Problem breakdown · No. 033 · November 18, 2026
Co-op members dispute payouts they cannot trace to a delivery
Producer co-operatives lose member trust when a payout arrives as one number with unexplained deductions. The fix is a deterministic ledger that links every rupee back to a weighed, tested delivery, not a smarter chatbot.
The problem
A dairy or produce co-operative collects from many small members every day. Each delivery is weighed, often quality-tested for fat, moisture or grade, and priced by a formula. Every payment cycle the member receives one net figure, minus deductions for input advances, feed, loan instalments, testing fees and share capital.
When that figure is lower than expected, the member cannot see why. The secretary answers the same question many times a cycle, from paper registers and a spreadsheet. Some disputes end in a manual adjustment nobody records properly. Others end with the member quietly selling to a private trader at the gate, which is the real cost: the co-op loses volume and the member loses the patronage benefit.
Existing tools stop short. Collection-centre software records weight and quality, and accounting software records the payout, but the link between them is a batch export. The member-facing output is a printed slip or a text message with a total. Nothing lets a member trace a deduction to the delivery or advance that caused it.
Constraints
- Budget is small and recurring fees face a committee vote, so anything paid must be cheap and obviously useful.
- Members use basic phones and messaging apps, often in a regional language, and literacy varies.
- Data already exists but is messy: collection-centre exports, a ledger spreadsheet, paper advance slips.
- Co-op accounts are audited and governed by cooperative law, so the system cannot rewrite figures, only explain them.
- Trust is the product. One wrong explanation does more damage than no explanation.
- The staff operating it are the secretary and a clerk, not engineers.
System design
The core is plain deterministic code.
- An importer reads the collection export and the ledger into an append-only table of events: deliveries, quality readings, price-chart versions, advances, deductions, payouts.
- A payout calculator recomputes each member's cycle from those events using the published formula, and flags any cycle where its total disagrees with the paid amount.
- A statement builder renders a short, line-by-line statement per member: each delivery with weight, reading and rate, then each deduction with its source document and date.
The LLM belongs at the edges. It translates the fixed statement template into plain regional language once, reviewed by staff, not per member. It also reads incoming member questions on the messaging channel and classifies them: which cycle, which line. The answer it sends is assembled from ledger rows, with the model only allowed to choose and phrase rows it was given. If the question does not map to a row, it routes to the secretary.
Scanning paper advance slips with a vision model is tempting, but I would keep it as a suggestion queue that the clerk confirms, never a direct write.
Trade-offs
The design does not fix errors in source data. If a reading was entered wrong at the centre, the statement faithfully shows the wrong reading. That is acceptable, because a visible wrong number becomes a dispute about one row instead of a dispute about the whole payout.
It also refuses free-form answers. Some members will want conversation and get a routed ticket instead. Narrow answers that are always correct are worth more here than fluent ones that are sometimes wrong.
What I would build first
The importer and the reconciliation report, for one collection centre and the last few cycles. The output is a list of members whose recomputed payout differs from what was paid, plus a printable itemised statement for each member.
If the secretary finds even a handful of real mismatches, or answers a week of questions by pointing at a statement line, the value is proven. Messaging and the question classifier come after that, on top of a ledger that already holds.
Before adding a conversational layer, make every number a user disputes traceable to a single source row, and let the model only quote those rows.