Why Do Departments React So Differently to the Same AI Rollout?

On this page

1. Which kinds of work AI is closest to
2. Companies and employees are keeping different ledgers
3. What gets assessed is quantity, which may not capture the result you want
4. Run one role through first, then talk about everyone
5. Management Advice

In March 2025 the company did something: it built a locally deployed DeepSeek into the office software everyone opened every day, opening it to employees as a conversational AI. Except for field sales and frontline production staff, every other department was uniformly required to ask at least 200 questions per person per month, counted toward their KPIs.

The move looked solid enough at the time. The tool was built into the software employees used daily, training was done, requirements were issued, and it was tied to assessment.

After it had run for a while, the picture I saw became more and more awkward.

Product and software development departments were relatively positive, because AI could directly help with the work they already did. Finance, administration, production management, and quality looked more like they treated the uniform requirement as an extra burden, using it to fulfill the KPI.

The same thing, the same requirement, landing in different departments produced results so different they barely seemed like the same initiative.

These differences came from observations in several places: direct conversations with employees, feedback in management meetings, actual usage behavior, and real question counts and completion. Behind them all was one common point: the conversational AI of that time had not been well connected to the internal data, business systems such as ERP, and formal workflows these departments commonly used.

Looking back now, the reactions of different departments and roles cannot be attributed only to whether employees were willing to use it. What mattered more was how far AI was from their core work.

Which kinds of work AI is closest to

Laid out plainly, there were three distances between different departments and AI.

The closest was product and software development. The things these roles dealt with daily were already code, requirements, software, and various digital materials — closer to what the conversational AI of the time could process directly. Precisely for that reason, they more easily felt help within their existing tasks.

The middle distance was research, marketing, and compliance. A large part of their work was finding external material, organizing, researching, and writing documents. The conversational AI could help with some of it, so their reaction was fairly calm — neither especially excited nor especially resistant.

The farthest was finance, administration, production management, and quality. What they actually had to handle mostly sat inside internal systems: business data, ERP, permissions, and formal documents and processes that had to be completed one by one. The conversational AI of the time stood inside a chat box in the office software and could not reach these. Employees could use it to write some generic text, but on much of their core work it could not get a hand in. The 200 questions a month turned from helping people work into extra work.

So attitudes toward AI diverged markedly across departments and roles. And that divergence came largely from the distance between AI and the objects of their core work.

The three distances, in one view:

• Closest — product and software development: the work objects are code, requirements, and digital material.
• Middle — research, marketing, compliance: finding, organizing, researching, and writing documents.
• Farthest — finance, administration, production management, quality: business data, ERP, permissions, and formal documents and processes inside internal systems.

Companies and employees are keeping different ledgers

In managing, a manager must consider human nature. Here that has a more specific meaning: accept that the unit of value calculation for the organization and for the individual is different to begin with.

The company counts the overall account: whether investing in this tool cuts cost and raises efficiency, and whether company-wide usage rises.

The employee counts their own share: is it reducing my workload, rework, and hassle, or is it adding learning cost, switching cost, and a layer of assessment pressure?

The company sees an AI usage rate; the employee calculates whether their own work has gotten lighter because of it.

This is not small-mindedness. Everyone has only so many hours a day, and anyone first reckons the account most directly relevant to themselves. The trouble is that managers easily assume that once the tool is installed, training done, and requirements issued, the employee’s account settles itself. The AI tool did go into the office software, but in employees’ actual work they had not yet felt any bit lighter.

What gets assessed is quantity, which may not capture the result you want

In March 2025, question counts still had to be tallied by employees themselves. By April, the office software was upgraded and the system could automatically count how many times each person asked.

From a management view this was progress: the count was more accurate, and employees had less to do.

But it improved the ability to measure, not the ability to judge value. The system can tell you how many times a person asked this month; it cannot tell you whether the thing in their hands got faster, better, or easier because of it.

A count measures an action; adoption needs value. Between those two lies exactly the wide variation in each role’s work objects, systems, and processes. The same usage count, in some departments, is a natural result of convenient use; in others it can only be forced out by assessment. If you take a uniform usage volume as uniform effect, what you measure is probably just the order you issued.

Run one role through first, then talk about everyone

If I went back to that situation, the more reasonable choice this time would not be to roll AI out to everyone from the start, but to pick one role and one real task and run it through first.

Concretely, by the role and the actual needs of that task, see at least four things clearly:

1. Can AI get the real data this task needs;
2. Can it get into the system employees actually use;
3. Can it take part in the real process, or does it only stand outside the system writing some text;
4. Can it reach the accuracy the delivery requires, how should permissions be controlled, who judges the result, and who bears responsibility when something goes wrong.

Put plainly, have employees run one complete piece of work themselves and see exactly which part AI helped with and which part is still missing. Without running that step through, there is no need to rush to measure adoption with a simple indicator like usage volume.

After it runs through, handle different departments differently. Where AI can enter core tasks directly, push around real tasks and let it do more of the actual work; where it can only help with peripheral and generic tasks, state its boundaries honestly and don’t set expectations at maximum; where it still cannot reach core data and systems, admit it has not arrived yet, and don’t use a pretty question count to pretend adoption has happened.

Someone will say that was a 2025 matter, and today’s agents are long past that.

That is only half right. Today’s agents have greatly lowered the barrier to generic office tasks like weekly reports, organizing material, and slide decks; employees pick them up faster and more easily feel direct help. But as long as the work objects are still in internal data, permissions, and formal processes, and as long as someone still has to sign for and bear the delivered result, how AI connects to that work still has to be solved role by role.

An easier-to-use tool does not mean it has entered the core work of the role.

Management Advice

For business leaders

Do not record the AI effort only as procurement and usage rates. Before asking whether employees have started using it, consider first whether the work in their hands has actually gotten lighter.

For managers

If those above look only at usage volume, you can propose a small trial run. Pick one real piece of work, run it to completion, and report both what AI helped with and what it did not. One trial run cannot explain everything, but it is closer to real value than simply reporting a usage count.

Next time, we will use management thinking to crack open another AI question.