The pitch right now is speed, and speed alone won't pay off
Most of what you'll hear about agentic AI marketing tools this year, and about AI in loyalty programs in particular, is about setup speed. Describe the campaign you want, and it's live. I'll be fair here, because anyone who has waited two sprints for a points promotion knows why that's appealing. And I agree with the bigger idea behind it. A loyalty program should keep changing as your members change, and faster setup makes that a lot more realistic.
But I'd push on it a little. Ron Kohavi, who led experimentation at Microsoft, and Harvard Business School professor Stefan Thomke found that at Google and Bing only about 10% to 20% of experiments produce positive results, and that "even experts often misjudge which ones will pay off" [4]. So if all we do is launch faster, we mostly roll more losing ideas out to every member.
Being able to test faster doesn't mean we should start putting half-baked tests out there.
This is where the human side matters most. AI is artificial intelligence, and it can help you design a test, set it up and pull the results. But someone still has to bring what I'd call authentic intelligence, the judgment to look at those results, understand what they're really saying about your members, and decide what the next test should be. The assistant can tell you redemptions went up in the test group. It takes a person who knows the program to ask whether those were members who would have come back anyway, and whether the reward is actually pulling in the behavior you want. I go deeper on how to set up those tests and show finance the math in this post on proving your loyalty program is working.
Why loyalty program owners wait so long for answers
As I said in my post on multi-brand loyalty, a loyalty program operator is basically a small business owner inside a business, except that when they get buried, they usually file a request with three other teams and wait.
Loyalty programs often struggle because of "under-resourced and misaligned teams," according to Forrester [5]. In a November 2025 survey of 435 marketing leaders, half said they wait one to three business days for answers to ad hoc data questions, and 45% said they often feel too rushed to analyze data properly [6].
50%of marketing leaders wait one to three business days for answers to ad hoc data questions.
45%often feel too rushed to analyze data properly.
Source: Supermetrics, survey of 435 marketing leaders, November 2025.
So what would have changed for me during that leadership changeover? An assistant doesn't get you around access you don't have, because it can only see what you're allowed to see. But the loyalty data was mine, and with an assistant connected to the platform I could have pulled reward cost by tier, redemption by benefit and spend by segment myself, that same day. And asking IT for read-only access to another system through a connection like this may be a much smaller ask than asking an analyst for their afternoon, so the answer I kept hearing might have been a yes.
Use AI for loyalty program analytics and reporting first
Robotics people have an old rule of thumb to automate the jobs that are dull, dirty or dangerous. I'd add drab, because a lot of loyalty work is just gray.
- Dull is answering the same handful of questions for five different stakeholders, like how many members joined, who redeemed and which tier moved.
- Drab is building the QBR deck, the board slide and the ELT update, and reformatting the same data three ways for three audiences.
- Dirty is chasing down why the loyalty platform says one thing and finance says another, and cleaning up duplicate members.
Most of this is loyalty program analytics, the questions and reports that eat a small team's week, and it's where an assistant can give you hours back this quarter.
The dirty work is where I think this helps most. Lee Arthur at Fospha, a marketing measurement company, shared that 48% of the marketers using its MCP in September used it at least once to check someone else's number, and he described what they wanted as "a referee" [7]. Loyalty teams have that same argument between systems all the time. So ask where the math happens. If the platform calculates the figures and the assistant explains them, every number traces back to the system of record.
Put a person on the changes that touch money, and let the assistant check your people too
The fourth category, dangerous, is any change to a live program, and I learned this one the hard way. On one program, as we were expanding internationally, we set the wrong exchange rate for points. It was a very large expansion and a completely innocent mistake. People were entering data across many, many cells and rows, and after a while your eyes start to cross. Then all of it went into the system as one mass upload in a JSON file, which is not a format most loyalty practitioners are used to reading. We caught it early, and the members who got that rate came out of it very well and were understanding about it when we corrected it, because it really was too good to be true. But it showed me how easily a person makes an error when the work is that manual.
This is an important piece to make note of. Points are a liability on the balance sheet. Milliman's actuaries looked at the three major U.S. airlines, where loyalty liabilities were larger than operating income, and pointed out that changes to those estimates "would have significant impact on the operating income" [8]. The math works the same way for a smaller program, so a points mistake shows up in finance's numbers too.
So for dangerous work, the assistant drafts the change, shows you exactly what it will do, and waits for a person on your team to approve it. Anthropic's own guidance on building AI agents describes agents that "pause for human feedback at checkpoints" [9].
But I'd be careful about the easy version of this, where a person has to approve every single thing. Researchers from Hugging Face and Data & Society call this "approval fatigue," where people stop paying close attention to what they're approving [10]. If you have to sign off on every report and lookup, you'll be clicking yes without reading by week three, and the approval that mattered will slide through. So I'd sort approvals by risk. Questions, reports and drafts can run without a sign-off, and anything that touches points, earn rates, tiers, reward costs or a member's balance gets a real review from someone who understands the money.
Sort approvals by risk| What the assistant is doing | Examples | Sign-off |
|---|
| Answering questions | Member counts, redemption by tier, reward cost by benefit | None needed, every action logged |
| Building reports and drafts | The QBR deck, the board slide, a draft offer | None until it touches the live program |
| Changing the live program | Points, earn rates, tiers, reward costs, a member's balance | A real review by someone who understands the money |
| Checking your team's manual work | A new exchange rate, a bulk upload | The assistant flags it, a person decides |
And the checking can run in the other direction, too. There was no AI anywhere near my exchange rate mistake. People made it, because people make mistakes. So something to consider is having the assistant pressure-test human work before it goes live, like comparing a new earn rate or exchange rate against the rest of the program and flagging anything that looks far more generous than it should. That way governance runs both ways, where your team reviews what the assistant drafts and the assistant reviews what your team builds by hand.
