Collections
How collections AI actually reduces PAR
Predictive collections only work when scores sit on top of a clean repayment ledger and a human workflow. Here is the operating model.
Efua Darko, Collections Scientist · 18 June 2026 · 7 min read
Portfolio at risk is a lagging number. By the time PAR 30 prints on a dashboard, the customer has already missed a cycle and a competitor may already have their next MoMo. Collections AI is useful only when it moves work earlier: who is likely to miss, who should receive which channel, and which officer should call today.
Features that actually predict
The strongest features in Ghanaian books are behavioural, not bureau-only: days since last deposit, declining susu intensity, repeated short-pays, changed phone numbers, and missed promises-to-pay. Combine them with product type and instalment size. Keep the model explainable so supervisors and credit committees can challenge it.
Orchestrate channels
A reminder SMS the morning an instalment is due is cheaper than a field visit. A USSD self-pay link is cheaper still if the customer has float. Reserve officer time for high-balance, high-probability accounts. OrbitMFI's Collections AI ranks the queue, logs every contact, and stops messaging once a receipt posts.
Measure the right thing
Do not celebrate model AUC in isolation. Track roll rates, promise-kept ratio, cost-to-collect, and complaints. If SMS volume rises while recoveries stay flat, the model is nagging the wrong people.
Worklists beat dashboards
A collector does not need another pie chart. They need today's list, ranked, with the last promise, the last channel, and the amount due. If the model cannot write that list, it is a research project. OrbitMFI's collections AI is only useful when SMS, USSD, and field visits share a contact log that stops messaging the moment a receipt posts.
Do not nag the current book
The cheapest way to destroy NPS and raise complaints is to remind customers who already paid. Reconciliation is therefore a collections feature. MoMo webhooks, Paystack, and teller cash must hit the instalment before the 6 a.m. SMS run. If they do not, pause outbound until they do.
Cost-to-collect as a first-class metric
Track SMS cost, officer time, and recoveries together. A model that doubles texts and holds PAR flat has failed. Tune channel by product: HP vehicle customers often need a call; susu savers need a collector in the market; salary-backed loans need an employer reminder, not a threat.
