Next-best-action systems promise to tell every seller which customer to contact and what to offer: a reorder reminder, a cross-sell, a retention offer, a service visit. Many are quietly switched off within a year. The model is rarely the problem. Sellers ignore recommendations they cannot explain, and a recommendation that is not acted on is worth nothing, however accurate it is.
The systems that last are designed around the questions a seller asks before acting: why this customer, why now, and why this offer? Answering those questions is a design decision, not a modelling one, and it shapes everything from the data to the way results are measured.
Why sellers ignore recommendations
A seller knows their accounts. A list of customers with a score next to each name competes with that knowledge and loses, especially when it is wrong in ways the seller can see: the customer complained last week, the account is on credit hold, the product was discontinued. Recommendations that come back after being dismissed teach the same lesson. After a few such mistakes, the list stops being opened.
Volume is the other problem. Dozens of recommendations a week for a seller who can act on a handful teach the seller to ignore all of them.
Build on signals the seller recognizes
Explainability starts with the inputs. The strongest next-best-action models are built from signals a seller would accept as reasons: an order gap against the customer’s usual buying cycle, a falling share of a category, a product that similar customers buy and this one does not, a contract renewal or price change coming up, open service tickets, a drop in usage or visits. Keep the list short enough for a sales manager to read in one sitting, and name each signal in the words sellers use.
The model can be sophisticated, but its output must be traceable to those signals for each customer. Techniques that attribute a prediction to its inputs make this possible, and they expose data problems early: a recommendation driven by a signal that makes no sense usually points to a data error. Most of these signals already exist in the order history, the CRM and the service system; the work is joining them at the level of the customer and keeping them fresh.
Show the reason with every recommendation
Each recommendation should arrive in the CRM the seller already uses, such as Dynamics 365, with four things: the action, the reason in one sentence built from the top signals, the expected value in plain words, and what happened the last time a similar action was tried with this customer. Language models are good at turning signals into a readable sentence; the facts in that sentence must still come from the model and the data, never from the language model’s own invention. A good reason reads like a note from a colleague: this store usually orders every few weeks, nothing has come in since early spring, and similar stores also carry the larger format.
Business rules belong next to the model. Accounts on credit hold, customers with an open complaint and products on allocation are filtered out before anything reaches the seller. The rest are ranked by value and capped at what a seller can act on in a week. Fewer, better-targeted recommendations get followed; long lists do not.
Let the field answer back
Every recommendation needs an outcome: done, not done, or dismissed with a reason chosen from a short list, such as already ordered, wrong contact, not relevant or relationship issue. Make dismissing quick, one tap and a reason; a form that takes a minute to complete will be skipped and the feedback lost. Feedback also protects the relationship: a seller who can mark a customer as off limits for a while will not be asked about them again next week.
Dismissal reasons are some of the most valuable data the system produces. They reveal data errors, show which recommendation types to retire and teach the model what sellers know and the data does not. Sales managers should review them each month with the team that owns the model. When sellers see their feedback change the recommendations, adoption follows. When it disappears without a trace, the list is abandoned.
Measure uplift against a control group
The most common way to report next-best-action results is also the most misleading: revenue from customers who received a recommendation. Many of them would have bought anyway. The honest measure is uplift, the difference in outcomes between customers who received recommendations and a comparable group who did not, over the same period.
The simplest design holds back a random share of eligible customers, keeps that control group permanently, and compares conversion and revenue between the two groups by recommendation type. It is the same discipline that makes early-alert programs credible: measure the follow-up, not the flag. Some recommendation types will show strong uplift. Others will show none and should be retired, however good their model scores look.
Report uplift in money as well as in conversion, net of the discounts and incentives the offers carried, so that a recommendation that wins deals by giving away margin is visible for what it is. Agree on the size and length of the test before launch, so nobody is tempted to stop it in the week the numbers look good.
Start with one action and one team
Start with one recommendation type that sellers already wish they had, such as reorder reminders for lapsed buyers, one sales team and a control group from the first day. Show the reason, collect the feedback and measure the uplift over a full quarter before adding the next type. Choose a team whose manager will review the recommendations with the sellers every week at first; that attention is what turns a pilot into a habit. The trust earned on the first action carries the ones that follow.