How ABM teams are building an AI advantage that survives platform transition

Moving house is no mean feat. There are things you own, and then there are the things you’ll let go. The kitchen units, carpets, and fitted wardrobes will likely stay behind for the next owner. (Granted, they may rip it all out, depending on taste.) What you end up packing into the van is the stuff you own outright and is relatively portable: your furniture, crockery, books, framed photos, DVDs – if you still have them – because these are the things that make a home yours.

Most ABM teams have been busy building their AI advantage into the house, and not the van. The workflows and hard-won knowledge of how a campaign should be built reside in whichever platform happens to host it this year. If the platform changes, so does that advantage.

But there’s a small group of marketers doing the opposite – focusing on the van and building for portability, and the industry is beginning to take note.

Most ABM teams rely on the platform

This month’s live webinar with ForgeX featured the ABM platform Propensity and the digital experience platform Folloze. It provided a very useful illustration of where this is heading. Claude sits above both, called through MCP connections, orchestrating a campaign build in Propensity and a personalised microsite build in Folloze from a single chat. If you choose, neither platform is the thing being operated directly any more. Both are being operated through.

Most marketing organisations are still adding AI on top of a stack that wasn’t meant to be orchestrated this way. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls as the leading causes, not the underlying technology.¹

The stack is splitting into two layers, and a third is still up for grabs

Martech researchers Scott Brinker and Frans Riemersma, in their State of Martech 2026 report, describe a landscape that looks almost static. 15,505 products were tracked, up just 0.79% on last year: 1,488 new products added against 1,367 removed.² This surface calm hides a structural rupture that has appeared underneath.

Their take on it is that the stack isn't consolidating – it's stratifying. AI-native tools are winning the content-creation layer. Incumbent SaaS is holding the orchestration layer – the lead scoring, the dashboards, the tasks where the data already lives inside the platform. And then there's a third agentic AI layer that takes autonomous action, which they describe as still being negotiated, and nobody's settled it yet.² The ForgeX webinar represents that third layer. It showed Claude sitting above Propensity and Folloze as the AI agent taking autonomous action across both platforms from a single chat.

Given this is the direction the industry is heading, the question moves from ‘which platform should we buy’ to ‘what do we own that survives us changing platforms’. For most teams, today, the answer is very little.

This isn't isolated – we’re seeing it industry-wide

It would be easy to interpret the Propensity and Folloze demo as a single well-produced example rather than a genuine market shift. Except the same pattern is turning up everywhere you look. Anthropic's own Claude Connectors Directory already lists HubSpot, Webflow, Wix, WordPress, Zapier, Notion, Asana and Airtable among dozens of others, and independent registries now index more than 29,000 MCP servers in total.² It took the commercial martech space 15 years to reach 15,000 products. MCP-based integrations passed nearly twice that figure in about 18 months.² This trend shows the default way agents expect to reach your martech stack.

We saw the same pattern turn up last week on a call with Dreamdata, a B2B attribution platform. Their team confirmed they're launching their own MCP server so their attribution data can be queried directly inside Claude or ChatGPT, used to build ICP-matched audiences, or wired into automated alerts when a target account shows engagement. These three vendors – Propensity, Folloze and Dreamdata – have taken the same path: become the execution layer, and let an orchestrator sit on top. As we put it on the call, a headless attribution platform starts to look a lot like a headless operating model. So this is the direction the whole stack is moving in, and the 'third layer' is exactly where the fight for advantage is about to happen.

What actually moves with you is the thing nobody's pricing into the stack

This is where the webinar's second half got really interesting. Trey Harnden, GTM engineer at Folloze, is building his workflow with Claude skills, which he defines simply as an SOP for an agent – an ordered set of steps, specific tools and specific data sources – saved as a plain text file and version-controlled in GitHub rather than buried in a chat history. He's built roughly 220 of them, nested inside each other, so that a single prompt to build a one-to-one microsite for a named account automatically pulls in a brand-harvesting skill, an ABM strategist skill and a designer skill without him having to specify any of it by hand.

GitHub is traditionally used to store code – it tracks every change made to a project over time. This repository keeps working even if Folloze is swapped for another platform. It may be a bit more time-consuming to store and update skills this way, but there are real advantages:

• Skills survive vendor churn. A documented, version-controlled SOP doesn't care which execution platform it's pointed at. Swap the tool, keep the skill. Granted, the skill may need a few updates, but the foundational logic would remain.
• Skills compound, licences don't. A platform subscription resets in value the moment you stop paying for it. A library of 220 tested workflows keeps producing value indefinitely, and gets more valuable as it's added to.
• Skills are the thing a competitor genuinely can't buy. Anyone can purchase the same ABM platform, but nobody can purchase a year of accumulated, battle-tested SOPs for exactly how that team runs a campaign.
• Who did what, and when. Every skill change is tracked, so you can revert, compare versions, and see exactly who changed what and why. This provides an audit trail handy for compliance, particularly in regulated industries.
• Onboarding and continuity. A documented, versioned skill survives someone leaving the team. A new hire can read the actual history of how a workflow evolved, rather than inheriting tribal knowledge that walks out the door with whoever built it.

This changes what "senior" means on an ABM team

If the platform execution layer is increasingly commoditised and the durable value sits in how well a workflow is specified, the scarce skill on an ABM team stops being campaign production and starts looking a lot more like management. Trey's advice on building skills is to focus on the narrowest possible task, the kind you'd hand an intern – get that right, then nest it inside a broader skill once it's proven. That’s essentially describing the steps precisely enough that someone, or a robot, can execute it without you.

As execution tools keep getting swapped, absorbed and replaced underneath the orchestration layer, whatever isn't written down as a portable skill effectively belongs to whichever platform currently hosts it, and it moves out the moment that platform does.

Where we’d start...

We treat our own standard operating procedure skills much in the same way. Documented, version-controlled, and backed up outside any single platform, rather than sitting in one person's AI instance or laptop. The starting point is smaller than you may think, and we help clients build this discipline into how they work, so institutional knowledge compounds instead of resetting with every new tool or hire.

Going back to our removal van. The teams with the real advantage in twelve months won't be the ones who signed with the platform with the best-fitted kitchen. They'll be the ones who could move house tomorrow and take everything that actually matters with them.

Sources: ¹Gartner predicts over 40% of Agentic AI projects will be cancelled by end of 2027, 2025. ²Chiefmartec, State of MarTech 2026.