Tap CXM’s Agentic AI Hackathon 2.0: Bigger & Brainier

Pushing the Limits of Agentic AI in CXM

We launched an open-brief hackathon in mid-2025 to see what our people could cook up using Agentic AI. At the time, the tech was generating a lot of hype. But it was largely untested against real-world challenges.

Our team’s AI credentials range from undergrad electives in the ’90s to CompSci degrees in the ’00s and a Masters as recently as 2024, so a hackathon was inevitable. 

Four teams of Tap CXM tech heads spent four weeks coming up with prototypes and we saw that built-for-purpose AI agents had real potential. And so, a seed was sown.

Agentic AI Hackathon v2.0

Naturally, we needed to run a second sprint. 2025 was a big year for AI developments and we wanted to see what was possible. The second cohort had access to the same resources as the first:

• Paid OpenAI accounts.
• Microsoft Copilot.
• N8N.
• Permission to carve out time (as long as it didn’t interfere with client work!)

Three teams, comprised of people from our London and New York offices, worked on prototype tools to solve challenges of their choosing. 

Senior Tap CXM team members Dirk and Harry guided the groups’ progress. Then, in the final days of 2025, everyone gathered on a Teams call for the big reveal. As expected, what they’d achieved outdid everyone’s expectations.

One team built on an idea from the first hackathon round. One addressed a persistent headache in email marketing platform migrations. And one used enough complex maths to cause a headache for anyone trying to write a simple-terms blog about it. 

Adobe Campaign Intelligence Tool

Creators: Akhil, Chris, Daisy, Dimple and Marcin.

The Idea

Adobe Campaign workflows are often opaque to non-specialists. In our first hackathon, a team tackled this challenge with an “XML Explainer” that translates complex code into simple English.

This time, one of the teams built an agentic tool that translates the other way. They were prompted by a question from our fearless leader and co-founder, Andrew: “Wouldn’t this be great if we could flip it on its head and get it to actually build the workflows rather than answering the question of what’s in the workflow?”

What It Does

Non-technical users can prompt an AI agent to create XML workflows from scratch or make quick edits.

The tool takes a plain-English request and produces Adobe Campaign-compatible workflow code. Because it’s provided with examples of working code, the agent can retry if its first attempt fails. Users keep chatting to iterate until they get a usable result.

Why It’s Interesting

This isn’t the end of the Adobe Campaign specialist. But it could enable non-specialists to build workflows and ship campaigns faster, freeing up specialists to focus on complex stuff.
The prototype includes explicit thinking around validation, access control, and data privacy – all identified as risks when deploying autonomous AI agents. It’s also designed for scalability, although it’s only running on sandbox data so far. Swappable model APIs will let clients decide which AI model to use.

What Needs More Work

As of 2026, the prototype is functioning with a limited number of use cases. Further development will enable it to handle broader and more complex requests.

We also had a ‘Cybertruck moment’ in the live demo when Adobe Campaign returned an unexpected error. Hey, it happens. That’s why we experiment.

With ongoing iteration – and some clients who are keen to test the solution – Tap CXM’s Adobe Campaign Intelligence Tool could be a game-changer for enterprise marketing campaigns.

Email Migration Agent

Creators: Chantelle, Deb, Hamish, Jack and Prins.

The Idea

Email platform migrations are slow and manual. They’re prone to human error. Especially when the platforms speak different languages.

One team set out to build a tool that addresses all these issues. Rather than rebuilding each template and asset from scratch, martech specialists ask the AI agent to convert it.

No manual coding. No human error. Just lightning-fast conversion with platform-specific syntax.

What It Does

The tool converts email templates in seconds using a completely self-contained architecture. Users upload campaign templates, select the ‘from’ and ‘to’ platforms, input details like user IDs or licence keys, et voilà. A production-ready template appears.

N8N automations drive the middle of the process. A series of web hooks, API calls, middleware, and React components do the technical bits.

Instant preview capabilities allow users to check how things look on desktop and mobile, and verify the code and configuration before deploying.

Why It’s Interesting

The team made some pretty lofty claims about their AI email migration tool:

• Cut email platform migration timelines by 90%.
• Slash labour costs by up to 80%.
• Reduce risk and human error.
• Speed up time-to-market from weeks to hours.
• Make migration workflows scalable and repeatable.

