The End of the Loyalty Programme

In 1992, Hoover offered UK customers two return flights to the USA with any £100 appliance purchase. More than 200,000 people took up the offer, many buying vacuum cleaners they did not want simply to claim the tickets. The promotion reportedly cost the company £48 million. [1]

Hoover is the extreme case of a loyalty decision whose consequences became impossible to ignore. The risks in modern programmes are usually harder to anticipate.

Modern loyalty programmes depend on thousands of interconnected choices. What should the rewards be? How generous? Who qualifies? Which customers should receive an offer, through which channel, at what moment and at what cost?

None has a single correct answer. These human decisions, more than the loyalty platforms beneath them, are what separate one programme from another. They are where the competitive edge is created and where the risk accumulates.

The public record shows several ways that risk can surface. In 2016, Chase launched the Sapphire Reserve card with a 100,000-point sign-up bonus. Demand ran so far past forecast that the bank ran out of the metal used to make the cards. The sign-up bonus cost the bank roughly US$200 million from that quarter’s profit. [2]

In 2018, SoftBank launched PayPay, a payments app, across Japan. The PayPay team allocated 10 billion yen (roughly $90m) in cashback budget to fuel user acquisitions for four months. The money was gone in ten days. [3]

In 2022, Wells Fargo backed a card that rewarded rent payments. The bank was later reported to be losing as much as US$10 million a month after cardholder behaviour and programme economics diverged from the original model. [4]

Chase, PayPay and Wells Fargo employ some of the most capable loyalty teams in the world. That is precisely the point. These were reasonable calls built on defensible forecasts. The flaws stayed invisible until millions of customers behaved in ways the spreadsheets did not expect.

Our own work with banking, fintech and telco partners is consistent with this. We have watched similar bets play out in quieter forms. Programmes that exhausted their budgets rewarding customers who would have acted anyway. Cashback benefits with spiking redemption costs. The pattern is the same.

That is the reality behind many loyalty programmes. Enormous human effort is poured into thousands of quiet bets about customer behaviour.

Being good at the job is no longer enough to get every judgement call right. The natural response is to ask for more data. In our experience, that helps. But it also exposes the next problem. Every new signal creates more possibilities, more questions and more decisions for someone to make.

For years, the constraint in loyalty was data. There was never enough of it, or what existed was inaccessible.

That constraint is fading. Most organisations we work with can now connect customer product holdings, balances and transactions with app engagement and lifecycle events. This should make loyalty easier to run. But as organisations seek to act on more of those signals, it multiplies the decisions a team must make.

We see this dynamic play out clearly in banking. Many banks want to expand their credit card loyalty programme into a whole-of-bank loyalty programme that drives cross-sell and usage across the entire retail banking product suite – from deposits and mortgages to wealth and insurance. This strategy has an excellent business case. Accenture estimates that banks which deepen their primary customer relationships can lift revenue from those customers by up to 20 per cent. [5] A whole-of-bank loyalty programme is the most direct way to achieve this.

But consider what that ambition does to the data and decisions facing the loyalty teams tasked to design whole-of-bank loyalty programmes. A typical credit card loyalty programme already involves around forty programme design decisions. Grant each one just a handful of options and the theoretical number of configurations for a credit card loyalty programme exceed 1 nonillion. That’s 10³⁰ — more than all the stars in the observable universe. [6] Layer in loyalty mechanics for six more banking products; the possible configurations of a whole-of-bank loyalty programme exceed 1 googol. That’s 10¹⁰⁰ — an unimaginably large number that none of us have ever heard of. Have a play with the decision tree below to see how layering in more product logic within a loyalty programme impacts the possible universe of loyalty programme configurations.

The figures in our model are illustrative. Not every theoretical combination would be viable. The point is that every new decision multiplies the possible paths, while a human team can examine only a tiny fraction of them. Of course not every programme faces this decision volume day to day. A stable programme can largely be set and left to run; the decision problem grows as organisations move towards more frequent campaigns, granular personalisation, and signal-responsive engagement.

Now ask yourself what the chances are that any team, of any size, can make sense of all of that data and configure a loyalty programme that is well optimised. Yet this is the exact mission loyalty teams in hundreds of banks have been handed.

More data creates more places where judgement must be applied. Every new data signal introduces more possible actions. This is true not just in banking, but also for loyalty teams across telcos, airlines, retailers and any other enterprise with a large and growing customer dataset.

The bottleneck currently is in the ability to turn a growing volume of information into timely, commercially sound decisions. The problem with the operating model of loyalty is that it asks human judgement to work beyond human scale.

