Personalised messaging: AI for email, push and WhatsApp campaigns

Personalised marketing AI is often sold as a new platform to replace the email tool, the push service and the WhatsApp provider you already pay for. It should not be. The valuable part is the decision layer: for each customer, which products or content to feature, which offer if any, on which channel, and at what time. Those decisions can be injected into the tools you already have through a handful of fields, and the whole thing can be measured against a control group so the marketing team knows what it is worth. This article describes that decision layer, how it plugs into existing tooling, and what the WhatsApp channel adds and demands.

What personalised messaging changes

A conventional campaign picks a segment, writes one message and sends it at one time. Dynamic messaging keeps the campaign but makes three things per customer: the content block, drawn from the same recommender that ranks your storefront; the offer, chosen by a next best offer model that weighs margin against the probability of conversion; and the timing and channel, chosen from when and where that customer has actually responded before. The copywriter still writes the frame. The AI fills the slots and picks the moment.

Where the AI sits relative to your tools

AI email personalisation in practice

Per-user product picks

The simplest and most reliable form is a block of recommended items in an otherwise standard email, populated at send time from the same ranking used on the site. It keeps one recommender for all channels, so a customer sees consistent suggestions, and it lets the email tool do what it is good at. The engine produces a small list of item IDs per customer; the template renders them. If your recommender has good cold-start handling, even a customer with one purchase gets a sensible block.

Send-time optimisation

Customers open messages at habitual times. A model that learns each customer's response pattern and chooses a send window within the campaign's day is a low-risk, measurable improvement, and most email platforms expose a per-recipient send time or a scheduled API that the decision layer can drive. Keep a control that sends at the campaign's default time so the gain is proven, not assumed.

Subject lines and copy

LLMs can generate subject-line and copy variants, but the disciplined use is to generate a handful of variants per campaign for a human to approve, then let the decision layer assign variants to customers and learn which works for whom. Fully automatic per-customer copy generation is hard to review for brand and compliance, and the measured gain over a few well-chosen variants is usually small. Start with variants and a selection model.

Next best offer: choosing the incentive, or none

Blanket discounts give margin away to customers who would have bought at full price. A next best offer model estimates, for each customer, how much an offer changes the probability of purchase, and recommends no offer for customers who will convert without one. That is where most of the value of offer personalisation sits: not in cleverer discounts but in fewer of them. The model needs past campaigns with some randomisation in who received offers; if your history has none, the first campaigns must include a random offer arm to learn from. This is the same logic as next-best-action applied to promotions.

WhatsApp: what the channel adds and requires

In India and much of Asia, WhatsApp outperforms email for many consumer brands, and it supports conversation, not only broadcast. A personalised WhatsApp message can carry product cards, and a reply can be handled by an agent that answers questions, checks stock and completes an order. The channel comes with rules: business-initiated messages must use approved templates, customers must have opted in, and message categories carry different per-conversation charges. The WhatsApp Business Platform documentation sets out template approval and messaging policy; the decision layer must respect both, and the opt-out must flow straight back into the identity record so no other channel picks the customer up. What a conversational agent can do on the channel is covered in what a WhatsApp AI chatbot can actually do.

Frequency, fatigue and channel choice

Personalising content while sending as often as before wastes the gain. A frequency model treats every send as a cost as well as an opportunity: each customer has a tolerance, visible in unsubscribes and declining opens, and the decision layer should skip a send when the expected value is below the fatigue risk. Channel choice follows the same reasoning. Email is cheap and tolerated at higher frequency; push is immediate and tires quickly; WhatsApp is the most effective and the least forgiving, so it is reserved for the highest-value moments. The output of the decision layer is sometimes silence, and that should be counted as a decision.

Measuring it against a control

Every personalised programme needs a random holdback that receives the standard campaign: same template, default content, default time, standard offer. Revenue per recipient over a sensible attribution window is the primary metric; opens and clicks are diagnostics. Unsubscribes and WhatsApp opt-outs are tracked as a cost in the same report. The design of that test is described in why you must run a controlled test; the one addition for messaging is that the control must be held at the send layer, so the campaign tool sends the control group the default rather than the decision layer quietly substituting picks.

A worked example

A D2C brand was sending the same weekly email and a monthly WhatsApp broadcast to its whole list, with a discount attached to most of them. We built a decision layer on top of the brand's existing event data and the recommender already serving the storefront. Each week it produced, per customer, a block of recommended items, an offer decision that was most often "none", a send window and a channel. Those were written as fields into the email tool and as an audience into the WhatsApp provider. A random holdback kept receiving the old broadcast. Customers who replied on WhatsApp reached a conversational agent that could answer product questions and take an order, with a hand-off to a person for anything else. The measured result was higher revenue per recipient in the personalised arm and a marked fall in the share of orders carrying a discount, which the finance team cared about more than the lift. The build is described in the D2C personalisation and WhatsApp agent case study.

Team and timeline

The decision layer is built by an ML engineer and a data engineer with a marketing owner on the client side who controls templates and approvals. Where a storefront recommender and an event pipeline already exist, the messaging layer with a control takes around four to six weeks for email and push; adding WhatsApp with a conversational reply path adds a phase. The work is scoped under personalisation engines, from $21,000 / ₹13.6L; a conversational WhatsApp reply path is a customer service agent build, from $12,500 / ₹8L. Care Plans keep the models, templates and opt-out sync monitored; see the pricing page for tiers.

Before you start: a checklist

• Confirm a stable customer identity spans storefront, app, email and WhatsApp
• Check your email, push and WhatsApp tools accept per-recipient fields or audiences via API
• Decide the primary metric and attribution window with finance before the first send
• Reserve a random holdback that receives the standard campaign at the send layer
• Audit WhatsApp opt-ins and template approvals before any personalised broadcast
• Plan a random offer arm if past campaigns had no variation in who received discounts
• Agree who approves copy variants and how often they are refreshed
• Wire unsubscribes and opt-outs back to the identity record immediately

Questions clients ask

• Do we have to change our email platform? No; the decision layer writes fields and audiences into the tools you have, which is the point of the design.
• Can the AI write every message from scratch? It can, but approving a few variants and letting a model assign them is safer and usually performs as well.
• Is WhatsApp worth the per-conversation cost? For high-value moments, often yes; the decision layer should reserve it for those and use email elsewhere.
• How do we know offers are not just cannibalising margin? The control receives the standard offer; the report shows revenue and discount share side by side.
• What data do we need to start? Order history, message history with sends and responses, and an event pipeline with one customer ID.

Related reading

See event pipelines for the identity foundation and real-time ranking for the in-session side; TheEazy CXM, the group's CRM product, is one place these decisions can be surfaced to a marketing team. Prices and Care Plans are on the pricing page.

Keep the tools, add the decisions, hold back a control: that is personalised messaging that pays for itself and can prove it.