Practical notes from the engineers who build agents, copilots and AI products.
Track inference and retrieval cost per account against plan price so the AI tier stays inside its margin. AI cost per account shows whether a premium tier is a business or a subsidy: attribute every model call, embedding and index query to a tenant, price it at provider rates, and report it weekly beside plan revenue.
Reminder calls with one-touch confirmation and in-call rescheduling cut no-shows and re-fill cancelled slots the same day. Unlike an SMS blast, the patient can act in the call: confirm, move to a slot offered from the live schedule, or cancel, and the freed slot goes to the waitlist within minutes.
Vendor demos on curated data, no eval suite, single-model lock-in, no running-cost estimate and no security review are the warning signs. Each predicts a failure after launch: accuracy that collapses on real data, regressions nobody catches, a provider change that breaks the product, a surprise bill or audit findings.
A recommendation engine ranks products per shopper from behaviour and catalogue signals, with fallbacks for new users and products. It is layered: candidate generation, ranking, business rules and a controlled test that proves the lift, and it needs clean events and honest measurement more than a clever model.
Look for the measurement method, the control, the ownership of the system and whether it still runs; ignore round-number uplift claims. A case study that names its metric, its baseline and its failure modes tells you how the vendor works; one that leads with a percentage tells you how the vendor sells.
RBI expectations cover outsourcing risk, data localisation, audit trails and accountability for automated decisions. There is no single AI rulebook; the existing directions on IT outsourcing, payment data storage, IT governance and fair practices apply to any AI system a bank or NBFC runs, and the board is accountable.
Ask how they measure accuracy, who owns the code, what happens when a model changes, and to see a production system they still run. The answers separate teams that build evaluated, owned, maintained systems from teams that build demos. Ask the same questions of every agency and compare the answers side by side.
Policy gating lets an agent approve refunds or returns only inside rules you define, with everything else routed to a person. The rules live in configuration enforced at the tool layer, not in the prompt, so a persuasive customer or a confused model cannot move money outside the limits you set.
Deliver low-risk modules first, migrate data per module, train per team and run parallel until reconciliation matches. Phased ERP implementation trades one frightening cut-over for a series of small ones, each with its own data migration, trained team and reconciliation before the old process is switched off.