Don't compare it to ChatGPT. Compare it to the first hire.

People compare this stuff to ChatGPT.

That is the wrong object.

ChatGPT is a very good intern you have to remember to ask. A small business does not have a ChatGPT problem. It has a Tuesday problem. Invoices still late. Inbox still a pile. Nobody wrote the weekly numbers down. You were going to hire "someone for ops" in Q2. Q2 became Q4.

Compare the teammate to that hire. Then the price starts making sense.

The loaded number, not the salary line

A junior ops person at a small shop is often about $4,000 a month on the offer letter. That is not what they cost.

BLS still puts benefits at roughly 30.1% of compensation. Payroll tax, leave, the stuff that does not show up in the job post. Loaded, that same seat is more like $5,200 to $5,700 a month. Before a laptop. Before the three months where they are learning which vendor is always late.

Octopus is $99 a month. OSS models, a computer, a board. Bring ChatGPT, Claude, or Grok if you already pay for them. We do not mark up the tasks.

Three years of Octopus: $99 times 36 is $3,564.

Three years of that hire: about $180k to $206k, if they stay, if the role was even the right shape.

I am not saying fire people. I am saying most shops were about to hire a person to be a reminder system. That is a bad use of a human. It is a fine use of Clerk.

Three months of ramp versus a job description

A new hire needs a first week, a first month, a first quarter. You know this. You still pretend the calendar is the start date.

A teammate starts when you can write the job the way you would write it for a person.

"Every weekday at 9, pull unpaid invoices over 14 days. Draft the reminder. Stop before send. Bell & Rowe gets the short version. Never CC legal on the first nudge."

That is a lane. Clerk can run it this afternoon. You still tap yes in Needs you. The log is there in the morning.

Microsoft's 2025 Work Trend Index had 82% of leaders expecting to use digital labor to stretch the workforce in the next year or so. Anthropic's 2026 agent survey had 57% of orgs already on multi-stage workflows, and only 16% crossing teams. Small shops are not in that 16%. They are the ones who still have one person doing five jobs, and a tab they feel guilty about.

Do not wait for a cross-functional program. Hire one lane.

The honest list of what $99 cannot do

Write this on a sticky note. Put it next to the board.

Judgment. A teammate can rank, draft, and flag. It cannot decide that this customer is about to churn because of a tone in a hallway. That is you.

Relationships. Mail can sort 31 overnight emails. It cannot take the supplier to lunch and get the lead time you actually need.

Showing up in person. Nobody on the board can walk a job site, read a room, or stand in a warehouse at 6 a.m. If the work requires a body, hire a body.

Taste. Maker can fill a calendar. It does not know if the photo feels cheap. You still have to look.

Also: unsupervised money, unsupervised legal, unsupervised customers. Those stay behind a tap until you say otherwise. Approvals, a vault, its own email, a dedicated workspace. We never train on your data. Seven-day money-back if you put it on real work and hate it. Agencies that want a dedicated VM should talk to us, not buy the $99 and hope.

If a vendor will not say this list out loud, they are selling a miracle. Miracles do not chase invoices.

Start with one recurring lane

The failure mode is "AI for the business." That is a slogan. It produces a graveyard of chats.

Pick the job that already repeats.

Invoices is the cleanest. The rule is written. The exception is named. The cost of forgetting is cash. Clerk does the pull and the draft. You send, or you don't.

Inbox is second. Mail triages. Routine answers wait for you. The supplier that always blows up gets a flag, not a guess.

Monday numbers are third. Scout pulls the list. Watch keeps an eye on the thing that broke last month. Chief routes the pile so you are not the switchboard.

One lane for two weeks. Read the Tuesday. Then add the next.

The computer is already on. The teammate has its own access. Secrets live in the vault, not in a thread you will forget to delete. If the plan you brought runs out of room, stack another account or set a fallback. The job continues mid-task. You do not babysit a reset.

That is the whole comparison. Not "is AI smart." Is this cheaper and faster than the first ops hire for the work that hire would have been bored by.

What it actually costs to keep an agent running.