SI in Healthcare and Biotechnology: 2026 Outlook and Preparing for 2027

For SI in healthcare, the real question is no longer how impressive a model’s answers are, but which problem it solves in the real care process and how reliably it solves it. A radiology system flagging a suspicious finding, an assistant summarising a patient consultation and a model proposing a new molecule require different evidence, different responsibilities and different implementation paths.

The 2026 outlook makes this distinction more important. While large randomised trials are being published in some narrow use cases, broader claims of autonomy are still being tested through controlled research. For organisations, preparing for 2027 starts with choosing the right use case, validating it on local data, defining human and system responsibilities and monitoring the results.

Webtures’ view of this transformation is built on linking technology investment to measurable organisational value. In healthcare, this value emerges when quality of care, workforce capacity, access to reliable information and process traceability are addressed together. This guide offers a decision framework for healthcare organisations, life sciences companies and health technology leaders. It does not contain individual diagnostic or treatment advice.

What is SI in healthcare?

SI in healthcare covers systems that support specific tasks by extracting patterns from inputs such as images, clinical records, laboratory results, biological sequences and operational data. Use cases range from assessing disease risk to clinical documentation, and from matching research participants to capacity planning.

Not every system works in the same way. Predictive models can calculate the probability of a particular event. Computer vision models can classify images or delineate the boundaries of anatomical structures. Generative SI can create text and other content. Agent systems, meanwhile, can call tools, query records and run multi-step workflows within defined permissions.

The presence of a chat interface does not show that the system behind it is suitable for clinical use. Likewise, a process being carried out automatically does not necessarily mean that SI is being used. Rule-based automation can be a clearer and more economical option in explicit, unchanging processes.

What changed in 2026? Reading the evidence correctly

Describing the current picture through a single market size is not enough to explain the nature of the transformation in healthcare. More meaningful indicators are regulatory records, studies conducted in real clinical workflows and evaluations that extend to patient outcomes.

These indicators answer different questions. The device list shows the regulatory product landscape, the mammography trial a specific clinical strategy, the documentation research the effect on workflow and the drug trial the early clinical evaluation of a candidate. Grouping them under a single “SI success rate” is misleading.

The difference between clinical evidence and a technology demonstration

When assessing a healthcare SI project, the first question is the setting in which the result was obtained. A model that achieves high accuracy on historical records may not deliver the same result in a new hospital with different devices and patient groups. An assistant that works correctly in a controlled environment may behave differently with incomplete records or in busy ward conditions.

It is useful to read the evidence in the following order:

• Technical performance: The model’s success on a defined dataset, its error distribution and the limits of its reliability.
• External validation: Whether performance holds across different centres, periods, devices and patient groups.
• Prospective evaluation: Forward-looking testing before the system goes into use or while it is in limited use.
• Workflow impact: The effect on staff behaviour, waiting times, additional review and the correction burden.
• Patient and system outcomes: The impact on safety, quality of care, quality of life, access and total cost.

The same study design may not be appropriate for every application. Nevertheless, the standard of evidence should rise with the decision’s impact on the patient. A system that edits a text draft and a system that influences a treatment decision should not be assessed against the same acceptance criteria.

Diagnostic imaging: one of the most concrete areas of application

Radiology and digital pathology are important use cases for SI thanks to image analysis. Supporting image quality, segmenting structures, flagging suspicious regions and prioritising reviews are different tasks. Success in one does not mean that the others have been validated.

The MASAI trial shows why this distinction matters. In the trial in Sweden, SI was used to route mammograms to single or double reading according to risk level and to show radiologists suspicious findings. Earlier analyses reported an increase in cancer detection alongside a reduction in reading workload.

In the follow-up results that Lund University shared in January 2026, the number of cancers diagnosed in the period between two screening rounds (interval cancers) was reported as 82 in the SI-supported group and 93 in the standard group. The numerical difference is around 12%. On its own, this finding does not mean a reduction in mortality or that the same effect will be achieved in every health system; it should be read together with the trial’s design and statistical evaluation.

