Personalize a mobile app with rules before models

Personalization sounds like a data science project. In most mid-market apps it is not. A short set of rules, based on what you already know about a person, gets you most of the value, and you can explain each rule to a customer without a diagram.

Use what you already know

Start with facts the user has already given you or that your systems already hold: their role, their branch, their language, the equipment they service, the plan they pay for. Rules built on those are predictable. A technician at the Orlando branch sees Orlando's jobs. A customer on the premium plan sees the priority line. Nobody asks why.

Nielsen Norman Group defines personalization as identifying a user as a specific type or individual and delivering appropriate content and functionality to them. Its cautions are worth keeping in front of the team: personalization is a best guess, past behavior does not always predict future actions, and labeling someone the wrong type can leave them frustrated or drive them away.

Personalize functions, not only content

The cheapest wins are small and mechanical. Remember the last filter a dispatcher used. Fill the customer's address from the last visit. Keep a half-finished checklist until the technician returns to it. None of that needs a model. All of it makes the app feel like it was built for the person using it.

Ask before you infer

Some tailoring needs information the person has not given you. Ask for it, when it matters, with a reason. Apple's Human Interface Guidelines on privacy advise requesting only the data you need, and requesting permission when the app clearly needs it rather than at launch. A technician asked for location at the moment they tap "navigate to job" understands why. A prompt on the first screen looks like a data grab.

Keep a way out

Every rule should be overridable. If the app thinks a manager only wants Orlando, the manager should be able to switch to Tampa in one tap and see why Orlando was chosen. This costs little and prevents the worst failure of personalization: the user who cannot find something because the app decided they did not need it.

When a model starts to pay

Machine learning begins to earn its place when you have a decision that depends on many signals, enough history to learn from, and a way to tell whether it worked. Ordering a technician's route by traffic and skill is a candidate. Recommending the next best action for a sales rep might be. Before you build, write down the rule-based version, run it for a month, and keep the numbers. If a model cannot beat the simple rule by a margin you can measure, keep the rule.

What to decide up front

• Which three or four facts you will use, and where each one comes from.
• Who owns each rule, so it gets changed when the business changes.
• What the user can see and override.
• What you will measure, such as time to first action or how often people change the default.

We cover measurement for customer-facing tailoring in personalization you can measure. If your app needs a first set of rules, talk to us about mobile.