A member opens the app at lunch. The program has about a second to decide what to show, such as a bonus on today’s order or no offer at all. That choice is real-time decisioning.

What is real-time decisioning?

Real-time decisioning is choosing the next action for one person at the moment they act, using live data, rules and models. The action can be an offer, a message, a price or a block. SAS shows the parts in a telecom case. A customer hit the monthly data cap early. The decision point was an offer to top up the allowance. It posed three questions: whether to make the offer, how big to make it, and what it cost next to the overage the customer would pay (SAS).

The case also shows the trade. A rule that fires on one trigger for everyone is fast but blunt. Overnight analysis is sharper but slow. SAS proposed combining the two: models built on real-time data, run when a customer nears the cap. It reports about 400 percent more responses than the company’s existing methods (SAS). That is one vendor’s account of one company. It shows the shape of the gain.

Decisioning picks the action. A rules engine applies fixed logic, and a decisioning layer adds scoring and choice on top. Start with rules and add models where the data is rich. Rules are easy to explain and safe to change. Models earn their place when many offers compete for one slot and the rules cannot say which wins. Compare each model with the rules it replaces.

How does real-time decisioning work?

A decision engine runs the same loop for every request. The table lists the steps with a loyalty example.

StepWhat happensLoyalty exampleWhat goes wrong
EventA signal arrives: a login, a scan, a purchaseMember opens the appLate or duplicate events
ContextThe engine pulls the member’s profile and recent activityTier, balance, last visitStale balances
EligibilityRules remove actions that are not allowedOffer expired, consent withheld, daily cap reachedRules that conflict
ScoringModels rank what is leftChance the member redeems each offerScores built on old behavior
SelectionOne action winsBonus on today’s orderAlways showing the same winner
LearningThe result feeds the next decisionRedeemed or ignoredNo holdout to prove the lift

Guardrails sit in the eligibility step. Cap how many offers one member sees in a day. Set a floor on margin, so a model cannot pick a loss-making offer. Respect consent before anything else. Keep a small share of traffic for exploration, so the model keeps learning.

The learning step has a known method. Li and colleagues treated news recommendation as a contextual bandit. The algorithm picks an item for a user from the context, then adapts from the clicks. On a Yahoo! dataset of over 33 million events, it gave a 12.5% click lift over a standard context-free bandit. The authors also argue that any bandit algorithm can be tested offline on previously recorded random traffic (Li et al., 2010). The lesson for a program is to keep some random traffic, so the models can be checked.

How fast does a decision need to be?

As fast as the moment it serves. Loyalty Juggernaut’s own guide splits the work. Earn posting, redemption checks, fraud holds and the checkout offer run in real time, because a delay is visible to the member or costly to the program. Tier reviews, expiry runs and statements run in batch (Loyalty automation). A fraud hold that lands after the points are spent protects nothing.

Set a time budget for each decision. Count every step from the event to the answer: the data pull, the rule checks, the score and the delivery. Talkdesk lists real-time analysis as a core stage of orchestration (Talkdesk). If the total is too slow, drop the slowest step first. A late answer is the same as no answer.

Speed pays at the shelf too. A study commissioned by Google and run by 55 and Deloitte covered 37 brand sites and over 30 million sessions. For retail, a 0.1 second gain in speed brought a 9.1% rise in shoppers moving from product page to basket, and 9.2% more spend (Google and Deloitte, 2020). That study measured page speed. It did not measure decisioning. It shows how much a few tenths of a second matter at the point of choice.

What limits automated decisions?

The GDPR gives a person the right not to be subject to a decision based solely on automated processing. That covers profiling too, where the decision has legal or similar effects (GDPR Article 22). A choice between two discounts rarely reaches that bar. A choice to close an account or forfeit points is closer. Ask counsel where the line sits in each market.

Good practice goes further than the law. Keep a log of every decision with the inputs and the rule or model that chose. Give members a way to see why an offer appeared. Route high-impact actions to a person. Our guide to loyalty automation sets the same split between what software does alone and what needs sign-off.

How do loyalty programs use real-time decisioning?

Programs decide when a member earns and when a member redeems. The browse screen is a third point. Ulta Beauty says its loyalty programs give it a deep view of its guests and what they like. That lets it tailor experiences, recommendations and promotions (Ulta Beauty 10-K). Kroger says more than 20 years of work in data science let it use loyalty data to build personal experiences (Kroger 10-K).

The redeem moment is the least used. A decision there can check availability, test for fraud and set a points price. Our entry on dynamic award pricing covers the last of those.

Proof is the weak spot. IAB guidance for retail media points to randomized controlled trials, synthetic controls and matched-market tests for lift (IAB/MRC). The same tests fit a loyalty decision engine. Hold out a random share of members. Serve them a default, and compare the results.

For illustration, a program decides which of three offers to show 100,000 members a day. The model’s choice yields 6.0% redemption. A random 5% holdout, shown a default offer, yields 4.8%. The lift is 1.2 points, or 1,140 extra redemptions a day across the 95,000 members the model served. The program can then price that gain against the cost of the offers. Report the holdout size next to the lift.

How GRAVTY handles real-time decisioning

Loyalty Juggernaut says GRAVTY connects to booking, check-in and other systems through open APIs and real-time events. Average API latency is under 50 ms (GRAVTY for airlines). That is the speed a decision has to live inside.

On lji.io, GRAVTY’s agentic systems share one set of facts about members, the program, partners and the economics. They act within rules the enterprise sets. One picks the next-best offer, benefit or experience for each member. AI-Trust flags unusual earn and redemption behavior as it happens. Business teams change rules in GRAVTY Visual Rules, a patented no-code engine, without a release.

Frequently asked questions

What is the difference between real-time decisioning and batch decisioning?

Real-time decisioning chooses at the moment of the event, and batch chooses on a schedule from stored data. Loyalty Juggernaut puts earn posting, redemption checks, fraud holds and the checkout offer in real time (Loyalty automation). Tier reviews, expiry runs and statements run in batch. Each job needs its own timing.

What data does real-time decisioning need?

It needs the live event and a current profile of the person. In the SAS telecom case, the data was how far each customer was into the monthly allowance. It also covered the cost of an offer against the overage charge (SAS). For a loyalty program, the key items are balance, tier, recent activity and consent.

How do you test whether real-time decisions work?

Keep a random holdout and compare. Li and colleagues say a bandit algorithm can be tested offline on recorded random traffic (Li et al., 2010). The IAB guidance for retail media lists randomized controlled trials and matched-market tests for lift (IAB/MRC).

Is real-time decisioning the same as next best action?

Next best action is the choice, and real-time decisioning is the machinery that makes it as the member acts. See our entry on next best action for the choice itself. On lji.io, one GRAVTY agentic system picks the next-best offer, benefit or experience for each member.

Do automated decisions fall under privacy law?

Some do. Under the GDPR, a person has the right not to be subject to a decision based solely on automated processing. That applies where the decision has legal or similar effects (GDPR Article 22). A choice of offer rarely reaches that point, but closing an account is closer. This is not legal advice.