When a hospital assesses such a system, it needs to examine its own device fleet, imaging protocols, patient profile, workload and the additional procedures that false-positive results will create. Faster image assessment may not translate into patient benefit to the same degree if there is no capacity in the subsequent diagnostic and treatment steps.

Early warning, clinical decision support and alarm fatigue

In intensive care or ward monitoring, models can generate alerts about a patient’s risk of deterioration. However, the benefit does not depend only on the alert arriving early. Who will see the alert, how quickly they will respond and which protocol they will follow are also decisive.

When false alarms increase, staff trust in the system may decline. A model may catch more at-risk patients while also increasing unnecessary assessments. For this reason, alongside sensitivity and specificity, the number of alarms per patient, positive predictive value, response time and missed events should be monitored.

A threshold that works in one organisation may not suit another centre. When disease prevalence, recording habits and care processes change, re-evaluation is needed. In high-impact use cases such as sepsis, the performance of the software and the effect of the care protocol applied should be analysed separately.

Generative SI and clinical documentation

Ambient documentation tools (ambient SI scribes), which listen to the consultation and create a draft clinical note, are one of the concrete workflow applications of generative SI. The aim is to ease the physician’s documentation burden and make it easier to focus attention on the patient during the consultation.

In the randomised study that UCLA published in 2025, two commercial tools were compared with usual practice. In the Nabla arm, the reduction in documentation time was 9.5% greater than in the control group. The difference in the DAX arm did not reach statistical significance on the same measure. The study was conducted in a single academic health system over a short period.

This result shows that organisations need to evaluate products separately, even within the same category. The study also reported occasional clinically significant inaccuracies. A note that looks fluent and well organised does not mean that information on medication, dose, negation, timing and patient identity is correct.

Implementation design should define how patients are informed that the consultation is being recorded, the data processing terms, the retention period for audio recordings, how the note is verified and who is authorised to transfer it to the electronic record. Time savings should be calculated after deducting the time spent on correction, verification and error management.

Agentic SI in healthcare: workflows with defined permissions

Agentic SI refers to systems that use more than one tool and carry out process steps in line with a defined goal. A workable starting point in healthcare is preparing appointment options, detecting missing paperwork, searching internal policies or creating drafts of actions awaiting approval.

Each agent’s access should be limited to the fields required by the work it will do. Reading a record, writing a draft and executing a final action are separate permissions. Text generated by the agent, or a document it retrieves from an external source, must not be able to change its own access rules.

Safe agent design also covers failure states. There should be controls that prevent the same action from being carried out twice, transaction limits, a stop mechanism and a plan for reverting to manual operation. Human approval should be a control under which the decision can genuinely be assessed; an approval click that becomes automatic amid a heavy workload is not enough.

Patient communication and the digital experience

SI can be useful in appointment preparation, service information, internal routing and presenting explanations approved by the clinical team in an understandable way. However, the boundary between administrative information and personalised medical assessment should also be maintained in the interface.

A patient assistant can explain opening hours; allowing it to generate unlimited diagnostic suggestions in the same interface creates an entirely different risk. Passing topics that cannot be answered to the specialist team, routing people to appropriate services in emergencies and offering accessible alternative communication channels are part of the design.

Success should not be measured only by conversations closing automatically. Reaching the right unit, completing the task, the misrouting rate, repeat contacts and patient experience should be monitored together. Automation that makes it harder to reach human support can lower service quality, even if it appears efficient.

Drug discovery and biotechnology: better candidate selection

SI can help researchers prioritise biological targets, design molecules, predict protein properties and interpret experimental results. This process can help direct limited laboratory capacity towards more promising hypotheses.

Predicting a protein structure does not show that a compound will be safe and effective in humans. A molecule’s synthesisability, biological effect, toxicity, behaviour in the body, manufacturability and clinical outcomes require different tests. The speed at which a model generates new candidates does not mean that the whole drug development timeline shortens by the same proportion.

