What to do before buying a robot for your business
Most bad robot pilots fail before the robot arrives. The job was never written down, nobody measured how it ran with humans, and the machine showed up with nothing to plug into. This is the pre-flight: four steps in order, each cheaper than the next, so that by the time a humanoid is on your floor the only open question is whether it earns its rent.
Step 1: Write the SOP
If the job lives in one person's head, it is not a robot job yet. It is not even a good human job, because it disappears when that person is on vacation. Write it down: the trigger, the steps, the exceptions, what "done" looks like, what to do when something is wrong. One page. Plain language. Have the person who does the job read it and mark everything you got wrong.
This step costs nothing and it is the one most owners skip, because it feels like paperwork. It is not. The SOP is the spec. Every vendor site assessment, every integration conversation, and every AI employee you set up will run off this document. If you cannot write it, you have learned something important about the job before spending a dollar.
Step 2: Run the AI employee first
A humanoid is a body. The brain is an AI employee: a software role, with rules you can read, running inside your company today. For receiving, that is the AI that reads purchase orders and reconciles counts. For inventory, the cycle-count AI that owns the schedule and the variance report. For security walks, the AI that reads the feeds and dispatches a human for anything real. None of those need legs. All of them run off the SOP you just wrote.
Why this comes before the robot, and not alongside it:
- It pays for itself alone. The AI employee removes management overhead, catches exceptions and keeps the schedule whether or not a robot ever shows up. It is a good 2026 investment on its own terms.
- It is the integration. When the vendor asks "what system tells the robot what to pick," the AI employee is the answer. Without it, the integration line in your pilot budget becomes "build the system," which is the most common place a pilot stalls.
- It de-risks the robot. If the AI employee cannot run the job from the SOP, the SOP is wrong or the job has too much variety. Better to find that out in software than with a $2,000 to $4,000 a month rental (per the vendor, Agility Digit's Robots-as-a-Service band) standing idle.
The owners buying robots first in 2026 are the ones who already run AI employees. The ones who skip that step pay for a body that does not know their business.
Step 3: Measure the baseline
You cannot tell whether the robot is winning if you do not know the score before it starts. For at least a month, with the SOP in place and the AI employee running, track the job as humans do it: units per shift, error rate, hours including supervision, stalls and how long they took to clear. Keep it simple enough that someone actually fills it in.
Two things happen during this month. First, the job usually gets better on its own, because writing it down and running the AI employee fixed problems nobody had named. Second, you now have the number the robot has to beat. Without that number, the pilot decision at the end is a feeling. With it, the decision is arithmetic.
This is also when you do the loaded-cost math for the robot side: rental, integration, safety perimeter, downtime, and the person who will manage it. The full breakdown is in how much it costs to rent a robot worker.
Step 4: Pilot the robot, on a rental
Now talk to a vendor. You arrive with an SOP, a running AI employee, a baseline and a budget. That is a different conversation from "we are curious about robots." The vendor can scope a cell, write an SLA against your numbers, and tell you honestly whether your job is a fit.
Rent, do not buy. Per the vendor, 1X NEO lists at $20,000 or $499 a month, Figure 03 is targeting $20,000 to $25,000 for pilot customers, and Tesla Optimus is on 2027 consumer guidance. Whatever the sticker, a 2026 purchase puts hardware risk on you in a market where next year's unit will be better. A rental puts it on the vendor. The mechanics of a RaaS pilot, from site assessment to SLA to the decision at the end, are in Robots-as-a-Service, explained.
Run the pilot for a window comparable to your baseline. Judge it on uptime, throughput and error rate against the number from step 3. Not on how it looks when a visitor walks through.
The three disqualifiers
Before any of the above, run the job through these. Any one of them makes it a bad first pilot, no matter what the vendor says or how good the demo looks.
| Disqualifier | What it looks like | Why it kills a first pilot |
|---|---|---|
| No written SOP | "Ask Maria, she knows how it goes" | There is no spec to integrate against and no way to know what the robot got wrong |
| No work cell | The space, the objects or the sequence change day to day | Humanoids in 2026 are reliable inside a defined box and unreliable outside one |
| Public in the space | Customers, guests, patients or crowds share the floor | Safety, liability and variety all spike at once. Front desk cover and event setup are in the wait band for this reason |
Notice that the first disqualifier is fixable in a week (write the SOP), the second is sometimes fixable (redesign the space into a cell), and the third is not fixable in 2026. That order matters when you are choosing between candidate jobs.
Picking the job in the first place
If you have several candidates, prefer the one that is dull, in a box, and expensive to get wrong. Receiving, picking, inventory counts, security walks and lab sample handling are the five pilot-band jobs on the chart for exactly those reasons. A $1-50M business owner on an Optimus mastermind call said the job he regretted not starting with was the one his team complained about most and nobody wanted to own. That is usually the right one.
The chart pairs each of the twelve jobs with its band and with the AI employee that has to exist first, so step 2 is a lookup rather than a guess.
FAQ
What should I do before buying a robot for my business?
Four steps in order: write the SOP for the job so it exists outside someone's head, stand up the AI employee that will supervise the job, measure the baseline (throughput, errors, hours) for at least a month, then pilot a robot on a monthly rental against that baseline. Do not buy hardware until the pilot has produced your own uptime data.
What is an AI employee and why does it come before the robot?
An AI employee is a software role with rules you can read, running inside your company today: a receiving AI that reconciles counts, a security AI that reads feeds and dispatches a human. A humanoid is a body; the AI employee is the brain it plugs into. It costs a fraction of a rental, pays for itself on its own, and the robot does not know your business without it.
Which jobs make a bad first robot pilot?
Three disqualifiers: the job has no written SOP (it lives in one person's head), the job has no defined work cell (the space, objects or sequence change), or the job puts the robot in the same space as customers or the public. Any one of those makes it a bad first pilot, whatever the vendor says.
How long should a robot pilot run before deciding?
Long enough to compare against a real baseline. If you measured the human-run job for a month, run the robot for at least a comparable window after commissioning, and judge it on uptime, throughput and error rate, not on how impressive it looks on the floor.