Medha Thapliyal

Sheet 04 AI advisory Revised October 2026

Sheet 04. Advisory.

Most of the AI you are being offered would be better as a form redesign. Some of it is genuinely new.

I help public health organisations tell the two apart, and build the governance and evaluation around the cases that deserve it. The work is advisory and independent. I do not sell tools and I do not take vendor commissions.

A4 · The work

What an engagement looks like.

Engagements are short and written up. Most begin with scoping and end with something a program manager can hand to a board or a funder.

  1. 01

    Readiness and scoping

    Before any tool: which handoff is failing, what data exists, who owns it, and whether a form redesign would do the job. You get a written scoping note you can share with a funder or a board.

  2. 02

    Data governance and consent

    Where data lives, who can see it, how consent is recorded, and what happens when a vendor leaves. Aligned with WHO guidance on large multimodal models and with India's consent-manager approach under ABDM.

  3. 03

    Evaluation design

    Silent-mode pilots, error-review panels, and hybrid effectiveness and implementation designs with RE-AIM outcomes, so a tool has to earn the right to interrupt a health worker.

  4. 04

    Procurement review

    Reading vendor claims against the evidence, checking for lock-in, model-version drift, and offshored data, and writing the exit clause before the entry clause.

  5. 05

    Capacity building

    Short, practical sessions for program teams on what the tools can and cannot do, run with the people who will use them rather than about them.

B4 · Readiness

A six-question readiness check.

Not a score to pass. A way to see which conversation to have first. Nothing you enter leaves your browser.

  1. 01We can name the specific handoff or decision we want to improve, and we have evidence it is failing.
  2. 02The data a tool would need already exists, is reasonably complete, and we know who owns it.
  3. 03We have a consent and data-protection practice our frontline staff can explain to a client.
  4. 04The people who would act on a tool's output were involved in defining the problem.
  5. 05We could run a tool silently for a few months and review its errors with a panel before it changes any decision.
  6. 06If the vendor or model disappeared next year, we know what we would do.

C4 · Archive

Writing behind the advisory work.

  1. Sept 2026 The 0.9 percent Rapid qualitative analysis appears in less than one percent of implementation science studies. AI may finally make it common. The question is what we trade for the speed.
  2. Aug 2026 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?
  3. 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.
  4. 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.
  5. 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.