Medha Thapliyal

Sheet 03 Writing Revised October 2026

Sheet 03/10. 19 August 2026. 3 minute read.

The red alert and the ASHA

A clinical copilot in Nairobi cut diagnostic errors most when it interrupted. What would an interruption mean for a community health worker with no clinical backstop?

03/10 · Essay

Between January and April 2025, fifteen primary care clinics in Nairobi ran a clinical copilot across nearly forty thousand visits. The headline result was a sixteen percent relative reduction in diagnostic errors and thirteen percent in treatment errors. The detail I keep returning to is this: the effect was concentrated in visits where the tool fired a red alert and stopped the clinician mid-flow. Where it only offered quiet suggestions, the gain was smaller. And over the study period, clinicians triggered fewer red alerts. They were learning from being interrupted.

The study was not randomised, and the primary write-up came from the company that built the model. I hold the numbers loosely. But the pattern is consistent with something I have seen without any AI involved: the tools that change behaviour are the ones that make a bad decision harder to complete, not the ones that make a good decision easier to find.

Who absorbs the flag

Now move the same tool a few thousand kilometres east and give it to an Accredited Social Health Activist. India has around a million ASHAs. Each one is responsible for roughly a thousand people, paid by incentive, and frequently the only health presence in a hamlet. Several groups are building decision support for them: a chat-based maternal risk tool in Assam, where maternal mortality remains among the highest in the country, and a general ASHA assistant from a Delhi research group. The Assam study is well designed. It is a multiphase mixed-methods implementation study, which is the right shape, and results are not expected until 2027.

Here is the question I would put to any of these projects. When the tool flags a pregnant woman as high risk, who absorbs the consequence?

In a Nairobi clinic, a red alert lands on a clinician with a prescription pad, a laboratory, and a referral hospital an hour away. For an ASHA, a red alert is an instruction to escort a woman to a facility that may be forty kilometres off, where the staff may not accept the referral, where the woman’s family may not agree to go, and where the ASHA’s own incentive payment depends on the outcome. The flag is the same. The weight is not.

The mechanism is not knowledge

Much of the enthusiasm for frontline decision support assumes that the binding constraint is knowledge. ASHAs do not know enough, so give them a model that knows more. In the programs I have worked on, that is rarely the constraint. The ASHA often knows exactly who is at risk. What she lacks is a referral chain that works, a supervisor who treats a missed case as a system problem rather than a personal failure, and time.

A model that produces more flags into a system that cannot act on them does not improve outcomes. It transfers the moral burden of the gap from the system to the worker holding the phone.

What I would want to see

Before a maternal risk tool goes to ASHAs at scale, I would want three things. A silent phase, where the model runs and its flags are reviewed by a panel but never shown to the worker, so we know how often it is right before it can be wrong. A referral audit, so that every flag the tool would raise is matched against what the referral chain can absorb. And an incentive review, so that the tool does not quietly convert a woman’s refusal into an ASHA’s lost payment.

The Nairobi study is encouraging precisely because it was done in a place where interruptions could be acted on. That is the condition, not the result, that needs exporting.

Sources

03 · Adjoining

Keep reading.

  1. Jul 2026 Novices with a probe One day of training produced sonographer-level gestational age estimates in Lusaka. The hard question is what happens to the referral chain after the machine says high risk.
  2. Jun 2026 Silicon mothers Language models can produce plausible answers about pregnancy in Assam. Plausibility is not validity, and formative research cannot be outsourced to a model that has never missed an antenatal visit.
  3. May 2026 ABHA before algorithms India's bet is that identity, registries, and consent come before models. I think the consent manager, not the model, is the real AI policy.