In perioperative medicine, artificial intelligence is often discussed in terms of innovation, predictive performance and future potential. But in daily clinical practice, an equally important question is what happens when these tools actually enter the operating room. How are they used? How quickly do clinicians become familiar with them? And does their value change over time as confidence grows?

These were the questions behind our study on Assisted Fluid Management (AFM), an artificial intelligence–based decision support system designed to guide intraoperative fluid challenges using real-time stroke volume monitoring. We were interested not only in the technology itself, but in its implementation in routine care.

This question felt especially relevant in major abdominal oncologic surgery, where fluid therapy is a continuous balancing act. Too little fluid may compromise tissue perfusion, while too much may contribute to postoperative complications. In this setting, clinicians are constantly making time-sensitive decisions based on physiology, surgical conditions and patient-specific factors. A decision-support system may help structure these choices, but only if it is integrated into real clinical workflow in a meaningful way.

Our study was a retrospective observational analysis conducted in a high-volume tertiary referral center for abdominal oncologic surgery. We included adult patients monitored with AFM between February 2024 and March 2025 and compared two consecutive implementation periods. Rather than focusing on postoperative outcomes, we chose to examine the implementation process itself. We analyzed each fluid challenge at the bolus level, looking at whether it was initiated by the clinician or suggested by AFM, and whether it generated an effective stroke volume response.

What we found was, for us, the most interesting part of the story. Over time, clinician-initiated boluses decreased, while AFM-suggested fluid challenges increased. At the same time, the physiological effectiveness of AFM-suggested boluses improved, as did the overall effectiveness of fluid challenges. In total, the study included 59 patients and 404 fluid challenges, and the second implementation period showed a pattern consistent with increasing familiarity, confidence and more targeted use of the system.

This is why we believe the “learning curve” is central to the discussion around AI in anesthesiology. A tool like AFM is not simply switched on and instantly absorbed into practice. Its role evolves. Clinicians learn when to trust its suggestions, how to interpret them in context, and how to integrate them into their own judgment. In this sense, implementation is not a side issue; it is part of the technology’s real-world performance.

For us, this study also speaks to a broader issue in healthcare innovation. AI systems should not be evaluated only under ideal or early conditions. They should also be studied as they are adopted in practice, because maturity of use may shape results. Early implementation may reflect caution, novelty or incomplete integration, whereas later phases may better capture how a system truly functions in routine care.

At the same time, our findings should be interpreted within the limits of the study design. This was a retrospective, single-center observational analysis, and it was not designed to determine whether these implementation-related changes improve postoperative outcomes. That remains the next key question.

Still, what this work offers is a practical perspective on how AI enters clinical care: not as an abrupt transformation, but as a gradual process of adaptation between clinicians, workflow and technology. In perioperative medicine, that process may matter just as much as the algorithm itself.