Every next-best-action programme that fails, fails on the offer library — not on the model, and not on the delivery.
That sentence should change how Indian marketing leaders evaluate the current wave of artificial intelligence (AI) decisioning pitches. The Indian market in 2026 splits into two tiers on this capability. Telecom operators, large private banks and the bigger insurance and payments businesses have working real-time decisioning, because they own direct, high-frequency digital interactions with their customers. Almost everyone else — fast-moving consumer goods (FMCG), pharmaceuticals, spirits, most of retail — is being sold “AI-powered personalisation” while still running batch campaigns against static segments. The vendors are selling engines. The question nobody asks is what the engine will decide between.
Start with the framing, because it determines the conversation you end up having. Personalisation is a content problem: right message, right person. Next best action (NBA) is a commercial problem: at the moment this customer is interacting with us, what is the single best thing we can do — with full awareness of competing objectives, revenue against experience against regulatory constraint? Ask a leadership team “how sophisticated is your personalisation?” and you get a tour of the content stack. Ask “what is your next-best-action capability?” and the answer almost always reveals the truth underneath: batch campaigns, static segments, no real-time component. The NBA framing raises the discussion from content operations to commercial decisioning, which is where the money is.
Across more than a decade of deployments, from enterprise decisioning programmes in the mid-2010s to the machine-learning-native platforms of today, the architecture question has never changed. A working NBA capability needs four things: an offer library the engine can consume; real-time customer context; a scoring model; and delivery integration into the channels where interactions actually happen. The pattern in client after client is remarkably stable. Most have the scoring capability — a data science team that can build propensity models is now table stakes. Most have partial customer context, usually a batch-updated profile dressed up as real-time. Delivery integration is almost always the underestimated engineering problem, the item that becomes scope creep in month four. And the offer library — the candidate set of offers, messages and actions, each with eligibility criteria, business value weighting, frequency caps and channel exclusions — very few have in any form an engine can read.
This is where the brand lens and the evidence lens converge, which is rare. The brand argument says an offer is a promise, and a business that cannot state what promises it is currently making, to whom, under what conditions, does not control its own brand at the moment of interaction. The evidence argument says a decisioning engine given three stale offers to choose between will make an impeccably scored choice between three bad options. Both lenses arrive at the same place: the constraint is not intelligence, it is inventory.
Why does the offer library break first? Because it is not a technology artefact; it is a marketing operations discipline. Maintaining a live offer library means someone defines offers systematically, someone approves them, someone version-controls them, and — hardest of all — someone deprecates them when they expire or conflict. In practice, most Indian brands manage offers in spreadsheets, email threads and tribal knowledge held by whoever ran last quarter’s campaign. The engine vendors do not surface this in the sales cycle because it is not their product. It surfaces six months into implementation, when the decisioning layer is live, the models are scoring, and the business realises it has eleven eligible offers where it needs three hundred, half of them with eligibility rules that exist only in someone’s head.
The generative AI wave makes this constraint more binding, not less. Large language models can now write the individually tailored copy for a selected offer in real time — the decisioning and the creative are merging. But generation without governance is a liability, particularly in categories where the Telecom Regulatory Authority of India (TRAI), the Reserve Bank of India (RBI) or advertising codes constrain what may be said to whom. A generative layer needs a structured, governed offer library beneath it more urgently than a rules engine ever did, because it will otherwise manufacture promises no one approved. The businesses treating generative AI as a shortcut around offer management are automating the production of ungoverned commitments.
There is also a distinctly Indian dimension. The second tier of the market — indirect-distribution businesses without high-frequency direct interactions — is often told that NBA is simply not for them yet. That is half right. The real-time engine may be premature, but the offer library discipline is not. A pharmaceutical business managing its engagement programmes for prescribers, or an FMCG company managing trade promotions across distributors, already runs a de facto offer library — in spreadsheets. Building the governed version now is what makes the eventual decisioning investment land, and it costs a fraction of an engine licence.
So the diagnostic for a senior leader is a gap analysis in four questions, asked in the right order. Do we have an offer library in machine-readable form, with owners, eligibility rules and a deprecation process? Do we have customer context that is genuinely current at the moment of interaction? Do we have scoring — and is it connected to anything? Do we have delivery paths into the channels where our customers actually show up? If the honest answer to the first question is no, stop there. Do not procure the engine.
On Monday: ask your marketing operations lead to produce the complete list of offers your business could legitimately make to a customer today, with eligibility rules and expiry dates. Give them a week. What comes back — its completeness, its format, how many people had to be emailed to assemble it — is a more accurate assessment of your next-best-action readiness than any vendor demonstration you will see this year. Fund whatever fixes that list before you fund anything that scores it.