Building an AI-First Culture: How Signifyd Empowers Teams

Maybe it’s no surprise that Signifyd is big on AI. The company’s value to its customers and the solutions we offer are built on it.

But as important, so are the work days of our employees. “AI-first” culture is not a buzz phrase at Signifyd or feel-good chatter to assure the troops that company leaders are with it and not watching as technology’s latest social and economic transformation passes them by.

“I don’t start the day without AI anymore,” says Aimee McAllister, a Signifyd lead chargeback analyst working in Belfast. “It’s become an absolute necessity in the day to day.”

While the national narrative has ricocheted between “gen AI as economic boom” and “gen AI as societal doom,” Signifyd has built an AI-first culture at the same breakneck speed at which AI technology is advancing.

Teams are trying, testing and measuring various LLM platforms and tools and the workflows they can revolutionize. A Slack channel of hundreds of employees has blossomed where users share successes, failures, what-ifs and why-nots.

Gen AI acts as a mundanity crusher and innovation inspiration

Software engineers and developers have turned coding chores over to AI assistants, chargeback analysts have diverted repetitive tasks to agents, fraud and risk team members and machine learning experts have bridged their worlds and more rapidly discovered patterns that benefit both and Signifyd customers.

And in the process, they’re finding ways to save millions in the cost of delivering Signifyd’s commerce protection technology and services.

“I’m looking at what AI is doing today and projecting ahead a year from now,” says Signifyd co-founder Mike Liberty. “No matter where you stand on AI, the fact is, if you’re not adopting it, your competitors will. And if they’re increasing their productivity, then you’re going to be left behind.”

Signifyd co-founder’s mission includes advancing with AI

How serious is Liberty about encouraging Signifyd employees to experiment and apply AI to their existing tasks and to new ones they hadn’t contemplated? When he returned to Signifyd in 2025 as chief risk officer after a brief hiatus, he did so on the condition that adopting an AI-first approach would be a top priority.

“I’m really trying to increase adoption as much as possible,” he says, “not through mandates, but through removing blockers, whatever those blockers might be.”

Liberty calls it creating the “paved path” — doing things like signing contracts, quickly approving licenses, configuring computers, connecting existing applications and getting out of the way, while still being there to offer support.

“It has tremendous potential,” he says of today’s AI systems. “But it’s not a turnkey set up. So, you do really need to focus on that initial experience and simplifying that. Because otherwise, they’re going to try it and become disenchanted with it.”

If the mission is not yet fully accomplished, disenchantment has clearly been avoided. Let’s consider the experience of a few Signifyers who have put gen AI to use to improve work life for themselves and their teams.

The machine learning director

Harry Simon and his team of data scientists know how vital precision is in the business of developing and training machine learning models that instantly distinguish between fraudulent and legitimate online orders. The margin of error comes with a direct effect on a different margin — profit margin.

“That means when we push out a model, we have to make sure that, one, it’s the right model. It’s not broken in any way. And that the expected improvements are going to drive some good business outcomes,” Simon says. “Because a model that is 1% better could mean millions of dollars saved. A model that is 1% worse? Millions of dollars lost.”

Signifyd works at a scale of thousands of merchants and billions of transactions. Our machine learning models are large and complex. Operating at that scale and complexity while staying ahead of sophisticated and constantly evolving fraud rings requires non-stop innovation on Signifyd’s part.

Large language models plow through old limitations

So, Simon and team conduct hundreds of experiments each month, testing new machine learning features, refreshing models to keep in front of new fraud patterns and malicious attempts at working around defenses — all while remaining conscious of the high stakes.

“And that’s where you need to treat your experiments like code,” Simon says. “You need to scrutinize your experiments with the same rigor that you would scrutinize regular code.”

Without intelligent automation, that need for intense scrutiny can be a limiting factor. A human can only review so much code so fast. But gen AI, properly trained and instructed, can obliterate those limits.

The data science team has built a gen-AI-based experiment review agent, dubbed Signifyd Model Intelligence, that unlike any one human, can rely on all the historical and team knowledge about successful and flawed models as it goes about reviewing new, experimental models.

“You can use these experimental review agents to pull in all the context of how exactly our modeling system works,” Simon says. “Everything we’ve tried in the past. What has worked and what has not worked. It serves as this super-intelligent thought partner when you’re running experiments.”

