Blog

AI and IT jobs: what really changes

AI will not erase IT jobs overnight. It changes tasks, exposes roles and forces better decisions on training, control and work design.

Operations console with AI panels and open IT work queues

On a Monday morning, the applications manager opens their inbox and sees three things: a repeat incident no one prioritised, a report request that is “urgent for today”, and a draft produced by an AI tool that only partly works. The problem is not the tool. The problem is that the company still measures activity, not friction.

The useful question is not how many jobs will disappear, but which tasks change first and who decides how they are introduced.

My thesis is simple: in IT, AI changes tasks before it changes jobs. And when management treats exposure as replacement, it diagnoses the wrong problem.

What the ILO does say, and what it does not

In May 2025, the International Labour Organization (ILO) published an updated analysis on how generative AI may affect different occupations. The central point is not alarmist: exposure is calculated at the level of tasks within occupations, not as an automatic forecast of layoffs or a mechanical prediction of future employment. The ILO itself reminds readers that these indicators are early warning signals, not a crystal ball. ILO occupational exposure analysis

There is also an important limit: the ILO is talking about potential exposure. In other words, what could be automated or transformed if the technology were fully implemented. That is not the same as real adoption, immediate wage effects, or job destruction. The ILO also stresses that most jobs will still need human input.

For a business, that distinction matters more than the headline. If you treat “exposure” as “layoffs”, you will probably cut training too early. If you treat it as “task redesign”, you can reshape work before frustration turns into resistance.

Where it shows up first in an IT team

AI usually enters through the least dramatic door. It does not start by replacing an entire function. It starts by saving minutes on repetitive tasks:

  • drafting answers to common incidents;
  • summarising technical documentation and tickets;
  • suggesting tests or edge cases;
  • generating queries or scripts;
  • classifying requests arriving by email, chat or forms.

In an industrial or service company, this does not mean “AI does the job”. It means a person gets to the judgment-heavy part sooner. The difference is huge.

Think of a 12-person internal support team. Leadership sees that average resolution time drops by 18%. That looks like a win. But when the detail is reviewed, an awkward truth appears: the gain is concentrated in simple tickets, while complex cases pile up and escalations increase. The dashboard showed speed; it did not show mental load or quality of routing.

That is where the real friction appears: not between people and machines, but between incentives. If the team is rewarded for closing faster, it optimises closure. If it is rewarded for solving root causes, the kind of work changes.

The story the dashboard does not show

Imagine a distribution SME with a small development and support team. Management wants to “add AI” to IT because it has heard it speeds up coding. The team pilots an assistant for queries and documentation.

Two weeks later the side effect is clear: juniors resolve faster, but they also rely more on reviewing generated output. Seniors receive more validation questions and less time for design. The visible result on the panel is good; the hidden work is worse.

The lesson is not to abandon AI. The lesson is to separate three layers:

  1. Interpretation: AI proposes text, code or classification.
  2. Business rule: what is allowed, what is blocked and where exceptions apply.
  3. Human approval: who reviews before sending, deploying or closing.

Without that separation, the company is not automating. It is shifting effort into invisible validation work.

The uncomfortable decision: train or shrink

This is where the uncomfortable conversation starts. If a task becomes faster with AI, should headcount shrink? My answer is: do not decide from instinct or fashion. Decide after measuring three things for several weeks: workload volume, output quality and the time spent reviewing.

If productivity rises but quality falls, you have gained nothing. If productivity rises but junior staff learn less, you create dependency. And if productivity rises while the remaining work demands more judgment, the problem is no longer technical. It is organisational.

Training is not a soft perk. It is the mechanism that lets a company capture value without hollowing out the craft. That means teaching people how to ask better, review better and approve better. Not just how to “use a tool”.

A small experiment worth running

Before launching a big programme, test one reversible change in a single work queue: for example, repeated low-risk incidents or internal documentation drafts.

Define four things:

  • Input: what type of request enters.
  • AI task: summarise, classify or suggest a response.
  • Rule: what fields it cannot touch and when it must escalate.
  • Measured result: first-response time, rework and number of escalations.

If after two or three weeks neither time nor quality improves, change the task, not the story. And if it does improve, document which part was truly automatable and which part was still human.

That is the sensible starting point for a company that wants to use AI in IT without selling itself a false story of full replacement. At Codefuente, that is usually where we begin: less promise, more real friction, and a small test that can be undone if it does not add value.

Three actions for this week

  1. List five IT tasks and mark which are repetitive, which need judgment and which only look “automatable”.
  2. Choose one small work queue and set a clear human approval rule before using AI.
  3. Ask the team where the invisible rework happens: writing, reviewing, validating or escalating?