Why I want this work gone

Humans are general creatures. We do a thousand different things a day and we have mastered most of them: walking, cooking, talking a friend through a bad week. For most of our history you had to know a little about everything or you died. A modern career lets you specialize, and for a lot of people that specialization has narrowed to a couple of very specific tasks they have become very good at. That is why they still have the job.

Look at what those tasks are. Entering data. Checking that two spreadsheets match a compliance guide. Loading product into a retail dashboard. Reading reports. Reconciling an engineer’s specs by hand. Until recently, running code by hand. People can get very good at all of these, and machines are now better at every one of them on a computer.

This is not a fringe of the economy. About one in three US jobs is done at a desk. The biggest slice of that, office and administrative support, is the group the Bureau of Labor Statistics now projects to shrink fastest through 2035, and the BLS names AI as a reason.

So I keep coming back to the question of what a person is for. I don’t have a complete answer, but I know what the answer shouldn’t be. It shouldn’t be sitting in front of a screen for forty years doing the same narrow thing. These tasks take very little creativity and very little judgment, and they take up most of a life.

I don’t think preserving office work is a noble cause.

From the recording this post is based on

I used to say a lot of these jobs are pointless. The data doesn’t back that up, and I would rather be right. Only about one in twenty workers say their job is useless when you ask them. What the data shows instead is that the office day is shredded. Microsoft’s own telemetry has knowledge workers interrupted every two minutes, 275 times a day, across 117 emails and 153 chat messages. The shredding is what I want gone.

We have done this before. In 1950 one in thirteen working women was a telephone operator. Today there are fewer than two thousand. The economy absorbed it; the next cohort of women went into clerical and service jobs that had not existed. The operators themselves did worse, and a decade later they were more likely to be in lower-paid work or out of the workforce. I take both halves of that seriously. The labor market adjusts. Individuals need help.

Is this inevitable? I used to say yes. The honest version is narrower. Every major exposure study puts office and administrative work at the top of the list, and the measured effect so far is real but narrow: workers in their early twenties in the most exposed occupations are about 19 percent below where they would be, entirely through reduced hiring, while total employment keeps growing. Not yet in the aggregate. Already in the hiring data. And the capability curve is not slowing.

I also used to argue that a fast transition beats a slow one, because visible change forces policy. The history says speed hurts incumbents either way. Britain’s first industrial half-century saw output per worker rise 46 percent while wages rose 12. The China shock was fast and visible, and the exposed towns still had not recovered a decade later. What changes outcomes is whether help arrives with the shock. So the argument I hold now is this: the replacement of this work is coming from the labs and the software vendors whether or not we take part, and the useful thing a small firm can do is do it honestly, price it fairly, and push for the retraining and support machinery now, while the displacement still shows up in hiring data rather than unemployment data.

What people do with the time is simpler than I once claimed. Every experiment that gives people hours back shows the same thing. They sleep, they look after kids and parents, they see friends, and burnout falls. That is people getting their lives back, and it is reason enough.

What an AI employee actually is

An AI employee is an agent with three things a chatbot lacks. A standing job rather than a one-off prompt. Access to the real tools: the inbox, the books, the CRM. And a clear line between what it may do on its own and what it hands to a person.

The test we use before building one is simple. If you can write the job down as a checklist a new hire could follow on their first day, an AI employee can probably run it. If you cannot, no model will save you.

Three jobs pass that test today. Bookkeeping prep: categorization, chasing documents, reconciling against a rule. Data entry and reconciliation between systems that refuse to talk to each other. First-line email support with lookups into the CRM. A fourth, on-demand report analysis, is the job the models are best at and the one that does not price by the hour, so we sell it per report.

I want to be precise about where the models are not ready, because the honest version sells better to the people we work with. The best models still get one in five multi-variable financial calculations wrong, and the most common error is dropping the word million. On degraded documents, four of six frontier models made up figures. Reliable unattended work tops out around three hours on software tasks, and less on anything else. So an AI employee does not sign a close, does not run for a day without a checkpoint, and does not decide what the client meant. It does the volume. A person owns the judgment.

How one runs

Each client gets its own micro-VM. Not a container on a shared machine: the isolation literature is unanimous that a shared kernel is not a boundary between tenants. Inside the VM, a harness runs one task per session with fresh context, an evaluator with no write tools grades the result before anything leaves, and a task ledger holds checkpoints so a failed step can restart on a backup model. The models are the working tier: Claude Sonnet 5, GPT-5.6 Terra, Gemini 3.8 Flash. The flagships are escalation steps. When a task fails the evaluator twice, or lands on something the playbook marks as hard, a reconciliation that will not balance, a contract clause, a customer dispute, the employee can hand that one step to Opus 5 or Fable 5.1 and then drop back to the working tier. The client chooses how that works: escalation on request, where the employee asks first through an approval card; escalation on its own within a monthly budget; or no escalation at all. Either way, every escalated step shows in the activity log with what it cost.

