A COO at a 70-person spare-parts distributor called us in spring about an AI project. He wanted a model reading every inbound order email, checking stock, and confirming delivery dates without human touch. Twenty minutes into the call I asked where the orders lived. In the personal Outlook folders of three sales reps, plus WhatsApp photos from two large customers. I asked who decided delivery promises when stock ran short. A dispatcher named Frank, from memory, in his head. The call ended with no AI quote and a two-day mapping job instead. Six weeks later the distributor ran the same flow on plain rules against a shared mailbox, and the AI budget line went back into the business.
That call was not unusual. It was the normal one. So we turned the questions from it into a fixed check we now run before quoting any AI work: four areas, scored 0 to 2, done in about 30 minutes. The check below is yours to run before you talk to any vendor, including us.
How the 30 minutes work
Book a room with the person who does the target process today. Add whoever owns your IT or data, even if that means the external contractor on speakerphone. And bring yourself. Pick one process, the painful one you would give to AI first. Set a timer. Each of the four checks below ends with a score. Zero means the foundation is missing, one means it exists but needs work, two means it is solid. Add them up for a total out of eight. What each total means comes at the end.
Answer from evidence, not optimism. If someone says the data is "basically in the ERP," ask them to open the ERP and show you. The whole value of the exercise is the gap between what management believes and what the screen shows.
Check one: can software reach your data?
A model is only as useful as the data it can read and write. Forget quality for now. Nobody cares yet whether the data is clean. The question is access: can a program get the inputs without a human copying them, and can it write results back somewhere the business trusts?
The distributor failed here on day one. Orders arrived as free text in personal inboxes nobody else could open, and stock lived in a spreadsheet on Frank's laptop. Customer-specific pricing lived in notebooks. No model fixes that. The fix was unglamorous: one shared mailbox, a nightly export from the ERP into PostgreSQL, and pricing moved into a table. Two days of plumbing, and suddenly the process was automatable, with or without a model.
Score it this way. Zero if key inputs live in heads, paper, personal inboxes, or chat threads no system can read. One if the data is digital but scattered across four systems with no export routine, which covers most 50-person companies we meet and is an honest starting point. Two if one system holds the truth and a program can read it today through an export or an API.
Most companies land on one, and one is enough to start. Zero means your first project is data plumbing, not AI, and any vendor quoting a model before the plumbing is selling you the second floor without stairs.
Check two: is the workflow written down?
Ask the person doing the job to describe the process while you write. Get every branch and every exception on paper. Then ask about the silent "it depends." Then apply the new-hire test: could someone competent run this from your notes on their first day? If not, the process is not documented. It is folklore.
Folklore kills AI projects in week three. The model gets built against the happy path, then real traffic arrives with all the exceptions nobody mentioned, and accuracy collapses. We watched this happen with a returns process at a mail-order company: the documented flow had six steps, the real one had nineteen, and eleven of them lived in one team lead's head. Mapping the real flow took two days of sitting with her. The build that followed took three weeks and worked, because the specification was finally true.
Score it this way. Zero if the process has never been written down and varies by who handles it. One if a rough description exists but the exceptions are missing or outdated. Two if the notes pass the new-hire test, including what happens when things go wrong.
This check doubles as the cheapest project accelerator available. A day or two of mapping with the person doing the work produces the specification and the test set for whatever you build next. Skip it and you pay for the same discovery inside the AI project at triple the rate.
Check three: what does a wrong answer cost?
Models guess. Good ones guess well, but every deployment needs an answer to the same question: who catches the wrong guess, and what does it cost before they catch it?
Sort your process steps into two piles. In the first pile are reversible annoyances: a misfiled inquiry, a support ticket in the wrong queue, a draft that needs one more edit. A model step is acceptable there, provided a person or a rule verifies the output before it matters. In the second pile are wrong payments, wrong delivery promises to key accounts, and wrong entries in records a regulator reads later. Those need deterministic behavior and an audit trail naming the exact condition that fired. We keep those steps in code even when a model could do them faster, and we say so in writing during discovery. Clients occasionally push back. They stop pushing once we ask who signs off on the first wrong payment.
