Next Best Actions

What are next best actions?

Next best actions are the ranked recommendations a system produces for what should happen next in a customer relationship, computed from that person's account context, case history, entitlements, and the intent expressed in the current conversation. Each candidate is scored, ordered, then surfaced to a human agent or executed automatically.

The phrase came out of marketing automation, where the ranked output was usually an offer. In support the candidate set is operational: issue the refund, reset the credential, order the replacement part, escalate to a specialist, or ask one clarifying question before committing to any of them.

How next best actions work

A next best action system runs as four layers over a feedback loop.

Context assembly comes first. The system pulls the customer record, open cases, entitlements, and recent events into one working state, then applies intent recognition to establish what the person is actually trying to accomplish on this contact.

Candidate generation comes second. The system enumerates the actions that are permissible for this customer at this moment, filtered by policy, plan, and system availability, so an action nobody is entitled to never reaches the ranking stage.

Scoring comes third. Each candidate gets a predicted value combining likelihood of resolution, expected cost, and the risk carried if the interpretation is wrong. A confidence score on that interpretation gates whether the top candidate is offered at all.

Arbitration and execution close the loop. One action wins, it is presented or executed, and the outcome is written back as training signal, which is why the same machinery drives call routing decisions on voice as it does replies in chat.

Types of next best actions

Support deployments generally rank across four classes of action.

• Informational actions: The system answers the question or explains the governing policy, the lowest-risk class, though a wrong explanation still generates a second contact.

• Transactional actions: The system writes to a system of record, issuing a refund, changing an address, or cancelling a subscription, which makes reversibility the central design question.

• Routing actions: The system moves the conversation to a queue, a specialist, or a supervisor, correct whenever the best next step is a different person.

• Commercial actions: The system offers a retention credit, a plan change, or an upgrade, and these need hard separation from service actions so an offer never delays the fix.

Next best actions vs next best offer vs recommendation engines vs decision trees

Teams conflate these four because all of them end with something being suggested to somebody. Next best offer ranks commercial propositions by expected revenue for a customer who may not have asked for anything. A recommendation engine ranks catalogue items by predicted affinity, with no notion of entitlement or case state. A decision tree encodes a fixed path an author wrote in advance, identical for everyone who lands on the same branch. Next best actions rank operational moves against the live state of one open case, which is why entitlement and reversibility sit inside the ranking.

If the question is what this customer needs done on the open case right now, you need next best actions. What to sell is next best offer. A decision tree fits only where the sequence must be identical every time, such as a regulated disclosure.

Why next best actions matter for customer experience

When nothing ranks the next step, every agent improvises one from memory, and two customers with identical accounts get different outcomes on the same day. Neither agent is wrong; there was nothing to be right about. The cost surfaces as reopened cases, inconsistent credits, and escalations that reach the specialist without the diagnostic step that would have made them solvable.

A ranked step also compresses handle time on cases that stay with a person, because the expensive part of most contacts is deciding what to do, not doing it.

The tradeoff is real. A ranker optimises whatever it is scored on, so scoring it purely on containment teaches it to suppress the escalation a customer needed, which is why a human in the loop checkpoint stays on high-consequence actions long after accuracy looks acceptable.

How are next best actions measured?

Four measures matter, and they only make sense read together.

Acceptance rate is the share of recommendations a human agent takes when one is offered, which reads the ranker's credibility on the floor. Completion rate is the share of accepted actions that finish without failing in a downstream system. Reversal rate is the share of executed actions later undone, refunds clawed back or escalations bounced, and it is the honest quality signal. Uplift against a holdout, where a random slice of traffic receives no recommendation, is the only measure showing the ranker beat doing nothing.

Set expectations against adoption context: the Stanford HAI AI Index reported organisational AI use climbing from 55% in 2023 to 78% in 2024, so most peer figures you are quoted describe young, unstable deployments. Build your own baseline first, in the spirit of these support quality metrics from Uber.

How AI agents change next best actions

Classical next best action engines required someone to enumerate the candidate set in advance, which is why they lived inside banks and telcos with analysts on staff to maintain them. An AI agent generates candidates at runtime: it reads the conversation, retrieves the governing policy, inspects the tools it can call, and proposes a fitting action, including combinations nobody wrote down.

That moves the control problem. The enumerated list is no longer the safety mechanism, so AI guardrails become the enforcement layer, constraining which tools may be called, with which arguments, and under whose approval.

The consequence is that recommendation and execution collapse into a single step. An agent holding a refund API does not stop at suggesting the refund, which is the design shift covered in these backend action automation tools.

How to implement next best actions

Implementation runs in five steps.

1. Mine resolved outcomes. Pull closed cases and record which action actually resolved each contact reason; that ranked list is your first candidate set.

2. Encode entitlement before ranking. Plan, region, and account state decide what is permissible at all.

3. Ship in suggest mode. Let agents accept or reject for several weeks, and read acceptance and reversal.

4. Promote reversible actions to execution first. Address changes before refunds, refunds before cancellations.

5. Hold out a slice of traffic so uplift stays measurable after launch.

Integration surface decides feasibility: an engine that cannot write to your CRM and billing system produces advice and nothing more. Governance decides trust, since regulated buyers ask who approved a scoring change and how that approval is evidenced, which is where ISO 42001 and SOC 2 Type II enter procurement conversations. The constraint teams underestimate is reversibility: every executable action needs a defined undo path before it is ever ranked.

Next best actions and agent assist

The clearest place to watch next best actions in production is agent assist, where the top candidate appears in the sidebar alongside the policy that justifies it, and the agent's accept or reject becomes the signal that retrains tomorrow's ranking.

Ticket prioritization answers the neighbouring question one level up: it orders which case a person picks up, while next best actions order what happens once that case is open. Teams that tune one and ignore the other work the right cases in the wrong sequence.

What do next best actions mean in plain terms?

Think of a next best action system as a triage nurse standing behind every conversation: it looks at who walked in, what has already been tried, and what the room can actually do, then says handle this one thing first.

Without it, the choice falls to whoever picked up the case, so the outcome depends on how long that person has worked there and what they happen to remember about the policy. Two people with the same broken order get a replacement and a store credit on the same afternoon, and both agents will tell you honestly that they followed the rules.

The tradeoff is that a ranked suggestion is a strong nudge. Agents tend to take the top item even when a case is unusual, so the value of the ranking depends on people keeping the willingness to override it, and on the system making each override easy to record.

Common next best action mistakes

Four patterns account for most disappointing deployments.

Ranking on the easiest metric comes first. Containment and handle time are convenient objectives, and a model scored on them learns that the cheapest action is often refusal, which shows up months later as complaint volume nobody attributed to the ranker.

Skipping the entitlement filter comes second. When permissibility is checked after scoring, agents get handed actions the customer cannot have and spend the call walking them back, which destroys acceptance rate faster than any accuracy problem.

Training only on historical agent behaviour is third. Past decisions encode staffing pressure, tenure gaps, and workarounds for systems that have since changed, so a model fitted to them reproduces yesterday's constraints as today's recommendations.

Launching without a holdout is fourth. With no untreated slice of traffic, every seasonal shift and shipped product fix gets attributed to the recommendation engine, and the first time results dip nobody can tell what actually moved.