(It’s important to point out here that the Hackathon is a friendly competition. Some friends are more competitive than others…)

So, does it deliver on those promises? Early signs point to yes.

In a live demo, the team converted email templates back and forth between Marketo and Salesforce Marketing Cloud. The structure and formatting carried over correctly. Even the variables, dynamic content, and tokens were tickety-boo.

The team also developed enterprise features, including collaboration tools, template management, and conversion history tracking.

What Needs More Work

The current prototype converts bi-directionally between four major email automation platforms:

• Salesforce Marketing Cloud
• Marketo Engage
• Adobe Campaign
• Braze

An obvious next step is expanding that capability to more of the big email automation platforms. In theory (and oversimplifying), that’s a matter of mapping syntaxes, structures, tokens, and formatting.

Later expansions could include:

• API integration: One-click creation, direct deployment, automated validation, rollback capabilities, and fewer manual steps in setup.
• Batch migration: Enabling bulk uploads, scheduling, building a progress dashboard, and automating quality checks.
• Asset migration: Image and media library transfers, content block libraries, and underlying dynamic content rules.

If you’re interested in trying the tool for yourself – well, it’s not quite client-ready yet. But that day isn’t too far off.

ML-Driven Prediction Engine

Creators: Chauncey, Dhanesh, Jeffrey, Uday

The Idea

Most businesses struggle with fragmented data. It’s a constant source of frustration, especially after investing in a CDP (Customer Data Platform) and finding it doesn’t ‘automagically’ create a single customer view.

This team set out to prove that machine learning (ML) can improve segmentation and thereby boost CDP performance. They wanted to develop a process that uses a ML model on existing data sets to drive and determine segmentation.

What It Does

The prototype clusters user behaviour based on signals coming from engagement platforms like web and email. In the middle, the ML model uses an ‘unsupervised’ learning technique to determine the central point between clustered data without needing pre-defined labels.
This means that behaviour determines segmentation. People aren’t wedged into pre-defined segments.

There’s some complicated maths going on behind the scenes. K-Means, centroids, regression models, and other words that make us sound really smart.

The upshot is that the model outputs very tight segments and individual profile data based on real behaviour signals. A CDP then takes these outputs and uses them to instruct activation platforms, leading to improved personalisation at scale.

Why It’s Interesting

Aside from the impressive maths behind the model, there are several things to be excited about:

• CDPs get better segmentation and customer profile data based on evidence rather than assumptions.
• The CDP doesn’t need to re-ingest data output by the model, reducing licensing and resource requirements.
• It’s configurable for any CDP, including the major platforms or a composable CDP layer.
• The architecture is scalable and low-cost, so it’s not just for enterprises.
• Clusters evolve as more data comes in, meaning the model isn’t static or based on assumptions.

Above all, the potential for better personalisation has us excited to see what’s next for this prototype. It shows what a machine-learning-driven marketing platform might look like.

What Needs More Work

In its current form, the model creates segments like:

1. Abandoned cart
2. Page views
3. On-page engagement
4. Email recipient
5. Purchaser

This is already valuable for teams to understand behaviour and optimise customer experiences.

Adding a logical regression model would level things up. It would enable predictive scores at the individual level, feeding into a highly optimised decisioning engine for next-best-action marketing.

Other potential upgrades include a persistent ML training model (for cluster re-use and continuous improvement) and AWS implementation (for enterprise scalability).

Oh, and you might be asking, “But where’s the AI agent?”

Machine learning is what enables AI systems to ‘learn’. So, we can add Agentic AI to the list of potential improvements, aiming to automate the in-between steps using an agent.

What’s Next for Agentic AI in CXM?

Over eight months and two Hackathons, we’ve seen AI evolve at a blistering pace. It’s undoubtedly going to impact the way we work, shop, search, interact, and learn.

The big question for us is whether Agentic AI will disrupt marketing and customer experience management as we know it. Or will the tech fall in step with how we work, speeding up time-consuming processes and automating manual tasks?

What we do know is that Agentic AI is an emerging solution for emerging problems. It delivers the most ROI when applied to specific use cases that can’t be solved through better data management, removing bureaucracy and bottlenecks, or investing in customer intelligence.

In short, a lot remains to be seen. But you can rest assured we’ll be watching keenly while tinkering with our own AI-enabled tools, getting them ready for clients to use.