The security design assumes the model will be fooled. Last year one crafted email pulled a company’s files out through Microsoft Copilot without a click. In April a coding agent found an over-scoped token and deleted a company’s production database and its backups in nine seconds, against explicit instructions. Instructions are advisory. So the agent never sees a credential; a proxy substitutes secrets at the network edge, and only toward allowlisted domains. Scopes are granted read first, write later. Sends, payments, and deletes stop on an approval card. Backups sit outside anything the agent can reach.

Onboarding takes about two weeks and works like onboarding a person. The employee asks for access, the client grants it a scope at a time, and for the first stretch it recommends rather than acts. We have watched this ramp with a retail client already: the system writes recommendations into a spreadsheet, the owner signs off, and the decision on record is automation later, once it has earned trust. Two weeks assumes the client’s IT can grant access in that window. In two of our projects this year the client side took longer than the build, so the free period has a date on it and the access list is written down up front.

The client sees a portal: one conversation, a task list, editable memory, an activity log of every action and every grant, and the approval cards. If they already live in Claude Desktop, the employee is reachable there through a narrow interface: ask, status, approve. Every deployment has a named person at ThirdAxis who is accountable for the outcome.

How it learns

An AI employee is only as good as the job written for it, so most of the work is training, and training is where the business compounds.

Week one is discovery. Every person at the client gets a link to a short AI interview that captures how they actually do the work, what they touch, and where it breaks. We already run this on every engagement; on a recent accounting firm project it produced six role-specific interviews across a nine-person roster in a day. Out of that comes the job description: the checklist, the systems, the rules, and the cases that need a person.

Week two is shadow mode. The employee runs the job against real work and writes down what it would have done. We build an evaluation set from the client’s own material, their chart of accounts, their support policy, their last month of tickets, and the employee has to pass it four times out of four before it runs unattended. A benchmark score says a model can do a task once. Four out of four on the client’s own cases says it can be trusted with it.

Then it goes live with the scopes it has earned, and the evaluator keeps grading. Every task leaves a record: what the job was, what went right, what went wrong, what the client corrected. That record becomes the playbook.

The playbooks are what we keep. Not the client’s data, not their customers, not anything that maps back to a person; that stays theirs, the contract says so in plain language, and there is an opt-out. What we keep is the process: how a month-end reconciliation is actually done in QuickBooks for a trucking client, which document-chasing email gets a reply, which exceptions a bookkeeper wants to see. The next client with the same job gets an employee that arrives knowing it. This is what the labs do with post-training, done in context, one role at a time.

The models underneath are interchangeable and get better every few months. The playbooks and the evaluation sets are ours regardless of which model runs them.

What it costs, and what we charge

The models that come up in most conversations about this, including my own recording, are the wrong tier. GPT-6 Astra and Claude Fable 5.1 cost roughly $15 to $22 per agent-hour at realistic usage. The tier one rung down, Sonnet 5 and Terra, costs $3 to $5 an hour with prompt caching on. Without caching the same work costs about five times more and eats most of a $15 hour. Caching is the margin.

Escalation is how the flagships stay affordable. A single hard step on Fable 5.1 costs about four times what the same step costs on Sonnet 5, so an employee that escalates one task in twenty adds roughly 15 percent to its token bill and still lands well under the price. An employee that runs on the flagship all day does not. That is why escalation is a per-step decision with a budget attached, and why the client can see and set it.

Tokens turn out to be the second-largest cost. The largest is our time. At two hours of supervision per client per month, bookkeeping prep, data entry, and email support run at 50 to 65 percent gross margin, in line with what AI companies report. Double the supervision and bookkeeping falls to 24 percent. That is the number we manage.

On price, I started with billing by the minute under $20 an hour, because that is honest. It still is, but it puts the product next to the wrong comparisons. A loaded US bookkeeper costs $35 to $42 an hour, but small businesses already buy offshore assistants at $3 to $9 an hour and Bench at $199 to $599 a month. So we price a role covered: $499 to $699 a month with a fair-use cap, and the minutes the employee worked printed on the invoice so you can see what you used. The training period is free and you start paying when it works, which is how every ThirdAxis engagement already runs.

