Clinical AI enters the workflow: what is changing in medical practice

The useful unit is not a function: it is the complete flow

A consultation does not begin when a recording is activated or end when a note is generated. It begins with background information, documents, questions and previous decisions. It continues with the clinical conversation. It then leaves notes, instructions, orders, references, summaries, and follow-up tasks.

That’s why a tool can seem impressive in a demonstration and still add work in practice. If it forces you to copy information between systems, remember what function to use, or correct results outside the usual flow, the cost of adoption competes with the benefit.

The experience of building Itaca in Latin America points to a simple rule: AI is adopted better when it reduces a transition. The professional should not stop paying attention to “talking to AI”; they should be able to continue with their work and review a useful result at the appropriate time.

Before the consultation: arrive with context, not another inbox

Pre-planning can consume a significant portion of the meeting: locating background information, reviewing documents, remembering the reason for follow-up, or re-asking routine questions. AI can help organize that material, but only if it is presented in a way that the professional can verify it quickly.

The most meaningful uses at this stage are specific: organizing responses from a questionnaire, summarizing documents provided, retrieving the context of previous inquiries, and making pending ones visible. The data Itaca show that sending a questionnaire alone is not enough to incorporate it into the workflow: starting with more context also depends on coordination, the message to the patient, and reminders.

In a clinic, this point also reveals an operational issue: who requests the information, where is it stored, and what actually reaches the professional? Incorporating AI without solving these three questions can digitalize the fragmentation instead of reducing it.

During the consultation: document without turning attention into data capture

Assisted documentation is one of the most visible uses of clinical AI because it connects with a repetitive task in almost every encounter. The predominance of live recording in data sets Itaca places this aid within the realm of care: preparing the draft while the professional is attending and leaving a note ready to be reviewed, rather than reconstructing the consultation from scratch.

The evidence is still growing and should be read with caution. Quality improvement study published in JAMA Network Open He followed 46 professionals from 17 specialties in a U.S. academic system. After incorporating an environmental writer, he observed associations with less time per note, more closings on the same day, less off-hours work, and a perceived lower mental workload. The design was pre-post, without a control group, and the setting was not Latin American; its results show potential, not a transferable guarantee to any clinic.

The practical conclusion is not that every automatically generated note is better. It is that preparing a first draft can take a task of composing from scratch. The benefit depends on three conditions: that the structure corresponds to the type of query, that reviewing is faster than rewriting, and that the professional retains the final approval.

This principle is evident in the AI-assisted clinical notesThe result should remain a revisable draft. The professional edits and decides what is part of the record; the automatic generation does not convert the content into final documentation.

After the consultation: use the context once, don't reconstruct it several times

At the end of the meeting, tasks often appear that repeat information already discussed. The concentration observed in recipes and orders places the use in routine and actionable outputs: to re-use the context that the professional has already verified, rather than re-writing it.

This continuity reduces two frictions. The first is to rewrite data that is already in the note. The second is to lose the relationship between the documented information and what is delivered to the patient. A coherent flow keeps both elements close together and makes it clear what needs revision before sharing.

The same principle applies to the following query. A useful longitudinal record is not a larger file, but a way to revisit treatments, results, decisions, and pending tasks without having to reread the entire history. For the professional, the important question is whether the system returns the correct context when needed.

Documentation is the first pillar, not the only one

Starting with documentation makes sense because the problem is common, visible, and easy to prove. It is also a use where professional supervision fits naturally: a draft is prepared and reviewed before it is incorporated into the file.

However, a clinical platform is incomplete if it only writes notes. Doctors also need to resolve specific questions, review a recommendation, compare sources, and prepare documents related to the encounter. These tasks share context; treating them as separate applications forces the professional to repeat them.

A coherent clinical architecture connects documentation with clinical answers with sources, a deeper investigation when the question requires it and medical documents prepared from the revised context. The practical distinction is important: not every doubt requires an extensive review and none of these layers should delegate the clinical decision-making.

Visible sources: a practical difference compared to generic AI

A clinical response is more useful when it allows one to open up the references and understand where an assertion comes from. This does not eliminate the need to assess quality, current status, and applicability, but it changes the interaction: the professional receives not just a conclusion; they receive a starting point that they can verify before deciding how to use the information.

The WHO guide on multimodal models in health It recommends designing systems for well-defined tasks, involving professionals and other stakeholders, and maintaining evaluation and monitoring throughout the life cycle. OPS It places AI within a broader transformation that includes governance, interoperability, capabilities, and data protection.

That framework helps avoid superficial comparisons between tools. It’s not enough to ask what model a platform uses. It’s more important to ask what task it solves, what information it receives, how it displays the sources, who reviews the output, and how the result is integrated with existing work.

