction from one person, its owner. It keeps its own memory, runs scheduled tasks and acts inside the apps it is connected to. As of October 2026 a first dot is included in the Pro and Business Premium ChatGPT plans, per OpenAI’s documentation.
Is a ChatGPT dot safe to connect to my email and files?
Connecting a ChatGPT dot to email and files is as safe as the access it is given and the rules it is set, and OpenAI says those rules are “instructions your dot tries to follow, and it can make mistakes”. Connect the fewest apps the job needs, with the least access each allows, and set anything that sends, deletes or shares to ask first. OpenAI’s built-in review reduces the risk without removing it. As of October 2026, treat every connection as a decision.
How do I turn off a ChatGPT dot?
Turning off a ChatGPT dot takes more than one step, because as of October 2026 there is no single switch. Pause stops the current task, but OpenAI’s docs say it “doesn’t stop every delegated task or cancel future scheduled runs”, so those are stopped separately. Disconnect its apps yourself, since removing its workspace access does not disconnect an app or sign it out of a website. Delete is final. It cannot be undone and does not undo changes already made in connected apps.
Can you set up or audit my ChatGPT dot?
Setting up or auditing a ChatGPT dot is work Innovation Theory does in two sizes. Dot Setup is one dot with one named owner, least-access apps, written rules, anything that sends, deletes or shares set to ask first, training off, a written stop procedure, a walkthrough and a two-week check-in. Dot Audit checks the dot you already have against our six rules and puts the findings in writing within two days. Either one starts with the free assessment.

Some jobs
should stay human.

We can tell you that before we build.

A seat fits when the work is bounded, repeats, has a clear done, and has a wrong answer that is cheap to catch. It does not fit when the work is a judgment call, when a wrong answer costs a relationship, or when the job is simpler as plain software with no model in it at all. The free assessment says which of the three you have. Only the first one gets a seat.

Five seats that have the shape. Each one is written as what it is for, because what it has done in your operation is a record that starts on its first day.

Front desk inbound, first reply, handoff

For the email and the form that arrive at six in the evening. A first reply from your own answers, the record opened, and the thread handed to a person with the context already in it.

Follow-up the second note, the renewal

For the thread that went quiet and the renewal that is sixty days out. The note is drafted while the conversation is still warm. A person still makes the call.

Back office the work between systems

For copy this there, reconcile that, file it, schedule it. Nobody was hired to do it and someone does it all day. The seat does it and logs each move.

Quotes and bids first-pass pricing

For the estimate that sits in the inbox for three days. A first pass from your own rate card and your own past quotes, ready for a person to check and send.

The reading the field, overnight

For the competitors, the filings, the forums and the papers nobody on your team has time to open. A brief in the morning, sources attached, in a form your team reads.