What the research found
Williams and colleagues reported longer-term mental-health outcomes from eIMPACT in the Journal of Affective Disorders. The analysis included 216 primary-care patients with depression and elevated cardiovascular risk. Participants received usual care or a twelve-month collaborative-care intervention incorporating internet or telephone cognitive behavioral therapy and/or selected antidepressant medicines. Improvements in several mental-health measures were greatest at twelve months and attenuated by twenty-four months. The article appeared online in 2025 and in the January 2026 issue. [1]
What the design can establish
This was a secondary analysis of a trial conducted in a safety-net system during 2015–2020. It evaluates a blended clinical service, not the independent effect of an app, an AI system or a particular channel. Attenuation during follow-up raises a maintenance question; it does not prove that a specific extension would preserve the benefit. The sample and setting also limit broad claims about other populations. [1]
My operator interpretation
I would ask a behavioral-health business to explain what happens after the funded treatment episode ends. A person may improve, retain residual symptoms, struggle to sustain progress or need a different level of care. Each situation calls for a defined route that the clinical team can adapt. An administrative discharge date should not be the organization's only answer.
The capacity model should therefore include initial assessments, active follow-up, clinical review, maintenance contacts and re-entry. I would make those workloads visible separately. Counting enrolled lives or completed sessions alone can conceal an accumulating group of people who are no longer improving and have no clear next step.
For an integrated primary-care program, the operating responsibility is to make nonresponse visible and bring it to someone authorized to change treatment. A registry can support that work, but its usefulness depends on reliable information, professional judgment and available treatment. Digital convenience should expand the ways a person can receive care without implying that software itself provides the full clinical service.
The commercial agreement should reflect this continuity. I would compare contracts that fund only initial engagement with those that support appropriate follow-up and renewed access. Any claim that maintenance lowers medical spending would require separate evaluation. Sustained symptom improvement, function and patient experience are already meaningful objectives and should have their own measures.
The decision I would make
I would fund a maintenance design as a testable clinical operating hypothesis, with qualified oversight, clear response responsibilities and a comparator where feasible. Expansion would depend on sustained outcomes and equitable access to follow-up, alongside workload and cost. The eIMPACT findings motivate this question; they do not answer the local service-design choice.
What this evidence cannot settle
Recent publication should not be confused with recent data collection. The operating interpretation is not a tested maintenance protocol or a claim about autonomous AI treatment.