Who else is doing this

More people than I thought, and fewer than it looks. Marblism, Sintra, Podium, and Motion all sell “AI employees” to small businesses, self-serve, for $15 to $50 a month. Viktor, which I had written off as a Slack chatbot, raised $75 million from Accel in May and passed 2,000 organizations within three months. The labs already say coworker rather than assistant: OpenAI’s Workspace Agents, Anthropic’s Claude for Small Business, xAI’s Grok Bot. Anthropic bundles QuickBooks, HubSpot, and PayPal workflows into a $20 plan and is running a ten-city small business tour. And Anthropic now sells the architecture above as an API, for eight cents per session-hour plus tokens.

None of them sells a role with a person accountable for it. That is a services business, and nobody funded is doing it for companies with twenty employees. The closest are enterprise vendors with six-figure implementations, and the managed service provider channel, which Pax8 and ConnectWise only started arming in June.

Artisan taught the positioning lesson for everyone. They put “Stop Hiring Humans” on billboards, got the attention, and retired the slogan in August because their customers’ staff would not adopt a tool marketed as their replacement. Our clients have staff, and those staff are the people who grant the agent access and answer its questions in week one. So we sell a role covered, for the roles a small business cannot hire or keep, with the approval gates visible and a person’s name on it.

Why this could work

Five reasons, in the order I weight them.

The economics clear. At the working tier, the three high-volume jobs run inside the margins AI companies actually report, and nothing in the plan depends on prices falling. They are falling anyway: the price of a fixed level of capability has dropped five to ten times a year, so today’s margins widen without a repricing.

The gap is real. The self-serve vendors sell a subscription. The labs sell a tool and a coworker story. The enterprise vendors sell six-figure implementations. Nobody sells a twenty-person company a role with a person accountable for it, and the channel that could, the managed service providers, was handed its first agent catalog three months ago.

We are not starting from a landing page. The discovery interview runs on every engagement today. Our own CRM, New Axis, which we fork and customize per client, mounts a tool server so every endpoint is already something an agent can call with that client’s permissions; the employee plugs into a system we built. A retail client has had an agent connected to live inventory and sales data since August, recommending what to buy and what to mark down, with the owner signing off before anything moves. An accounting firm re-scoped its project with us toward an AI bookkeeping platform for its trucking clients after seeing the demo. A legal services group whose teams already use Claude asked us to build the agent layer under their case software. The shipped work is on the site, and behind it are about forty open conversations across accounting, staffing, construction, hauling, produce, pharmacy, auto sales, distribution, retail, lighting, fiber, legal, and real estate.

Every vertical, on purpose. The AI-employee companies that narrowed picked one job, sales, and their agents got measured against the best human in it. Our clients are defined by workflow, not industry. Every twenty-person company has the same five back-office jobs whatever it sells, and a playbook written for a trucking accountant’s month-end transfers most of the way to a fence distributor’s. Growth is one role at a time: cover the first, earn the second, and each new client starts further along than the last because the playbook already exists.

The buyers are reachable. Small businesses choose vendors by referral; 57 percent found their accountant through a peer and 3 percent through advertising, and 86 percent say reaching a human matters. Our leads come through the discovery interview, through accountants, through clients we have already shipped for, and next through the MSPs that already sit inside hundreds of small businesses. The retention data says the same thing from the other side: AI products under $50 a month kept 23 percent of their revenue last year, products at $250 and up kept 70 percent, and services alongside the software is one of the named reasons.

What could kill it

Supervision creep. If a client needs four hours of our time a month instead of two, the margin halves, and if the free training period never ends, billing never starts. Onboarding is time-boxed and the client’s access obligations are in writing.

An injection incident. One exfiltrated inbox would end the trust story for every client. The controls above exist because of that, and they go in before the first deployment.

Bundling. If our delivery layer is thin, a $20 Claude plan and a half-day training session replaces it. The answer is depth: the role definitions, the evaluation sets, the integrations, and the person on the phone.

Lab cutoff. OpenAI terminated Cursor’s model access in August with 76 days’ notice, at over two billion dollars of revenue. We keep one internal tool contract and thin adapters per provider, so a policy change is an afternoon of work rather than a rebuild.

Framing. Replacement language wins attention and loses adoption. See Artisan.

I am bullish anyway. Very few small business owners can build a secure automation system themselves, and if ours is already running at the price they already pay for the job, there is no reason to switch.

Start with one job

ThirdAxis is a small firm that will sit with you, pick the role, build the employee, and put a name on the outcome. Pick the job your team complains about most, the one that reads like a checklist, and we will show you what it costs to cover it. We should talk.