AI agents keep up instantly and constantly

Think of it: Unlike humans, AI agents are able to keep up with thousands of merchants instantly and constantly. AI agents can sort through inputs and outcomes, connect findings to the Signifyd Model Intelligence layer and understand what to do next, propose improvements and write machine learning features or run machine learning models for the team.

Yes, the human is still in the loop. But now, thanks to what Simon calls “the super computer of all Signifyd model intelligence,” the human is working at a much higher and more efficient level. The AI-driven review agent allows team members to write project requirements the way a product manager would and then review the resulting code the way a senior engineer would, Simon says.

Patterns are key in detecting and preventing fraud. But no human can be all-seeing all the time. So, Simon’s team has built models that gather observations from Signifyd risk intelligence experts and models that can sort transactions by product categories. Combining the collective knowledge gathered and organized through LLMs means the team can create more granular predictions, which can better inform the models that sort legitimate transactions from fraudulent ones.

Finding new ways to work is a team focus

The exploration and innovation around LLMs has become so ingrained, Simon says, that his team creates weekly challenges that push the boundaries of deploying AI to tackle everyday tasks and routines.

“I’ve started doing AI weeks with my team where the specific goal is to see how complicated a task you can get the AI to do for you,” Simon says.

Maybe the goal for a week or two would be to not hand write any code, to “try your best to make the AI work for you.” It encourages team members to break big tasks into smaller tasks — tasks that are small enough for AI to competently tackle, while the humans review results and look for other large tasks to break down.

“We’re seeing big productivity gains with it,” Simon says. “We’re seeing people on our team being able to do more. They’re being able to stretch into new systems that they haven’t really been able to do before.”

The lead chargeback data analyst

Lead Chargeback Data Analyst Aimee McAllister lists the availability of and access to gen-AI tools among Signifyd’s most valuable employee benefits.

“No other company is going to offer me the same resources that Signifyd does,” she says. “The things I can learn by using these tools are incredible.”

McAllister spends her days managing and digging into chargebacks — disputes that seek to reverse a consumer’s credit card charge based on a real, perceived or concocted shortcoming. Signifyd handles the disputes for merchants, including presenting the case to overturn an unwarranted refund based on specific evidence.

The voluminous evidence needs to be in a standard form to be efficiently processed by Signifyd’s systems. But the thousands of merchants Signifyd serves don’t provide information in the same, standard form.

Making sure everything is squared away, and editing that which isn’t, can take up a lot of a human investigator’s time. But McAllister and her team don’t rely on humans to do the work.

“What AI has helped us do is create these micro-apps or scripts to do really repetitive tasks over and over again.”

AI brings benefits beyond speed and volume

The team has automated the process of building representment packages, essentially the merchant’s case to have a chargeback overturned. Chargeback investigators still need to review and assess the packages, but because they are not creating them, each investigator can review more packages and review them faster.

The time savings is obviously valuable. Not as obvious is the value that AI has brought to the team in terms of accelerating learning that helps team members do their jobs, McAllister says. It’s not just that the gen-AI tools work; it’s that they teach humans how they work as they do.

“I didn’t realize how much you could continue to learn whilst having a tool do something for you. So AI is obviously building out a lot of this. And we’re using plain English to ask it to build this out. But whilst you’re doing it, you’re really understanding the architecture of things. Like, yes, I wanted to do one thing, but it needs to do four different things before it gets to that point.”

And the learning doesn’t stop there. Each team member now has access to the collective knowledge of the entire team through what is essentially an AI answer engine.

“I was sort of like a knowledge hub for things,” McAllister says. “And now the knowledge is getting spread out more, so people aren’t dependent on just me for answers anymore. So it’s kind of eased up my day-to-day as well.”

The chargeback analyst

Chargeback Analyst Aldo Díaz works closely with McAllister and shares her enthusiasm for the educational potential of LLMs.

“Every time I want to understand something, or learn something, I just go with Claude,” he says. “Whenever I research something, I iterate a lot and I ask lots of questions.”

Claude’s summaries are helpful, he says. But the real gold is in the sources from which the LLM creates the summaries.

“I can go to the sources and start learning and building my own knowledge base.”