The rentosertib example is important because it shows SI-assisted target and molecule development moving into clinical evaluation. In the phase 2a trial published in Nature Medicine, 71 participants were assessed over 12 weeks on different dose arms or placebo. The primary endpoint related to treatment-emergent adverse events; measurements of lung function were among the secondary endpoints.

The findings should therefore be read as an early safety and efficacy signal. No conclusion about long-term safety, definitive treatment success or marketing authorisation can be drawn from a small, short study. In corporate investment decisions, experimental validation, development stage and reasons for failure should be monitored alongside the number of candidates.

Patient matching in clinical trials

Identifying patients who are eligible for a trial requires assessing a large number of inclusion and exclusion criteria. Language models can compare these criteria with patient records and provide research teams with prioritisation support.

In the TrialGPT pilot user study reported by NCI, a 42.6% reduction in experts’ screening time was reported. This result does not show that trial participation rates increase to the same degree or that participant diversity is guaranteed. The study evaluated a matching process that supports human experts.

In practice, missing tests, out-of-date records and misinterpreted exclusion criteria should be checked separately. Final eligibility, the invitation to participate and informed consent should be handled in processes managed by the research team. A candidate being ruled out by the model must not turn into an automatic exclusion whose rationale cannot be examined.

Digital twins and alternatives to animal testing

A digital twin is a computational representation of a specific system that is updated with data. Modelling a hospital’s capacity and representing an individual’s biology are not at the same level of scientific maturity. When assessing digital twin claims in healthcare, it is worth asking which variables the model covers and for which decision it has been validated.

The FDA’s April 2026 progress announcement describes work to expand the use of new approach methodologies in drug development. Cell-based systems, human-derived platforms and computational modelling are part of this framework. The FDA also reported that the first SI-based drug development tool had been qualified for a specific regulatory use.

Qualifying a tool for a specific context of use is not the same as approving a drug. The spread of alternative methods also does not show that all animal studies or human clinical trials have disappeared. What is decisive is the validation of the method and its suitability for the scientific question it will be used to answer.

The role of SI in robotic surgery

Robotic surgery and autonomous surgery should be distinguished from each other. A surgical robot may be a platform directed by a human surgeon; the use of a robot does not show that the procedure is carried out independently by SI.

SI can be researched, or used within a product’s authorised scope, in areas such as recognising anatomy, image support, analysing operative records and motion planning for specific tasks. Transferring laboratory demonstrations involving high autonomy into the clinical setting requires additional evidence on safety, unexpected situations and human intervention.

In surgical assessment, it is not correct to use a single general “complication reduction” percentage. The procedure, patient selection, surgeon experience, the method used for comparison and the follow-up period should be specified. Organisations should assess both patient outcomes and training and maintenance costs.

Remote monitoring, biosensors and public health

Wearable devices and remote monitoring systems can allow some measurements to be tracked outside healthcare facilities. SI can assess patterns in this data. However, the accuracy of the measurement, the way the device is used and the capacity to intervene after an alarm determine the outcome.

A sensor producing a risk signal does not automatically reduce hospital admissions. If there is no team to monitor the data and no care plan, the system can create an additional alarm burden. Alternatives should be maintained for patients who cannot access a device or who find digital tools difficult to use.

In antimicrobial resistance and outbreak surveillance, laboratory, clinical and environmental data can be analysed for research and early warning purposes. Data coverage, sampling bias and field validation are important in this area. A signal detected by the model and a confirmed public health event should be reported separately.

How should healthcare SI infrastructure be built?

The model is just one component of the health information system. Sustainable implementation requires data sources, authentication, access control, the knowledge base, workflow and monitoring mechanisms to work together.

Support for standards such as FHIR and DICOM can help integration; on its own, it does not guarantee meaningful data interoperability. Units, codes, dates, patient identity and local workflows also need to align.

The source retrieval approach known as RAG can allow the model to draw on the organisation’s approved documents. However, finding a source does not guarantee that the answer has been interpreted correctly. The document’s currency, access permissions, conflicting instructions and whether the answer is actually grounded in the source should be checked.