The question is no longer how to help a team make better decisions. It is what kind of a system can search a space this size continuously, within goals and boundaries set by people.

Loyalty systems have historically been good at execution. They calculate points, apply rules and send offers. What they have not done is decide. People set the objective, choose the audience, design the reward and define the limits. The system carries out the instruction.

AI agents move the division of labour between loyalty teams and loyalty systems. An agentic loyalty system has four defining characteristics:

The first three reflect what strong loyalty teams already do. The difference is the scale and speed of the learning loop. A team can review a limited number of campaigns. A system can learn from every intervention across millions of customers.

Put simply, agentic loyalty is customer loyalty run by AI agents that perceive continuously, pursue goals set by the programme owner, act within governed limits and learn from every outcome. It is a system people direct rather than operate.

Consider an organisation trying to reduce disengagement among customers who hold several products.

Today, a team might analyse the data, define an audience, agree an offer, secure approvals and launch a campaign. In Pulse’s client work, this process can take up to 20 weeks and every customer in the segment may receive roughly the same response.

An agentic system works differently. It observes the signals continuously, identifies which customers appear to be disengaging and works towards a goal the bank has set, within a fixed budget and clear eligibility rules.

For one customer, the right action may be a service prompt. For another, it may be a merchant reward. For a third, it may be no action at all. Costly or sensitive decisions can be escalated for approval.

As outcomes arrive, the system measures what changed. It learns which signals matter, which interventions work for whom and where an incentive would have been wasted. That learning shapes the next decision, not the next quarterly planning cycle.

The same logic applies across loyalty. A campaign that took weeks to assemble can become a continuous decision process. A reward that once drained a budget can be withheld from customers likely to act anyway. Any data stream can become another input.

The questions have not changed: who needs action, what might influence them, what is that change worth, and what should happen next?

What changes is how often, and how precisely, they can be answered.

The loyalty programme decisions that a team once placed under pressure become millions of small, governed decisions made continuously. AI agents will enable loyalty to operate at the scale of the problem.

If agentic loyalty changes how decisions are made, it also changes the role of the loyalty team. The role of the loyalty team does not become smaller. It becomes more strategic.

Most loyalty professionals did not enter the field to spend their time configuring reward rules, pulling lists and building segments. Those jobs are necessary, but they often crowd out the work that matters more: deciding which customers to prioritise, what behaviour the business should influence, what value it is prepared to exchange and what is fair to the customer.

We see this in client teams today. Much of the week is consumed by getting campaigns out of the door. The harder questions are pushed to the edges because the mechanics come first.

An agentic model reverses that balance.

The team sets the goals, budgets, guardrails, and trade-offs. The system handles more of the high-volume decisions within them. The role moves from operating the programme to directing the system.

That also changes the unit of work. Today, many loyalty teams run on a campaign calendar: brief, build, launch, review. But customers do not experience loyalty quarterly. They experience it whenever they open an app, make a purchase, miss a payment or consider switching.

In the new model, the system runs continuously towards goals the team has set. The team watches performance, intervenes when something drifts and resets the strategy as priorities change. The operating rhythm moves from a campaign calendar to something closer to a control room.

A system that makes more decisions does not remove human responsibility. It increases the need to define it clearly.

People own the goals and the limits. Agents handle more of the high-volume choices within them.

Goals, budgets, brand promises, fairness and eligibility should remain human decisions. People decide what the organisation is trying to achieve, which trade-offs are acceptable and where the system must stop.

The agent operates inside those boundaries. It can choose among approved actions, work within set thresholds and escalate decisions that carry greater financial, customer or reputational risk.

That is especially important in regulated industries. Introducing a new customer journey can require legal, risk, product, marketing and brand approval for good reason. Agentic loyalty does not make those controls disappear. It moves them upstream, into the rules and limits that govern every action.

Every decision should be logged, explainable, reviewable and reversible. The business should be able to see what the system did and why. Regulators should be able to examine the decision process. Customers should receive a clear explanation when an outcome materially affects them.

Personalisation must also remain prominent. A loyalty brand can promise to recognise customers and reward them fairly without giving everyone the same journey. The promise stays consistent even when the mechanics differ.

Adoption will be gradual. Agentic loyalty systems will take on high-volume, lower-risk decisions first. People will keep control of decisions that carry greater weight. Different organisations will draw that boundary in different places and move it over time as confidence grows.

The future of loyalty is not a system left to run itself. It is a system capable of making millions of decisions, directed by people who remain accountable for the goals, limits, and consequences.