For instance, Díaz recently wanted to bulk up his peer review skills — the process of reviewing code written by others. With the help of Claude he created his own guide to best practices for code review, which he describes as a work in progress.

Díaz is grateful for the opportunity to try different gen AI platforms and to explore different use cases to make his work better and more efficient. Of course he is familiar with the discussions around how advances in AI could disrupt jobs and careers.

Humans in the loop are a key to success

He is not one to object to progress and he believes, as many do, that AI will increasingly be shaping the future.

“The only thing to do is to evolve with that and learn to work with these tools,” Díaz says.

And he is a firm believer that humans will be in the AI loop for the foreseeable future.

“I always say to people, AI is like a two-year-old that knows everything in the world. So a two-year-old may know everything, but the two-year-old will not know how things work. It has all the information, but it will not know how to use it.”

The senior staff engineer

The way Stephen Hopper sees it, gen AI has a multiplier effect when it comes to promoting efficiency. The senior staff engineer and his team can access the tools they need to hand over mundane tasks to AI. So rather than become tied up in the mundane, they can work on more valuable projects and review AI’s work when it’s completed.

“A lot of the time, I’m able to take whatever it is that I’ve built with AI, package it up in a way that then other people can use that to apply the same leverage to improve their own productivity.”

Testing is a thing in software development. A big thing. So are deadlines. Sometimes the two are in conflict. More testing means more time, more time means deadlines are in jeopardy. But AI is a time condenser.

There are instances when a senior engineer has to decide how many tests are enough tests. Should they request more tests to avoid the chance that tweaks could be needed later? Even at the expense of potentially missing a deadline?

How speed drives higher quality

“It would sometimes take a day to write additional tests,” he says of the old days, like a few months ago. “Now, it’s minutes. And it all gets done. So, it’s allowing us to write better code.”

Or consider a recent project that required moving a significant amount of data from one database to another.

“Before it would have taken a team of engineers three months to do it,” Hopper says. “We were able to have one or two people work on it, just managing a team of AI agents, and do it in three weeks.”

It is also changing the way of working for the team and the way of thinking.
Hopper explains the old way of working: “You would generally have a team of engineers and one of them would be the lead engineer on the team.”

The lead would set the technical direction, communicate back-and-forth with stakeholders throughout the company, all while the other engineers would write code as quickly and efficiently as possible and maybe review peers’ code.

The new way of working? “Now every engineer is kind of seen as a lead engineer, and they just kind of get to manage these AI agents that handle coding for them.

“So, on a given day, instead of me just focusing on one task and writing code for a single task, I’m managing multiple AI agents. And the language that I would use to describe a task that needs to be completed to a human engineer is the exact same language that I can now use when giving that same task to an AI agent.”

Offloading some tasks can lead to a clearer focus

And about the new thinking, Hopper explains that different tasks require a different mindset. Writing code means thinking a certain way, a different way from working with a colleague on project requirements, for instance. It’s the difference between thinking like a computer and thinking like a human.

“And so switching between those two mindsets can be difficult,” he says. “It’s a mental shift. But now that I don’t have to write code entirely by myself, I don’t have to switch to that computer-brain mindset as much.”

Hopper says he doesn’t write code by hand anymore. He calls that change the most amazing in the last year of the current AI revolution. Which is not to say, AI is on its own. Far from it.

“Every line of code that AI is going to write for us, a human is going to review,” he says. “We are still very carefully reviewing the actions that AI is taking. But we’re seeing huge productivity gains.”

And while it all seems so logical, this leaning into an AI-first posture, Hopper doesn’t take it for granted. He’s seen in his seven years at Signifyd that the company strives to provide the support and tools employees need to do their jobs as effectively as possible.

He’s worked at other companies where a suggestion that teams try different approaches or invest in different tools to get the job done have been met with indifference or resistance.

“They’d fire back and say, ‘Well, this is the way we’ve always done it,'” he says of his past experience. “And that would kind of be the end of the conversation.”

The conversation continues at Signifyd

As quickly as the ongoing gen AI revolution is bringing change to the way Signifyers work, we know the transformation is still in its early days. The paved path ahead is lined with opportunities to become more efficient, more productive and more fulfilled in our day-to-day work.

The challenge now is a good one: Figuring out which among the many potential initiatives to choose in order to get there.

---

Interested in an AI-first approach? Come join us.