Data security, privacy and model updates

Rules for accessing, retaining and reusing health data should be defined before use begins. Whether data processed to deliver care can be used for model training or another purpose is assessed separately. Buying an enterprise licence does not mean that all legal and technical requirements are automatically met.

Data minimisation, role-based access, logging, encryption and retention period management should be addressed together. Removing identifiers does not always ensure anonymity; free text and rare disease records in particular can carry a risk of re-identification.

A change of version by the model provider is also an operational event. A claim that the new version is more powerful may not show that it is safer for the organisation’s tasks. Critical use scenarios should be retested, and there should be a rollback plan for performance degradation or unexpected behaviour.

Regulatory framework: intended use is decisive

Not every piece of health-related software falls into the same regulatory class. The product’s intended use, the decision it affects, its target market and whether it qualifies as a medical device should be assessed. The requirements for an administrative assistant and for software that influences a diagnostic decision may differ.

The FDA’s AI device list brings together specific devices that have marketing authorisation in the United States. The FDA states clearly that the list is not exhaustive. In addition, the clearance, approval and authorisation pathways are not the same. Being on the list does not mean that the product can be used for every task or that it is superior to all alternatives.

The European Commission’s explanation, accessed on 29 September 2026, sets out application dates of 2 December 2027 for high-risk systems and 2 August 2028 for the relevant high-risk SI rules concerning regulated products. This timetable should not be read as a postponement of all SI obligations or of existing medical device and data protection rules. The scope and transitional conditions of each product should be examined separately.

The joint principles of the FDA and EMA on good AI practice in drug development also highlight context of use, risk-based assessment, data governance, multidisciplinary expertise and life cycle management. The readiness of healthcare organisations should rest not only on a deadline but on the need for safe operation.

How should SI investment in healthcare be measured?

The value of the investment is not limited to frequency of use or the number of texts generated. Clinical safety, workforce capacity, patient access and total cost of ownership should be assessed together.

Questions to ask when choosing a vendor

Procurement discussions should start with the intended use, before the product demo. When the problem the organisation wants to solve, its current performance and the acceptable level of error are not defined, the comparison gets stuck on price and marketing claims.

• What is the product’s authorised and intended use, and which tasks does it not cover?
• With which populations, centres, languages and devices was the evidence generated?
• Where is the data processed, who has access to it and is it used for training?
• Can the source of outputs, the model version and user interventions be traced?
• What are the processes for reporting critical errors, for updates and for rollback?
• What is the total cost, including human verification?
• How are data and workflows transferred when the contract ends?

Within the organisation, an assessment by the technology team alone is not enough. Representatives from the clinical team, information security, data protection, quality, procurement and operations should take part in the process, depending on the use case. In systems that make clinical decisions, specialist assessment plays a central role.

A 90-day starting plan

The goal of the first 90 days is not to start using SI in every area but to produce a controlled decision for a specific workflow. Where clinical research or product authorisation processes are required, this timeline may be extended; 90 days is not a guarantee of proving clinical effectiveness.

In shadow mode, system outputs can be assessed without changing care decisions. Stopping criteria should be defined before the pilot begins. It should be clear who will intervene in events such as an incorrect patient record, loss of critical information or an unauthorised action.

At the end of the pilot, the finding that “users liked it” is not sufficient on its own. Safety, net time, cost and the targeted business outcome should be examined together, and negative findings should also be included in the management report.

Readiness priorities for 2027 and beyond

Assessments of 2027 are a forecasting and planning framework; it should not be assumed that all organisations will move at the same pace. The strongest preparation is to build organisational capabilities that can be reused regardless of changing products.

First, an SI inventory should be compiled. Alongside officially procured products, the tools that employees use on their own initiative should also be made visible. The priority for assessment can then be set according to clinical impact and data sensitivity.

The second priority is the organisation’s own evaluation set. Common tasks, rare but critical errors, different languages and patient groups, incomplete records and unexpected inputs should be included in this set. As models change, it should be possible to compare them on the same tasks.

The third priority is workforce design. Employees need to learn not only to use the tool but also to notice errors, to know when not to trust an output and to report problems. It should be planned which care or research activities the freed-up time will be directed to.