Some of the rules that protect your program may live only in the admin screen, like a cap on how many points one person can be given or a second approval on large adjustments. An assistant working through MCP never sees that screen. Shubhendu Tripathi made this point about finance software recently [11], and it applies to loyalty too. So before an assistant gets write access, try to break your riskiest rules through the connection it will use, and ask your vendor to enforce any rule that doesn't hold inside the platform itself.
Roll out the assistant the way you'd roll out a program change
BaselineWrite down reporting hours, time to answer leadership and data requests to other teams.
Read-onlyRun it for a quarter while your team still makes every change in the platform.
ComparePut the quarter's numbers next to your baseline.
DecideWith finance in the room, decide whether it earns the right to draft changes.
Before you turn anything on, write down how many hours a month your team spends on reporting, how long it takes to answer a question from leadership, and how many data requests you send to other teams. Then run it read-only for a quarter while your team still makes every change in the platform. At the end of the quarter you compare the numbers and decide as a team, with finance in the room, whether it has earned the right to draft changes to your live program. I'd also add those numbers to your loyalty program KPIs, because they show leadership how much faster you can now answer the questions they care about.
Questions to ask your loyalty platform about AI and MCP
- When the assistant gives me a number, is the math done in the platform or by the assistant?
- Can it only read data, or can it change my program? Which changes, exactly?
- Can I require approval for the changes that matter, like points and tier rules, without approving every lookup?
- Does the assistant inherit my user permissions, and are limits like points caps enforced in the platform when a change comes through it?
- Can the assistant check a change my team made by hand before it goes live?
- Is there a log of every action, and is it off by default until someone in my company turns it on?
- What member data does the assistant see, and where does it go?
So this is something I've been close to for a while. Our product team has been building the TrueLoyal MCP, and read access is now in beta with a group of our clients, with the ability to make changes coming next. If you're a TrueLoyal client and want to join the beta, or you're working through these same questions with your own team, I'd love to talk it through with you.
Frequently Asked Questions
What is MCP in a loyalty program?
MCP (Model Context Protocol) is an open standard that lets AI assistants like Claude, ChatGPT and Copilot connect to business software, including loyalty platforms. Depending on how the vendor built it, the assistant can read program data, make changes to the program, or both. It's the connection behind many of the AI loyalty program tools launching this year.
How is AI used in loyalty program management?
AI assistants can answer questions about members and program performance, build reports for business reviews, help reconcile numbers between systems, check changes before they go live, and draft program changes for a person to approve. For loyalty teams, the biggest benefit is getting to the answers leadership asks for, on the data you already have access to, without waiting on other departments.
Can AI help with loyalty program analytics?
Yes. With an MCP connection, an assistant can answer questions about members, tiers, redemptions, spend and reward costs in plain English, pulling from the source system, without filing a data request.
How do I keep AI changes to my loyalty program safe?
A safe setup gives anything that touches points, earn rates, tiers or reward costs a real review by a person, limits the assistant to the same permissions as the person using it, enforces limits like points caps in the platform itself, uses the assistant to check manual changes like a new exchange rate or bulk upload before they go live, keeps a log of every action, starts with the feature switched off, and checks your contracts and data agreements before any member data is connected.
What loyalty program KPIs should I track when using AI?
Keep your usual program KPIs, like active members, redemption rate and repeat purchase rate, and add a few that show what the assistant gives back, such as hours spent on reporting, how long it takes to answer a question from leadership, and how many data requests you send to other teams.
Sources
- Anthropic, "Introducing the Model Context Protocol," Nov 25, 2024.
- TechCrunch, "OpenAI adopts rival Anthropic's standard for connecting AI models to data," Mar 26, 2025.
- Anthropic, "Donating the Model Context Protocol and establishing the Agentic AI Foundation," Dec 9, 2025.
- Ron Kohavi and Stefan Thomke, "The Surprising Power of Online Experiments," Harvard Business Review, Sept–Oct 2017.
- Forrester, "How To Build A Loyalty Team That Scales With Your Program."
- Supermetrics, "The AI readiness gap," survey of 435 marketing leaders fielded November 2025, published July 14, 2026.
- Lee Arthur (Fospha), LinkedIn post on how marketers used the Fospha MCP in September, Oct 13, 2026.
- Milliman, "Loyalty programs: The crown jewels," Oct 7, 2021.
- Anthropic, "Building effective agents," Dec 19, 2024.
- Margaret Mitchell, Avijit Ghosh and Samir Passi, "AI Agents Push Humans Out of the Loop," arXiv 2608.23642, Sept 2026.
- Shubhendu Tripathi, "Headless ERP Removes the Screen. Some of Your Controls Only Existed There," The Integration Layer (LinkedIn newsletter), 2026.
About the author
Amanda Boshell is a Product Marketing Manager at TrueLoyal. She spent twelve years as a loyalty practitioner before moving to the platform side, across agency, tech, and retail, in roles spanning program operations, strategy, and marketing. She has sat on both sides of the table, as the operator running the program and as the buyer evaluating the platforms.
