Depending on who you talk to, we’re heading into the fourth or fifth year of LLMs hitting the mainstream.
Which makes it a great time to reflect on where AI truly does “take the boring, repetitive tasks off your plate so you can focus on your best work” and where it ends up creating more work for you and your team in the long run.
Getting a sense of what’s working for other people can help you figure out what’s worth automating and what’s not.
In her role as Senior Manager of Affiliate Programs at Manychat, Alessandra Silveira has used AI to help transform its partner program, including a review that removed more than 25,000 fraudulent and inactive accounts. She’s also learned plenty of lessons the hard way.
She sat down with us to share how she uses AI in her program today, her best tips (and no-nos) and what she hopes to use AI for in the near future.
What to use AI for in your partner program
If you’ve got a huge volume of partners, you’ve got a huge load of work on your plate. Particularly if you’re short-staffed, like so many partner programs are.
With so much to do, it’s easy for things to fall through the cracks. Silveira’s biggest motivation for using AI came when her team realized that nine of their top 10 partners were fraudulent.
“That prompted us to investigate the program more deeply, and AI became a turning point because it gave us a way to verify partner activity at a scale we couldn’t manage before,” she explains. “AI supported our investigation and gave us stronger evidence for our decisions.”
Since then, she’s used AI to:
- Surface suspected fraud
- Weed out bad-fit applicants
- Research potential partners
- Track progress (and get suggestions)
- Highlight inactive partners
- Audit commissions
Surface suspected fraud
After finding out that most of Manychat’s “top” partners were a scam, Silveira’s team built a detailed fraud playbook, outlining exactly what they saw that made them do a double-take.
It identifies eight types of behavior the team considers suspicious, along with the patterns they see in daily applications, which gives AI a clear benchmark for what’s normal and what isn’t.
The team uses AI to flag potential fraud and suspicious behavior against that playbook. They don’t run formal accuracy tests on the analysis. They just keep refining the playbook as they go.
Along with a broader program review, the playbook has helped her team remove a whopping 25,000-plus fraudulent and inactive accounts.
Weed out bad-fit applicants
PartnerStack’s AI beta gives Silveira a first pass on new applications.
It doesn’t score applicants. Instead, it passes through the ones it thinks are a good fit and tags the rest to exclude. She can still check the excluded ones if something catches her eye.
Without it, she’d have to scan through all of the applications manually.
See more: Red flags to watch for when recruiting partners for your program.
Research potential partners
Silveira has a pinned chat with a specific prompt in ChatGPT that she uses to scan “maybe” applications: ones in the PartnerStack pile that she wants to do some extra due diligence on.
She pastes in the application contents, and ChatGPT spits out a score and a recommendation: approve, approve and monitor or reject the partner outright — along with specific reasons based on its independent research and interpretation of the applicant’s answers.
Equipped with that background, Silveira can make a much quicker yes/no decision.
The workflow also means less time on repetitive operational work and more time for relationships with affiliates, especially her top 10. She uses that time to understand how she can help their businesses grow.
After all, if they succeed, she probably will too.
Track progress (and get suggestions)
You never know when you’ll get that dreaded, “Hey, can you give me a status on the partner program overall? Like the main big wins lately? The CRO’s asking for an update before her meeting at the top of the hour” Slack.
You know that CRO isn’t going to want a qualitative answer.
She’s looking for numbers. If those numbers are good, she’ll want to know what you’re doing to keep them that way. If those numbers are bad, she’ll want to know your plan for improving them.
To avoid a mad scramble and stay on top of her program, Silveira’s set up a daily check-in with Claude, using PartnerStack’s MCP to pull in her partner data. It reports on her paying accounts, tells her how she’s tracking against her goal and suggests where to focus her time to hit it.
That way, she isn’t waiting on numbers she can’t easily see, and she gets practical tips for hitting her targets.
Highlight inactive partners
If a partner hasn’t brought in a new account in months, they probably won't reel in a whale anytime soon, and they’re expensive to keep around.
But digging through your partner base to find them is highly manual. So Silveira has Claude do it for her.
Periodically, she’ll export her partner performance data from PartnerStack, upload it to Claude and ask which partners haven’t brought Manychat a deal in the past three-plus months. Then she tags and excludes them.
She uses the same export to ask Claude other questions, like which partners are declining and who’s performing best over a given period.
Clearing out inactive accounts was part of the same broader review that removed fraudulent ones, which Silveira says improved the program’s overall health. Manychat also cut the channel’s CAC by 42%.
You might also like: Expert tips and tricks to revive inactive channel partners.
Audit commissions
A wrong attribution or double payout can cost you big time. Not just in the money you lose, but in the goodwill of a partner who feels shortchanged.
Before, Silveira prevented this by double-checking everything by hand, combing through commissions and invoices. Now, she uploads them to Claude and asks it to find anything that looks off.
Catching an error before it goes out is way easier than trying to claw back money later, and it’s another way she’s keeping her CAC down.
4 dos for adopting AI in your partner program
The six main use cases Silveira relies on today weren’t some one-shot setup. It took a lot of massaging to get these workflows to work the way she wanted them to.
Her experience points to a few ways you can save yourself time and frustration as you hand more over to AI:
1. Starting with a well-established manual process
This might sound like a “duh” thing, but when you’re handed a new task, it’s really tempting to just clack away on your keyboard, shoot off a prompt and see what AI can do.
That can be a fun experiment, but you’ll get the most trustworthy results if you’ve worked through the problem on your own. You know what “good” looks like, you know the right metrics to track and you know about weird scenarios you might need to account for.
Take Silveira’s fraud playbook.
Her team had already mapped out the eight types of suspicious behavior and the patterns they see in daily applications, so AI had something concrete to check partner activity against. Without that groundwork, there’d be nothing to measure against.
