Practical notes from the engineers who build agents, copilots and AI products.
African fintech, telecom and logistics companies use the same agent, voice and modernization playbooks with regional partners. The conditions are familiar to Indian teams: mobile-first customers, WhatsApp and USSD as channels, many languages in one market, patchy networks and regulators who want data kept at home.
Copilots that act, retrieval over the customer's data and natural-language reporting are the features users pay more for. A chat box over your help centre is not one of them. The tier holds when the feature removes recurring work inside the customer's own records and is priced with its running cost known.
The method is identical; the constraints are not. An AI discovery sprint for startups optimises for speed to one shippable use case. An enterprise sprint spends a third of its time on access, security review and integration depth, and ends in a document that has to survive a committee.
Australian clients overlap with IST mornings, get fixed-price programs and ownership terms designed for distant teams. Sydney's afternoon is Bengaluru's morning: three to four shared hours a day; Perth shares most of it. The rest runs on written hand-offs, demos in your hours and a contract handing over everything.
AI drafts lessons, quizzes and summaries from approved sources; educators review before anything reaches students. AI course content generation is a review workflow with a model inside it, not a publish button: sources are fixed, outputs are structured, every claim is traceable and a teacher signs off each item.
AI collections combine reminder calls and WhatsApp with payment links, within RBI conduct rules and consented communication. The agent handles early-bucket reminders, promises to pay and payment links; hard cases and disputes go to people, every contact is logged, and calling hours and frequency are enforced in code.
Chatbots are cheap because they do little; agents cost more because they integrate, act and are evaluated; compare on resolution, not price. A chatbot answers from a knowledge base. An agent reads the account, takes gated actions and is measured per intent, which is where budget and value both sit.
Log inputs, outputs, model version, user and reviewer for every automated decision in an append-only store auditors can query. Add the prompt version, retrieved sources, policy checks and the action taken, keep it for the retention period your regulator sets, and make sure a non-engineer can reconstruct any decision.
Agents replace brittle screen-scraping RPA where judgement is needed; keep RPA for fixed, structured steps with no API and no ambiguity. Most estates end up with both: agents deciding and coordinating, RPA bots executing the last mechanical step into a system that has no other door.