Finally, demands for greater autonomy should be assessed together with the available evidence and the capacity to intervene. The capability that can create a competitive advantage in 2027 is less about owning many tools and more about the ability to evaluate and operate reliable applications in a repeatable way.

SI visibility for healthcare brands

Digital trust in healthcare starts with the accuracy of the information an organisation provides on its own website. Specialties, physician information, service scope, facility addresses, contact channels and content update dates should be consistent. The boundary between educational content and personalised advice should be shown clearly.

From the perspective of SI-assisted search and assistants, clear, sourced and up-to-date information is also important. Structured data, an accessible page layout and mutually supporting organisational records can make the information easier to interpret. They do not guarantee being recommended or cited by a model.

From Webtures’ perspective, measuring this area does not rely on visibility counts alone. Which questions the organisation is mentioned in, whether it is represented with incorrect information, how current the sources are and whether users reach the right service information should be addressed together. Clinical success claims should not be turned into marketing messages without expert review and appropriate evidence.

Frequently asked questions

What are the most important uses of SI in healthcare?

Image analysis, clinical decision support, documentation, patient communication, operations planning and drug research are the main areas. Each has a different level of evidence and different implementation conditions; they cannot be assessed with a single general success rate.

Can SI replace doctors?

It can automate or support specific tasks. Medicine involves many tasks, such as clinical assessment, communication with patients, managing uncertainty and responsibility for care. High performance in a single task does not show that the whole profession can be carried out independently.

What does it mean for an SI tool to be on the FDA list?

It shows the record that a specific device has received marketing authorisation in the United States. The authorisation pathway, intended use and product version should be examined separately. The list is not an authorisation for use in every country or a guarantee of superior results in all patients.

Is it appropriate to upload hospital data to a general-purpose chat tool?

Patient data should not be transferred without assessing the organisation’s data policy, the legal basis, the provider contract and the access and retention terms. Deleting some of the identifiers may not, on its own, provide sufficient protection.

Is a molecule designed with SI an approved drug?

No. Molecule design, experimental validation, clinical research and marketing authorisation are different stages. The publication of a phase 2 trial does not mean that marketing authorisation has been obtained or that long-term efficacy has been established.

How is success measured in clinical documentation?

The total time taken to prepare and verify the note, the critical error rate, the need for re-editing, user experience and record quality should be measured together. Tracking only the number of notes generated is not enough.

What should the first investment be for 2027 readiness?

At the outset, the use case inventory, data quality, responsibilities and measurement infrastructure should be addressed. Controlled testing can then take place in a workflow whose impact and risk have been defined. The choice of the first product depends on the organisation’s real bottleneck.

How does Webtures approach this transformation?

Webtures’ approach is shaped around presenting organisational information reliably, assessing SI visibility and designing digital workflows with measurable goals. Clinical validation, medical decisions and regulatory compliance are the domain of the relevant specialist teams.

To assess your SI strategy in healthcare alongside your goals for reliable information, a measurable digital experience and organisational readiness, get in touch with Webtures.

Get in touch →

Sources and methodology

This guide is based on the institutional announcements and published studies of the primary sources below; it is not a systematic literature review or a clinical study conducted by Webtures. The sources were accessed on 29 September 2026; the publication dates of the studies are given separately.

• FDA, list of AI-enabled medical devices: 1,614 entries at the time of access; the most recent decision date shown is 29 June 2026. The list is not exhaustive and is updated periodically.
• Lund University, MASAI follow-up results (30 January 2026); primary article.
• UCLA Health, randomised study of ambient AI scribes (26 November 2025); primary article.
• Nature Medicine, Rentosertib phase 2a trial.
• NCI, TrialGPT pilot user study (18 December 2024); primary article.
• FDA, roadmap to reduce animal testing, year one (20 April 2026).
• FDA and EMA, joint principles on good AI practice in drug development.
• European Commission, AI Act explanations: 2 December 2027 for the relevant high-risk systems, 2 August 2028 for the rules concerning regulated products.