2. Doing a bunch of testing
We all know at this point that AI is pretty confident even when it’s pretty wrong. Which is why you need to pressure test it — a lot.
Originally, Silveira’s team brought in an outside developer to build an AI agent for them. At first, it looked fabulous.
It scored each application and estimated fraud risk. If it couldn’t tell whether a partner was real, it flagged them for manual review.
“So great, right?” Silveira says. “But it was not working at all.”
The team caught the issues during a fraud review, and the deeper they dug into the workflow, the more problems turned up.
Especially with partner-facing activity
When it came to communicating directly with partners, the agent’s performance went downhill.
It was designed to personalize partner outreach, telling affiliates how many accounts they’d referred, how much they’d earned and what their potential earnings could be.
But while Silveira was asking one partner to verify their activity because of potential fraud concerns, the agent sent them an upbeat sales message about how much more money they could make.
“The messages contradicted each other and made the conversation confusing,” she says. “The agent wasn’t distinguishing a verification conversation from a commercial or activation conversation.”
And with token limits
Failures aren’t always about behavior. They can be about capacity, too. Sometimes, AI just can’t finish the job.
Recently, Silveira asked Claude to tag 245 inactive partners for exclusion. It came back saying it had only gotten through 100 and asked whether that was OK. She didn’t have enough tokens for the rest.
You might need to ask IT to bump up your allowance, tighten up your prompts or accept the ceiling and finish the rest of the workflow by hand.
3. Sharing ideas with vendors you already use
Like you, vendors are being asked to use AI in their day-to-day, but they’ve got even more pressure to build AI into their tool. Giving them good ideas is a win-win: they get real-world feedback, you get features that make your job easier.
That’s how Silveira ended up in PartnerStack’s AI beta. She showed the agent to PartnerStack’s product team and gave them her honest take: it wasn’t working for her, but the idea was amazing. As it turned out, it was the kind of thing they wanted to build, too.
They started working together from there. Today, Manychat is testing the beta, and Silveira shares regular feedback with PartnerStack’s product team as new capabilities take shape.
4. Being picky about what you automate and how
You don’t need to use AI for everything, nor should you.
Silveira estimates that about half her work is automated today, which has helped her cut manual tasks while still maintaining strong relationships with her partners. She’d like to hand off more, but her team’s experience has made it clear where AI doesn’t belong.
“The experience showed us that sensitive conversations need strong context awareness and clear boundaries,” she says.
Where AI does shine is the operational side. “The best part is being able to assess partners faster and with more confidence, analyze commissions more efficiently, and spend less time on repetitive work,” she explains. “That gives me more time to focus on relationships with affiliates.”
As for how she automates the parts she does hand off to AI, Silveira doesn’t lock herself into one tool. Not only does that help with token management, it also lets her play each model to its strengths. After all, models are changing all the time.
In practice, that means using PartnerStack’s built-in AI features to manage partner data, Claude for partner performance and commission analysis and ChatGPT to score potential partner applications.
Related: The AI maturity curve: Where partner teams actually stand in 2026.
3 don’ts for adopting AI in your partner program
On the flip side, here’s what to steer clear of to save yourself a headache (or two):
1. Automating everything at once
Instead of big-bang-ing your AI project, try taking one step at a time. That gives you plenty of room to test and tweak.
Silveira’s taking that approach with her program as a whole, moving in waves.
Wave one was cleaning up the partner base and getting to know who’s in it. That turned out to be eye-opening: many of Manychat’s partners aren’t content creators, but educators and consultants, who need a completely different approach.
Wave two is figuring out the best way to reach them. That starts with a new email onboarding flow she’s testing this quarter and the program’s first partner newsletter.
2. Letting AI completely speak for you
Sending AI slop is no way to establish a relationship with new partners, or keep a healthy one going. It reads as “I don’t care.”
Silveira uses AI to help respond to standardized partner messages. But for anything more nuanced, she’s seen how quickly an automated message can go sideways. And she wants the content she sends partners to sound like her team, in their own tone of voice.
Make sure any communication coming from you represents your company and brand and reflects the true value of the partnership. Until a tool can reliably sound like you, you should have the final say on anything partner-facing.
3. Forgetting to do a before-and-after analysis
Without tracking what you did before, what numbers you got and how long something took, it’s impossible to say (with any precision) how much AI is helping you and whether you should keep using it.
That’s important. Not only in terms of knowing how to prioritize your daily activities, but in terms of unlocking partnership budget. Asks like “I need more budget for a new tool,” or “This can’t really be automated, so I need another partner manager for my team, and here’s why,” need to be backed by hard numbers.
What you should think about using AI for next
With about half her work automated, Silveira has more time for developing content, meeting new partners and working closely with her top 10 affiliates on things like events and the occasional co-marketing project.
But she’d like to get even more time back. Ideally, her AI workflows would be more personalized and built right into PartnerStack. On her wishlist:
- Personalized content by partner type. Wave one showed just how different Manychat’s partners are. Her dream is to segment them in one click and send each group content and flows built for them.
- Answering routine partner questions. This one’s going to need some testing and finesse, especially after her experience with the previous agent. But questions like “how do I get paid?” and “what counts as a referral?” are a natural place to start.
- An agent that works like part of the team. Over time, she’d like to train an agent that understands Manychat’s program well enough to act “as an extension of my team,” handling repeatable tasks so she can focus on affiliate relationships.
Most of what’s on Silveira’s wishlist (maybe yours, too?) comes down to spending less time on routine tasks and more time with partners.
PartnerStack was built to jumpstart that journey, with AI that helps you find the right partners, writes on-brand recruitment emails, captures leads from email and Slack, and connects your partner data to the AI tools you already use through the PartnerStack MCP.
Book a demo to see what PartnerStack can do.
