Three coloured paths branch from a common point on a layered paper map.
Original illustration by AI Boss Limited.

A rental business is considering a two-payment option for weekend equipment bookings. The sales team wants to know how much the new option could add to revenue. The service manager wants to know whether a smaller first payment helps customers complete a booking. An AI tool can organise the records for that discussion. The team first has to agree on which routes a customer can take.

Start with one booking decision

Pick one rental package. Ask product and finance to confirm who can book it and what each payment option costs. The proposed test gives one group the usual full-payment offer and another group a choice of payment plans. Either group can leave without a booking. Adrian Dunkley's Customer Equity Trees proposal maps paths like these and asks how each could affect a customer's financial position.

The split plan leaves the customer with J$5,000 more cash on the day they book. If they make both payments, the plan costs J$1,000 more. Finance still needs payment, refund and cost records to see what the business kept.

Payment schedules in the booking example
Payment planDue at bookingDue 30 days laterTotal due
Pay in fullJ$10,000J$0J$10,000
Pay in two partsJ$5,000J$6,000J$11,000
Illustrative prices in Jamaican dollars.

Check what the customer was offered

In a parallel circuit, opening a switch stops current through that branch. A checkout rule can close off an option. The assigned group determines what the page should offer. A customer’s choice still depends on their circumstances, and you need a record of what the page showed them.

Expected checkout routes

Group 1

  • Pay in full
  • Split plan unavailable
  • Leave without booking

Group 2

  • Pay in full
  • Pay in two parts
  • Leave without booking
Compare the two checkout setups before any customer responds.

Give the AI a dated rules file, a group assignment export and a checkout event export. Use an approved study ID to join the records and a session ID and timestamp to tell repeat visits apart. Ask the tool to list the routes the rule permits for each customer. For each session, ask it to cite the event that recorded which options appeared and the separate event for the plan the customer selected. With no display event, mark the options shown as unknown, even if a selection has a record. Keep names and card details out of the export. Ask a product analyst to check a sample against the source records and the split plan against the billing schedule.

Compare groups over the same period

Ron Kohavi and his colleagues describe a controlled experiment in which “live users are randomly assigned to one of two variants”. In a pilot, randomly assign each eligible customer before showing the checkout page. Keep their assignment fixed across repeat visits. Count a customer once if they book the chosen package and their first payment clears within seven days of assignment. Divide the number of customers who booked by all assigned customers in each group. People who leave or have a missing display event stay in that denominator. Customers who choose the split plan may have different needs at the start, so keep the comparison at the assigned-group level.

Allow seven days for the second payment to clear after its due date. Check collections 37 days after each qualifying booking, using only the first qualifying booking per customer. For each group, subtract refunds and direct fulfilment costs from the cash collected, then divide by everyone assigned. Keep costs the business incurred even if a qualifying booking was cancelled. Separately, record the share of assigned customers who made at least one complaint about the price or due date within 44 days of assignment. These measures show what happened at booking and over the payment period. The effect on net worth needs a longer study with evidence on customer assets and debts.

Agree on the thresholds before launch

Before the pilot, the product manager owns the decision on whether to extend the offer. Finance sets the minimum acceptable result for collections, and the service manager sets a complaint limit. Agree how to judge uncertainty in both results. Ask the product analyst to reconstruct the checkout for one eligible customer in each group before launching. If an offer cannot be traced to a display event, fix the logging before collecting pilot evidence.

Sources & further reading

Adrian Dunkley, Customer Equity Trees and the choices a lifetime value score hides ↗Kohavi, Henne and Sommerfield, Practical Guide to Controlled Experiments on the Web ↗OpenStax Physics, section 19.3, Parallel Circuits ↗

Sources checked 2026-11-17.