Personalisation at Scale: What AI Can Actually Change in Your Marketing

AI personalisation is usually sold like this: every customer receives a message written for them alone, automatically, without hiring anyone. That much is technically possible now, and what decides the result is the state of your customer records.

Personalisation runs on data, so the work starts before any model is involved. For a growing business, the version that pays back is almost always narrow: a handful of segments built from data you already collect, with the drafting and sorting handled automatically.

What personalisation means once you strip the pitch out of it

Personalisation means sending a different message based on something you know about the recipient. What separates one implementation from another is how fine the distinction is and where the information comes from.

The first level is segment-based. Customers are grouped by simple rules: first-time buyers, regulars, people who have not come back in six months. Each group gets one message written for them. Almost any business can run this with the data it holds today.

The second level is behaviour-triggered. A message goes out because something happened: a cart abandoned at checkout, a warranty approaching its end date, a third order completed. All this needs is one event recorded with a date, plus a way to send from it.

The third level generates copy per person, drawing on their history and preferences. Here AI does something plain rules cannot. For a list of a few thousand names this level rarely qualifies, because the cost per send climbs faster than the return.

The customer data you need, and the state it is usually in

Three things have to exist. Identity, so you know these two transactions belong to the same person. History, so there is something to base a distinction on. Recency, so a date can act as a trigger.

What we usually find looks like this. Orders sit in the point-of-sale system or in a marketplace account. Customer conversations sit in WhatsApp, split across two people's phones. The email list lives in a separate tool and contains names collected two years ago. This is normally the point at which an owner starts thinking about a system of their own, because the customer records are scattered across several spreadsheets and tools that do not talk to each other.

One customer can appear three times with their name spelled three ways, and no customer number links the three records. Which identifier gets used and how records are merged, we write down alongside the code, so the next person knows why two records were joined.

Without a shared identifier there is nothing to personalise. All you can send is one message to everyone with a first name pasted on the front. Joining those records, picking a single identifier, clearing duplicates and connecting the sources usually takes the largest share of a project like this, and that part has to happen whatever tool you end up using.

Where AI does real work in a marketing workflow

AI is worth applying where manual effort is the bottleneck. Four jobs pay back most often.

Drafting variants takes one offer and turns it into three versions for three segments, with different tone and emphasis. Your people still decide the offer and approve the copy, but the first draft no longer eats a working day.

Summarising incoming conversations turns a forty-message thread into five lines. The summary sits at the top of the thread, so the salesperson reads those five lines before replying and already knows what this customer asked about and complained about.

Sorting inbound requests sends price enquiries, delivery complaints and large quote requests to different people without anyone reading through them first.

Tagging and cleaning records consolidates product names written four ways and labels customers by category from their purchase history. The work is dull and high in volume, which makes it a good fit for a machine. A wider set of examples for smaller businesses is in the piece on practical AI use cases for growing SMEs.

Where it does not, and the honest version of that

Generated copy does not fix a weak offer. If customers are not coming back because the price does not work or delivery is slow, a thousand personalised variants will still be ignored, and the offer itself is what has to change. Where generative AI earns its place and where it runs out is covered in what generative AI does well for business.

Generating copy per person carries a cost that adds up quickly once you count it per send. That cost goes into the written Discovery plan, so the number is in front of you before you decide to move up from segments to per-person copy. For a list of a few thousand names, the difference in result between three segments done seriously and per-person copy is usually smaller than the difference in cost. Start with segments and move up if your send volume gets large enough.

If an email tool or CRM already on the market covers the segments and triggers you need, we will name the tool and you build nothing. Building your own only enters the picture when your data sits across several systems that the tool cannot connect, or when the segment rules are specific to your business. The broader view of when AI and automation are worth building is in the guide to AI and automation for businesses.

A small first version that is worth running

Take one segment and one trigger. An example where the material almost always exists: customers who bought once and have not returned in ninety days.

Four pieces need building for that. The data join, which means picking one identifier, clearing duplicates and connecting the sources so one customer reads as one person. The segment rule, stored separately from the code, in a sentence a non-technical person can read and change without a deploy. The message, one or two versions the owner has approved. Then the measurement: how many were sent, how many opened, and how many bought again, on a single page the operations lead can open themselves.

Keep the size small so the result is readable. Running six segments at once in the first month leaves you unable to tell which one worked, while one segment running for two months gives you an answer you can use to decide whether a second segment is worth it.

What it looks like to commission this

Our first stage is Discovery, which produces a written plan and a fixed price for the stage that follows. That document is yours whether you continue or not. Scope and price for every stage are agreed in writing before the stage starts, so you can stop after any stage.

During the build there is working software every week that you can open yourself. You see the segment logic and sample sends before any budget goes behind them. If the rule turns out to catch the wrong customers, you find that out in week two.

One thing is worth settling at the start, whoever does the work. The customer data, servers, domain and every account behind that marketing system are registered to your business from day one, not to a vendor.

Bring one segment and one trigger to a first call

You do not need to clean your data before this conversation. The join is still the largest piece of work later, but mapping it is what we do with you. Name one customer group you want to address differently, one event that justifies sending a message, and where your customer records currently live. We will work through whether the material is already there and what needs joining. The first call is free, no deck, straight with the engineers.

If something already on the market covers it, we will point you at that tool and you build nothing. If there is real work in it, Discovery gives you a written plan and a fixed price for that stage, and the document is yours whatever you decide afterwards. Use the contact form, which opens WhatsApp, or email hello@arktik.id.