Score it this way. Zero if errors reach customers or money with no checkpoint, and nobody can say where a checkpoint would even sit. One if a review step exists in principle but nobody staffs it, or the reviewer just approves everything, which is the same as no review. Two if a named person checks flagged output before it counts, with a queue they actually clear daily.
If you scored zero here, your project starts with the review queue, not the model. Fifteen minutes a day on a review screen is the control that makes every later AI step safe to switch on. We describe the full boundary between guessing and rules in our piece on workflow automation without AI: variance that rules cannot capture is the only thing a model is strictly better at, so buy the model only for that variance.
Check four: who owns the system after launch?
This is the check vendors skip, because the answer threatens the sale. A model step needs weekly care for as long as it runs: someone clears the review queue, keeps the test set of real inputs current, tunes the prompt when accuracy slips, and re-validates behavior after every provider upgrade, because upgrades change behavior without asking. Plain automation needs none of this. It sits still until someone edits it.
We inherited a chatbot pilot at a wholesaler that had no owner. The vendor shipped it, trained nobody, and left. Six weeks later the answers had drifted, the review queue had 400 unread items, and staff had quietly gone back to phone and email. The technology worked. The ownership did not exist. The relaunch cost more than the original build, and the only structural change was naming one operations employee as owner with four hours a week reserved.
Score it this way. Zero if no one has time reserved and maintenance would fall to "whoever notices." One if an owner exists on paper but has no weekly hours or no authority to change the system. Two if one named person holds both the hours and the authority, with a deputy for holidays.
Be strict here. A one is a zero wearing a nicer suit. If you cannot name the owner and their hours, you are not ready for a model. You are ready for plain automation, which runs without an owner the way a staircase needs no pilot.
What your total says
Totals of 0 to 3 mean fix the foundations first. Move the data somewhere reachable and write the process down. Install the review queue. That work usually takes weeks, costs less than any AI pilot, and returns hours on its own. Many companies stop here happily with a finished deterministic system and never need a model. Our implementation playbook for smaller firms describes exactly this sequence: one painful workflow, the boring 80 percent automated first, measured hours back.
Totals of 4 to 6 mean run one supervised pilot. Pick the single step with genuine input variance, put a model on that step only, and keep the human checkpoint behind it. Measure error rate and review minutes per day for a month before expanding. This is the shape most of our AI automation delivery takes: a deterministic flow with one model step inside.
Totals of 7 to 8 mean you are ready to scale. The data is reachable, the process is written, errors get caught, someone owns the system. Spend your budget on the second and third workflow, not on re-proving the first.
One warning about high scores: a perfect eight on a process nobody measured is suspicious. If you cannot state current error rate and hours spent, knock a point off checks two and three and start measuring. Models amplify unmeasured processes in both directions.
If you ran the check and the result surprised you, that is the normal outcome, and it is a good discovery conversation to have. Ours runs from Berlin with senior engineers behind it. Send us your score and we will tell you honestly whether your next step is plumbing, mapping, a pilot, or nothing at all.
Frequently asked questions
How long does an AI readiness assessment take for a small company?
The check in this article takes about 30 minutes with three people in the room: the COO, the person who does the work today, and whoever owns IT or data. It answers whether you should pilot AI now or fix foundations first. A full vendor assessment with system audits takes weeks, and you only need one after this check says you are ready.
What is the most common reason companies fail AI readiness?
Data nobody can reach. Orders live in personal inboxes, stock counts live in spreadsheets on one laptop, and customer history lives in the heads of two senior staff. A model cannot work on data it cannot read, so access beats quality at this stage: scattered digital data you can export is enough to start, locked-away data is not.
Can we run an AI pilot if our processes are not documented?
You can, but you will document the process during the pilot at higher cost and lower accuracy. Writing the workflow down first takes a day or two with the person doing the job and gives every later step a specification and a test set drawn from real work. Undocumented process is the reason most pilots stall in week three.
Who should own AI inside a mid-size company?
One named person with weekly hours reserved for it, not a committee and not an external vendor alone. The owner clears the review queue, keeps the test set current, and signs off on model or prompt changes. Without that owner, run plain automation with near-zero maintenance instead of a model that needs weekly care.
