Window-Aware Scheduling for Health Messaging under Platform Delivery Constraints: WhatsApp Maternal and Infant Health Chatbot in Nigeria
A · Strong evidence HelpMum Africa researchers analyse 98,607 production messages with a regression-discontinuity design, a replication export, bounded offline evaluation and stated limitations; strong evidence on deliveries, from one service.
Summary based on the full text·Original: about 40 minutes to read
In short
An arXiv preprint (2 November 2026) by HelpMum Africa on its WhatsApp maternal health chatbot in Nigeria. WhatsApp delivers free-form business messages only within 24 hours of the user's last message; outside that window only pre-approved templates arrive. Logs of 98,607 messages to 3,696 mothers show free-form failures jump by 93 points at exactly 24 hours. A clinician referral offer, chained after a template, reached almost no one. Sending it as a template would recover most of the loss, and reinforcement learning only rediscovers that rule.
Key numbers
- 93.0 pointsjump in failure rate of free-form messages at exactly 24 hours since the mother's last message (95% CI 87.0 to 99.1)
- 0 of 35,500free-form messages that failed while the WhatsApp window was open
- 0.3 in 100closed-window mothers reached by the chatbot's offer to speak with a medical professional, as built
- 81.5 to 86.8closed-window mothers per 100 estimated to be reached if the offer were sent as an approved template
- US$0.47simulated messaging cost per mother over six months under the template rule, vs US$0.24 as logged
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- No free-form message failed with the window open (0 of 35,500); failure jumped 93.0 points at 24 hours, while templates showed no jump (1.0 point). The replication export shows 0 of 9,546 failures with the window open.
- Two flows caused 66.1% of 21,456 failures: an error follow-up (44.4%) and the medical-professional offer (21.7%).
- Of 4,869 offers, 95.6% failed; the offer could reach only mothers who had written in the previous day.
- Mothers return slowly: median 51 days to the next conversation; 13.6% wrote again within 7 days.
- Offline evaluation for closed-window mothers: template 81.5 to 86.8 reached per 100, waiting 16.6, a reply prompt 11.6.
- Tabular Q-learning in a calibrated simulation converged to the template rule and did not beat it in any of 13 stress tests.
- Opting out did not rise with message volume; messages sent after 22:00 were read within 6 hours 23.0 points less often than morning sends.
- Cost under the template rule: about US$0.47 per mother over six months, or about US$1.43 if the offer were billed as marketing.
How they got it
Two operator log exports with timestamps, flow, message type and delivery status. The window effect is identified by a sharp discontinuity in failure at 24 hours, with placebo cut-offs and a second export analysed with unchanged code. Policies are compared with a model-based offline evaluation that bounds unconfirmed deliveries, then tested in a simulator calibrated to the logs.
Limits
- Exports contain only mothers still in contact, with left-truncated histories selected by an undocumented rule.
- Outcomes are deliveries, reads and replies; nothing shows improved clinical outcomes, or that a template offer leads to clinician contact as often.
- Template results assume a new template is delivered like existing ones; 3% of messages lack a receipt.
- One service, one country, one language; the simulator gives learning little to exploit by design.
Why it matters for companies
Any customer, employee or field-service bot on WhatsApp faces the same 24-hour rule, and messages sent outside it are logged as sent but never arrive. The first audit is cheap: failure rate by flow and by hours since the person last wrote. The paper is also a clean case for encoding known platform rules before reaching for machine learning: once the rule was respected, a learning scheduler added nothing. Check delivery logs before buying smarter scheduling.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org