Your AI agents are stalling. The plumbing is why.

90% of companies with AI workflows fully in production share one thing. Not Claude. Not GPT-4o. Not a particular vendor.

An integration platform.

That's the finding from MIT Technology Review Insights research, surfaced again this week as Celigo reported surging demand on June 9, 2026. The study surveyed 500 senior IT leaders and found that the thing separating companies running real AI workflows from those still stuck in pilot purgatory isn't the model they chose. It's whether they built the connective tissue first.

I've been turning this over all week because it matches what I see when I look at my own stack — n8n, HubSpot, Make.com, Clay — and the places where AI actually delivers versus where it just burns credits and produces garbage.

The number that should stop you

The inverse stat is the one everyone quotes: 90% of production AI shops use an integration platform. But the 1% number is the one worth staring at.

One percent. That means if you're trying to scale an AI workflow — a lead enrichment pipeline, an automated nurture sequence, a CRM agent that classifies inbound intent — without first making sure your tools actually talk to each other cleanly, you're essentially choosing to be in the 99% failure bracket. That's not a technology problem. That's an architecture decision you made before you opened the AI tab.

What's actually killing AI projects

Gartner made a quieter prediction that keeps getting louder: over 40% of agentic AI projects will be cancelled by end of 2027, due to escalating costs, unclear business value, and inadequate risk controls. The Gartner analyst framing is blunt: most are "early-stage experiments driven by hype and often misapplied."

The easy read is to blame the technology. The accurate read is to blame the order of operations.

Here's what I see in practice. A marketing team gets excited about an AI agent that can score leads, write personalised follow-ups, and update lifecycle stages in the CRM. They spend three weeks prompting. The agent hallucinates contact fields that don't exist, fires workflows against stale data, and writes emails that reference the wrong product tier because the HubSpot property it's reading hasn't been maintained since 2023.

The model isn't the problem. The data plumbing is.

The MIT research puts a number on why: 93% of organisations running AI with an enterprise-wide integration platform draw from three or more data sources. Zero percent of organisations without that platform achieve the same data complexity. Your agent is only as smart as the context it can actually see — and if your CRM, your enrichment layer, and your intent data aren't connected and clean, the agent is flying blind regardless of what model powers it.

The contrarian beat everyone skips

Most AI in B2B marketing gets evaluated at the output layer. "Did the email sound good?" "Did the lead score move?" What almost never gets evaluated is the input layer. Where did the data come from? How stale is it? Does the enrichment field in Clay actually write back cleanly to HubSpot, or does it partially overwrite a field that three other workflows depend on?

The teams I see pulling ahead aren't the ones running the newest models. They're the ones who spent six boring months standardising their HubSpot property schema, building reliable n8n error-handling, and making sure Make.com webhooks don't silently fail at 2am. That work doesn't go in a LinkedIn post. It's not a flashy demo. But it's why their AI workflows actually stay in production instead of getting quietly abandoned after a month.

Gartner estimates only about 130 of the thousands of agentic AI vendors are real — the rest are "agent washing", rebranding existing chatbots and RPA tools. That should give every marketer pause before they sign a contract with the latest AI-native GTM platform promising autonomous pipeline generation.

What this means for your stack right now

If you're running marketing ops on a SaaS stack and your AI agents aren't delivering, the checklist I'd start with isn't model selection. It's:

• Are your HubSpot properties consistently populated, or does enrichment logic vary by source?
• Do your n8n or Make.com workflows have error-handling that surfaces failures, or do they just silently drop records?
• Is there a single source of truth for contact lifecycle stage, or are three different tools writing to the same field?

None of that is glamorous. All of it is the reason the 90% stat exists.

The uncomfortable truth from the MIT research is that AI strategy is architecture strategy. The model is almost commoditised at this point. The integration layer — the thing connecting your CRM to your enrichment data to your activation workflows — is the actual competitive advantage. And most marketing teams are still treating it as an IT concern, not a marketing one.

I run marketing ops at a Copenhagen SaaS and the thing I spend more time on than any model evaluation is keeping the data plumbing honest. That's the work that makes the agents useful.

What does your AI stack actually look like under the hood — clean integration layer, or held together with hope and duct tape?