Next-Best-Action Models for Sales Teams

The problem it solves

An account manager at a building supplies wholesaler looks after 350 trade customers. On Monday morning she has a CRM full of activity, a list of overdue quotes, a promotions calendar and a vague sense that a few customers have gone quiet. What she lacks is an ordered list: who to call first, and what to say.

Next-best-action models exist to produce that list. For every account, they weigh the possible actions and recommend one, with a reason. Done properly they replace an hour of CRM browsing with ten minutes of reading.

How a next-best-action model is put together

Despite the name, it is rarely one model. It is a small decision system with three layers.

1. Candidate actions. A finite list the business actually offers: call to check in, send a reorder reminder, offer the bulk discount, introduce a new product line, escalate a service issue, do nothing.
2. Predictions per action. Models estimating, for each account, the likely outcome of each action. Probability of reordering this month, probability of responding to a cross-sell, risk of lapsing.
3. Business rules and value. Expected value of each action, minus its cost, filtered by rules: no discount to accounts on credit hold, no more than one contact a week, compliance restrictions on certain products.

The recommendation is the highest-value action that survives the rules. 'Do nothing' must be a genuine option. A system that always recommends contact trains customers to ignore you and salespeople to ignore the system.

What goes into the predictions

Most of these come from order and CRM data you already hold. The engineering effort is in computing them consistently every night, not in exotic modelling. If your CRM and ERP do not talk to each other, that integration is the first project; our post on CRM integration that actually helps is the place to start.

Why most of these projects fail

Accuracy is rarely the problem. The usual causes are duller:

• The recommendations live in a separate dashboard that nobody opens
• Reasons are missing, so experienced salespeople assume the system is guessing
• It recommends actions the salesperson cannot take, such as a discount above their authority
• It ignores what the rep already knows, like the customer being on holiday
• Nobody measures whether following the recommendation led to more sales than not following it

If a recommendation cannot be accepted, dismissed or corrected in one click inside the tool the salesperson already uses, it will not be used.

The dismiss button is not a courtesy. Every dismissal with a reason ('already spoke to them', 'customer is closing a site') is training data for the next version.

Measuring whether it works

The tempting measure is revenue from accounts where the rep followed the recommendation. It is misleading, because reps follow the recommendations that look obviously right, and those accounts would often have ordered anyway.

A fairer test holds a random share of accounts out: their reps get the usual CRM view, everyone else gets recommendations. Compare revenue per account over a quarter. It is slower and less flattering, and it is the only number a finance director should accept. The same thinking underpins uplift modelling, which is worth reading if you want recommendations based on what an action changes rather than who was going to buy.

When a next-best-action model is overkill

If your team handles fewer than about a hundred active accounts each, a well-designed report sorted by 'days overdue against normal reorder cycle' will capture most of the value. The same goes for businesses with only one or two possible actions, where the question is really just 'who to call', which is a single propensity score rather than a decision system.

It is also a poor fit where sales are dominated by a handful of large, relationship-driven accounts. Nobody needs a model to tell them to call their biggest customer.

A sensible first version

We would start with three actions, not twelve: reorder reminder, check-in call for drifting accounts and cross-sell of one product category. One prediction model per action, simple value rules, recommendations shown inside the CRM with a reason and a dismiss button, and a holdout group from day one.

At SpiderHunts that version typically takes six to ten weeks, most of it on data and CRM integration. If it moves revenue per account after a quarter, add actions. If it does not, you have spent relatively little finding out. For the wider picture of automation around sales, our AI agents service covers the cases where some of those actions can be drafted or sent automatically once the recommendations have earned trust.

Frequently asked questions