Artificial intelligence tools are multiplying in professional environments, but the measured productivity gains remain highly variable from one employee profile to another. A study published in the Quarterly Journal of Economics in April 2025, covering 5,172 customer support agents, observed an average increase of about 14 to 15% in the number of problems resolved per hour.
Beginner agents benefited the most, while more experienced agents saw a weak effect, or even a slight decline in quality.
Gap Between Perceived Productivity and Measured Productivity in Companies
The first reflex when faced with a new tool is to ask users if they feel more effective. Field data tells a different story: in several experiments, participants significantly overestimated their own gains compared to the objectively measured results.
This discrepancy has direct consequences on investment decisions. The companies most advanced in adopting AI did not produce more per employee than non-adopters. Early adopters even sometimes showed slightly lower productivity.
| Indicator | User Perception | Experimental Measure |
|---|---|---|
| Customer Support (QJE, April 2025) | Positive for all profiles | +14-15% on average, low among experts |
| Software Deliveries (DORA 2024) | Perceived individual productivity up | Decrease in delivery throughput and stability |
This table illustrates a recurring pattern: trust in the tool often exceeds verifiable results. Several recent IT innovations therefore deserve to be evaluated based on their real effects rather than their promises.
AI Assistants for Customer Support and Writing: Real but Targeted Gains
The most solid results concern repetitive and structured tasks. On the Planet X Tech IT site, the categories of tools related to word processing and technical support are among the most consulted, reflecting the current demand from professionals.
The study from the Quarterly Journal of Economics details a point that tool lists overlook: gains are concentrated among the least experienced profiles. For a junior support agent, the AI assistant provides template responses that speed up resolution. For a senior agent, the tool offers suggestions that are sometimes misaligned with the complexity of the case, which slows down processing or degrades quality.
In professional writing, the same phenomenon is observed. A writer who already masters their subject spends more time correcting the AI outputs than writing directly. In contrast, for producing first drafts, internal summaries, or standardized reports, the time savings are tangible.
Criteria for Evaluating an AI Assistant Before Deployment
- The average experience level of the targeted users: the more junior the team, the higher the return on investment according to available data
- The proportion of standardized tasks in the daily workflow, as AI performs well on recurring cases and weakens on exceptions
- The tool’s ability to integrate with existing applications (messaging, CRM, internal database) without creating additional friction
AI Programming Assistants: Speed vs. Delivery Stability
The Google DORA 2024 Accelerate State of DevOps report provides counterintuitive insights. Development teams that used the most AI to generate code saw their delivery throughput and stability decrease, even as individual developers reported feeling more productive.
The DORA 2025 report, focused on AI-assisted development, confirms that trust in generated code remains a problem for a significant portion of the professionals surveyed. The code produced by an assistant passes unit tests but introduces architectural fragilities that manual code reviews detect late.
For teams considering adopting a programming assistant, the relevant question is not “how many more lines of code?” but “how many fewer bugs in production?”. DORA metrics (deployment frequency, lead time to production, change failure rate, time to restore) remain the most reliable indicators to measure real impact.
Workflow Automation: Where the Real IT Gains Are
Automation platforms (connectors between applications, conditional triggers, data routing) represent a less spectacular productivity lever than generative AI, but often more predictable. Their principle is simple: connect existing tools to eliminate manual entries and file transfers.
Automation produces its best results on high-volume, low-variability processes: invoice validation, notification sending, database synchronization between a CRM and accounting tool. Teams that automate these flows recover time on low-value tasks.
What Distinguishes Profitable Automation from a Gadget
An automated workflow that handles ten cases a week does not justify the setup time. In contrast, a process triggered several dozen times a day, with stable rules, generates measurable returns from the first weeks. Data security in transit between applications remains a point of vigilance, especially when sensitive information circulates between tools hosted by different providers.
- Prioritize the most frequent and standardized flows for the first deployment
- Ensure that each connector complies with the company’s security policies, particularly encryption in transit
- Measure the actual time saved after 30 days, not just the theoretical time announced by the vendor
The most publicized IT innovations are not always those that measurably improve productivity. The gap between perception and reality, documented by experimental studies and DORA reports, suggests prioritizing concrete metrics over impressions. The most useful tool is not the one that impresses in demonstration, but the one whose impact withstands rigorous measurement after several weeks of daily use.
