Glossary
›
Hybrid recommendation engine

Hybrid recommendation engine

Glossary
Also known as:
Next-best-action engine, LLM-plus-recommender architecture

What is a hybrid recommendation engine?

A hybrid recommendation engine is a decision system that combines a recommendation model, which ranks possible next moves by their expected outcome, with a large language model, which interprets what a customer says and expresses the chosen move in natural language, so an AI agent can select the next best action during a live conversation.

The two halves solve different problems. Recommendation systems have ranked products, offers and content for decades, but they cannot hold a conversation. Language models hold conversations well, but they generate text by predicting likely words, not by estimating which move will produce the result the business needs. Joined, one decides and the other speaks.

How a hybrid recommendation engine works

Within a single turn of a conversation, the engine does four jobs. They are split between the two models on purpose.

StepDone byWhat happens
InterpretLanguage modelReads the customer's words, including partial answers and objections phrased in unexpected ways, and maps them to a structured situation
List candidatesRules or a flow graphSets out the moves permitted at this point: questions, responses, offers, a handoff
RankRecommendation modelScores each candidate on its likely outcome, given the customer's context, the conversation so far and what worked in similar situations
RenderLanguage modelTurns the top-ranked move into speech or text in the right tone and language

Outcome feedback closes the loop. When the conversation ends, its result (an application completed, a renewal, a payment kept) is attached to the decisions made along the way, and the ranking model is updated. Moves that win more often in a given situation are ranked higher the next time that situation comes up.

Why a hybrid recommendation engine matters

Choosing what to offer, and when, is where much of the commercial value in a customer interaction sits. McKinsey's research on personalization found that it most often drives a 10 to 15 percent revenue lift, with company-specific results spanning 5 to 25 percent. The companies that do it best use predictive models, including next-best-action algorithms, to decide what each customer sees.

A language model alone does not make that choice well. Asked what to offer, it produces a plausible answer, and plausible is not the same as most likely to succeed. Asked the same question twice, it may answer differently. Separating the decision from the wording makes the choice consistent, measurable and improvable, while keeping the conversation natural.

Where the engine makes the call

A hybrid engine is useful wherever an agent faces several permitted moves and the right one depends on context.

  • Offer selection. Which product, plan or term to present: an insurance coverage tier, a telecom plan, a credit product, a hotel package.
  • Question order. Whether to ask the next qualifying question now or address the customer's concern first.
  • Objection response. Whether to answer an objection directly, reframe it, or acknowledge it and move on.
  • Retention. Which save option to present, and whether to present one at all.
  • Payment options. In collections, whether to propose payment in full, a date, or an arrangement.
  • Handoff timing. When a conversation should move to a person.

Where there is one product and one path, a ranking layer adds little and simple rules are the better tool. The engine pays off when the choice is real and repeated thousands of times.

What a good hybrid recommendation engine looks like

  • Bounded candidates. The ranking model chooses only among moves already approved. It cannot invent an offer or a statement.
  • Outcome-based training. It learns from what happened after the conversation, not from clicks or sentiment alone.
  • Explainable decisions. For any choice, the system can show which factors drove it.
  • Fairness testing. Offers and options are checked for systematically different treatment across customer groups.
  • Controlled exploration. Some alternatives are tested deliberately, within approved limits, so the engine does not lock onto early winners.
  • Separation of concerns. Changing the wording of a move does not change which move is chosen, and vice versa.

How to measure a hybrid recommendation engine

  • Decision lift. Outcomes when the engine chooses, compared in a holdout with rule-based or random choice among the same approved moves.
  • Impact by decision point. Which decisions move results most, so effort goes where the value is.
  • Consistency. How often identical situations produce identical choices.
  • Downstream value. Not only whether an offer was accepted, but whether the accepted product stayed, was funded or was paid.
  • Drift. Whether performance holds as products, prices and customers change.

Rules that apply to ranking decisions in the United States

Once software decides which option a customer hears, the rules about fair and accurate treatment apply to that software's choices. The US considerations are set out below; regimes abroad differ.

  • Fair lending. The Equal Credit Opportunity Act prohibits discrimination in any aspect of a credit transaction, which includes the products and terms a customer is offered. Ranking models used in credit conversations need testing for disparate outcomes.
  • Steering and suitability. Presenting products that serve the company at the customer's expense can be unfair or abusive. Broker-dealers recommending securities to retail customers are also subject to Regulation Best Interest.
  • Insurance. Many states have adopted the NAIC model bulletin on insurers' use of AI systems, which expects governance, documentation and testing for unfair discrimination.
  • Use of personal data. Offers driven by personal data must stay within privacy notices and state privacy laws, such as the California Consumer Privacy Act.

Frequently asked questions

What is a next best action?

The move, at a given point in a conversation, most likely to advance the customer toward a good outcome: a question, a response, an offer or a handoff to a person.

Why not let the language model decide?

Language models predict plausible text. They are not trained on the organization's business outcomes, and their answers can vary from run to run. A ranking model trained on outcomes makes the decision consistent and testable.

How is this different from product recommendations on a website?

The ranking idea is the same, but the setting is a live conversation. Candidates include questions and responses as well as products, and timing is part of the decision.

Does the engine need a lot of data?

Ranking needs outcome-labeled history for each type of decision. Where history is thin, the engine can start from rules or from the observed choices of top performers and let outcomes adjust the rankings over time.

Can the engine choose something compliance has not approved?

It should not be able to. Candidates come from an approved set, and the engine only ranks within it.

Let the winning move make the decision.

Encore AI pairs a ranking engine with a language model, so every decision in a conversation follows what your top performers chose when it worked.

Scale your best to be the best.

Book a Demo