One Agent for Many Conditions: The Need Is Documented, the Evidence Is Not

The problem

In a Danish register study of 4,651,842 adults, people with six or more long-term conditions averaged 34.1 healthcare contacts in 2017, against 5.8 for people with none² (nationwide register data, observational). They also dealt with 6.9 different providers on average, against 4.0 for people with three conditions, and passed between providers 13.7 times, against 6.6². After adjustment, 20 or more transitions between providers were associated with a higher risk of death than none (hazard ratio 2.80, 95% confidence interval 2.71–2.90), and 20 or more contacts with more potentially inappropriate medication (incidence rate ratio 2.83)². These are associations in observational data and do not prove that fragmentation causes harm. The authors add that the findings may mainly generalise to gatekeeper systems such as Denmark's, where a general practice is the entry point to care².

In plain terms: the more conditions a person has, the more separate parts of the system they must deal with, and the more those encounters go together with worse outcomes.

The problem is also common and unevenly spread. A cross-sectional study of primary-care patients published in 2012 (older than three years) found that 23.2% were multimorbid, meaning living with several conditions at once, and that multimorbidity began 10–15 years earlier in the most deprived areas than in the most affluent¹ (cross-sectional analysis of patient records). It warned that the overall drug burden for patients prescribed treatments from several disease-specific guidelines is "difficult for patients to manage and potentially harmful"¹. The evidence opened covers two studies from high-income countries; no evidence from low- or middle-income countries was opened.

What already works

The coordination role already exists on paper, held by people. NICE guidance says clinicians should assign responsibility for coordination of care and agree an individualised management plan with the person³. It describes a multimorbidity-adapted approach for adults prescribed 15 or more regular medicines and suggests a screening tool such as STOPP/START to identify medicine-related safety concerns³. The 2012 study went further, concluding that "each patient needs a dedicated clinician to take responsibility for care coordination"¹.

The data side is moving. The European Health Data Space Regulation states that people should have free and immediate access to their electronic health data, subject to technical practicability, and may choose to restrict access to some parts of it⁶ (recitals 9 and 17; the operative articles were not read).

The AI building blocks exist but are mostly untested in real care. A 2025 preprint (not peer reviewed) reviewed 20 studies of AI agents in clinical medicine, meaning language-model systems that call tools and carry out multi-step tasks⁴. It found that all agent systems outperformed their baseline language models, with a median improvement of 53 percentage points in single-agent tool-calling studies⁴. Of the 20 studies, one was a randomised controlled trial, 65% relied on synthetic data and all were single-centre⁴. The comparison was against the same models without tools, not against clinicians. In plain terms: tool-using agents beat plain chatbots in controlled set-ups, but the review does not show them working in real care.

One study asks the multimorbidity question directly. A 2025 preprint abstract tested whether a team of language-model agents simulating a multidisciplinary team could give safer therapy recommendations for people with multimorbidity, and reported that a single agent acting as a general practitioner performed as well as the team⁵ (abstract; preprint).

What does not exist. No trial or deployment of an AI system coordinating appointments, medicines, tests and providers across a person's care was found in this research, and no evaluation of such a system with real patients.

The possibility

One way this could work is a coordinating service that a person, or their carer, chooses to use. If it were given access to the person's records from each provider, it could read letters and results, build a single list of conditions, medicines, tests and appointments, and flag conflicts for a named clinician to resolve. It could then prepare the individualised plan that NICE describes, remind the person of what is due, and draft requests to providers, with a human approving any change to treatment.

The role of AI. Only part of this would depend on AI. Reading unstructured letters, spotting conflicts between medicine lists, drafting plans in plain language and holding a conversation with the person are tasks where machine learning and language models could contribute. Record access, consent management, appointment booking, medicine supply, and above all who is accountable and who pays are ordinary software, data plumbing, money and policy. If shared records and a named clinician were delivered without AI, much of the benefit might arrive anyway; AI would then be a small part of the solution, and its added value would have to be shown against that baseline.

What exists and what would have to be invented. Today there are guidance on coordination, patient access rights in law, and tool-using agents in controlled studies. Still to be invented: cross-provider record access that works in practice, evidence of safe agent behaviour in real care, and a funded, accountable owner for the agent's output.

What would have to be true

1. Records flow across providers. If the agent could see only one provider's records, it could not join anything up. The Health Data Space Regulation gives people access rights⁶, but its recitals also allow Member States to provide an absolute right to opt out, under which data would not be made available through the shared services other than to the provider that gave the treatment⁶.
2. A named human stays accountable. NICE places responsibility for coordination with a clinician³. If an agent is to be safe, it would have to support that role rather than replace it.
3. The medicine reasoning is reliable on real records. Real records are incomplete and contradictory. The multi-agent abstract suggests extra complexity did not improve on a single agent⁵, so any claim would have to be tested on real records and against clinicians, not only against other models.
4. It clears regulation. The AI Act's recitals 50 and 51 describe AI in products such as medical devices that undergo third-party conformity assessment as high-risk⁷ (recitals only; operative articles not read). If the agent informed treatment decisions, it might fall under such rules; this research could not settle whether it would.
5. Evidence comes from real care. The agent review calls for prospective, multi-centre trials using real-world data to determine safety and cost-effectiveness⁴.
6. The people who need it can use it. Multimorbidity begins earlier in the most deprived areas¹. If the service relied on a smartphone and confident digital skills, it might reach least those with the greatest need.

What could stop it

What the evidence shows

• Of the 20 studies of AI agents in clinical medicine reviewed in a 2025 preprint, one was a randomised controlled trial, 65% relied on synthetic data and all were single-centre⁴.
• In one 2025 preprint abstract on therapy recommendations for multimorbidity, a single agent performed as well as a simulated multidisciplinary team⁵.
• The Danish evidence on fragmentation is observational and, in the authors' words, may mainly generalise to gatekeeper systems².
• The Health Data Space recitals allow Member States to offer an absolute opt-out, which would limit what any shared-record service could see⁶.
• A dedicated search for failed trials of human care-coordination programmes did not produce a source that could be opened, so no such failure is reported here.

Reasoning

The challenge question is whether the possibility would be less useful than it first appears. If the main harm is fragmentation, the simpler remedies are already written down: a named clinician, an individualised plan and structured medicine review³. If those were delivered with a shared record, an AI layer might add little. The case for AI would then rest on scale, for example if clinician time is too scarce to coordinate every case by hand, and that case has not been tested.

The risk would fall mainly on patients. A wrong reconciliation of medicines, accepted because it appears in a confident summary, could add to the harm the Danish study associates with inappropriate prescribing². If providers are paid per activity, a coordinator that reduced contacts could also work against their revenue, which would leave the question of who pays unanswered.

Signs to watch

• A prospective, multi-centre trial of an AI agent used in real care, with outcomes measured against usual coordination.
• Peer-reviewed publication of the agent review and of the multi-agent therapy study, which are currently preprints.
• The dates on which the Health Data Space patient-access provisions apply in Member States, and which opt-out rules Member States choose.
• Guidance bodies such as NICE naming a role for AI in multimorbidity coordination in updated recommendations.