What changes when adoption takes place in a clinic

For an individual doctor, a tool can be useful from the first consultation. For a clinic, the result depends on shared processes as well. Buying access does not equate to changing the operation.

Adoption question	What is useful to define
Where the flow begins	What information arrives before the consultation and who verifies that it is available.
What is being reviewed	What results does AI prepare, which professionals approve them and at what time.
What is shared	How to keep notes, documents, and appointments in order among professionals without duplicating work.
How it is evaluated	Real adoption, review time, tasks completed and issues detected; not just created accounts.

The OPS points out that digital transformation involves cultural changes, In addition to technology. In 2026, the OPS and the IDB also published AI-GUARD, a framework for evaluating strategic value, risks, governance, human oversight, and bias mitigation before implementing AI in healthcare. In practice, a clinic needs to explain what improves, what doesn’t change, who is responsible for each outcome, and what control the professional retains.

For teams, the adoption unit should not be a collection of licenses, but a shared clinical flow. The implementation must define how the context supports the care, what each professional reviews, and how the work ends without duplicating it.

What we can affirm today and what remains to be measured

The available literature allows us to conclude that environmental documentation has the potential to reduce some of the burden of note-taking, and that generative AI can support administrative and informational tasks. It also makes clear that quality, safety, and usefulness depend on the design of the task, the context, and supervision.

It does not allow us to conclude that all professionals will save the same amount of time, that all notes will be more complete, or that the results of a health system will be transferred unchanged to Latin America. It is also not enough to measure how many times a function is opened: to understand value, stable definitions of user, session, task completed, and review are needed.

The internal data of this report allow to describe observed use, not clinical efficacy or time savings. A completed action indicates that the flow reached a valid final state within Itaca; it does not in itself demonstrate that the result was incorporated into the record, improved a decision, or reduced the workload of the professional. Those questions require additional measurements.

How we conduct this analysis

• Period of internal data: From January 1 to September 24, 2026, UTC. September 24 is the last full day included.
• Observed population: External users with at least one valid action in one of five measured flows. The size of the cohort is not published. “Active user” does not mean registered user or subscriber.
• Action steps: Regular clinical responses deep research, documentation sessions with completed notes, completed documents, and shared questionnaires.
• Conversations: a response was classified as follow-up when it was linked to a previous question from the same user. The latency was calculated between the completion of a response and the completion of its follow-up response; negative intervals were excluded.
• Documentation turnaround time: Active days and weeks were calculated among users with at least one completed session. They are not interpreted as retention because each user had a different observation window.
• Exclusions: internal users, demonstration records, discarded, excluded or failed events, and events that did not reach the final state defined for each flow.
• Questionnaires: The seven-and-thirty-day rates use only questionnaires shared with a full thirty-day window. The median time was calculated only between the questionnaires completed within that timeframe.
• Country and specialty: These are fields declared in the profile. The country was available to 99.4 % % of active users. Mexico, Peru, Colombia, and Ecuador accounted for 64.4% % of those who declared a country; this distribution describes the cohort of , Itaca not the Latin American market. The specialty was available only to 28.5%, % so it was not used to compare behaviors across specialties.
• Privacy: The analysis was conducted using aggregations. No clinical content, patient identity, or user identifiers were consulted or published.
• External sources: WHO and OPS guidance on adoption, governance, and digital transformation, in addition to peer-reviewed scientific literature.
• Scope: observational analysis of use within Itaca. It is not a regional survey, a prevalence estimate, or an evaluation of clinical results.

Limitations

Evidence on generative AI in clinical practice changes rapidly, and a significant portion comes from healthcare systems outside of Latin America. The data are biased: Itaca they reflect those who used the platform and completed a flow, not all registered users or all healthcare professionals. Later enrolled users had less time to repeat activity; therefore, the active months figures were not interpreted as retention. Nor were full months compared to September, which only includes data up to the 24th. Flows, regulations, infrastructure, and documentation practices vary between countries and organizations. This report identifies patterns and useful questions; it does not replace the technical, clinical, operational, and regulatory assessment required for each implementation.

Sources consulted

• American Medical Association: 2026 Physician Survey on Augmented Intelligence.
• OPS: Regional study on artificial intelligence and the health workforce.
• OPS and BID: AI-GUARD, an assessment for the responsible deployment of AI.
• JAMA Network Open: Clinician Experiences With Ambient Scribe Technology to Assist With Documentation Burden and Efficiency.
• WHO: guidance on ethics and governance of multimodal health models.
• OPS: Artificial Intelligence for Health.
• OPS: Eight principles for the digital transformation of the healthcare sector.

---

This report is part of the Reports of Itaca over-reliance on AI in medical practice.