An AI rollout questionnaire: what to ask your department heads before talking to a vendor

Short answer: before an AI rollout, each department head needs to answer six questions: what work repeats most often, where the data for it sits, what has to come out at the end, who checks the result, what an error costs and whether that data may be handed to an outside service. From those answers a vendor can give a range, and the company can see which process to take first. The questions are below, ready to forward as they are, along with how to read what comes back.

This came out of an inbound request. A company wanted AI in sales, logistics and accounting, and asked whether we had a questionnaire for department heads. We did not, so we put one together.

We quote in roubles, so the figures below are given as proportions and timelines rather than converted amounts — convert at the rate of the day if you need an absolute number, or ask us for a quote.

Why the department heads and not the director

The decision to spend money is made by the owner or the director, but the people who know where the time goes are the department heads. A director will say “sales spend too long on quotes”. The head of sales will say quotes take long because equivalents have to be looked up across several price lists. You can put AI against the second sentence, not the first.

There is a second reason. Whoever fills in the questionnaire will later sign off the work. If nobody asked them at the start, requirements that were never in the brief will show up at acceptance.

Six questions for each department

Forward them unchanged, one questionnaire per department.

1. What work repeats most often, and how many times a day or a week. One or two concrete operations, not a description of the department.
2. How many person-hours a week it takes now. A rough estimate is fine.
3. Where the data for it sits: the accounting system, the CRM, email, shared spreadsheets, messengers, paper. If it sits in several places, which ones.
4. What comes out at the end: a letter to a customer, a filled-in document, a row in the accounting system, a reply in a chat. Two or three real examples attached are better than a description.
5. Who checks the result before it goes any further, and which kind of error would cost the most.
6. Whether that data may be handed to an outside service or has to stay inside the company.

Plus three general questions, answered once by one person: what the company does, how many people are in each department, and whether there is a deadline for the result.

How to read the answers

The first candidate is the process with many repeats and an easy check. An answer to a wholesale customer about stock and equivalents repeats dozens of times a day, and a manager can see in a second whether it is right. That process pays back first.

Data in one place is cheaper than data in five. If the answer to question 3 is “in the accounting system”, the project is shorter. If it is “in the accounting system, in email and in one employee’s head”, a large part of the budget goes on bringing the data together. More on that in what to prepare before the start.

An expensive error pushes the process down the list. If question 5 comes back with “a tax penalty” or “a wrong shipment worth a year of margin”, the process can still be automated, but later and with a mandatory human check. That is why accounting is usually touched last: the gain is smaller and the cost of a mistake is higher.

“It cannot leave the company” changes the architecture. If the data must not go to an outside service, the language model has to run on your own servers. That is a different estimate and a different timeline, described in a local LLM for B2B.

What usually comes out, department by department

These are observations from typical requests. Every company has its own processes, and the order can turn out differently.

Sales. Catalogue and equivalent lookups, answers to standard customer questions, assembling a quote. Many repeats, the result is checked by a manager, and the data usually sits in the accounting system and the price lists. This is where companies normally start.

Support and service. Answers from rules and manuals. A good fit if the rules are written down. If they live in people’s heads, writing them down comes first.

Logistics. Mostly summaries and reconciliation: shipment statuses, mismatches between documents, letters to carriers. AI helps less here than a plain integration between systems, and that is worth establishing before paying for a model.

Accounting. Recognising incoming documents and reconciliation. It works, with a human check at every step.

What to do with the filled-in questionnaires

Send them to the vendor whole, together with the examples from question 4. From those a vendor can give a range and propose one process for a pilot. The first working version on a single process takes four to eight weeks. How the budget for a sales assistant comes together is broken down in what an AI sales assistant costs.

Common questions

Can AI go into every department at once? Technically yes, in practice it comes out slow and expensive. One process in one department gives a result in a couple of months, and by then it is clear whether to go further.

Who should fill in the questionnaire, the head or the employee? The head, sitting with the employee who does the work every day. The head knows what matters, the employee knows how long it actually takes.

What if a head cannot name a repeated piece of work? Then that department does not need AI right now. That is a useful answer too.

The rule

A questionnaire is good when it shows one concrete operation, its volume and the person who checks the result. A vendor can settle the rest on a call.