TL;DR
- MCP went from spec to universal standard: now under the Linux Foundation with all major players on board
- AI agents got real, but so did their limits (vibe-coded apps created technical debt)
- Connectivity infrastructure attracted billions: $8B Salesforce-Informatica (data pipelines), $2.85B ServiceNow-Moveworks (AI automation)
- Integrations are critical path for AI: data integrations, MCP Apps, and MCP UI are foundational to making AI promises work
What we did: Launched the Unlimited Integrations Cohort, shipped Frigg 2.0, and learned that integration expertise + AI fluency is the combination that wins.
The Vibe-Coded Hangover
Let's start with the most important lesson from 2025, because it's the one nobody else is talking about yet.
Y Combinator's Winter 2025 batch revealed that 25% of startups have 95% AI-generated codebases. In March 2025, a vibe-coded payment gateway approved $2M in fraudulent transactions due to inadequate input validation. The AI had copied insecure patterns from its training data.
Research shows that 40% of AI-generated code snippets contain vulnerabilities. Google's 2024 DORA report found that a 25% increase in AI usage leads to a 7.2% decrease in delivery stability.
The pattern we saw in integrations specifically:
- People built integrations by prompting Cursor and Claude without deeply understanding the APIs
- Some worked great. Some created time bombs.
- The gap between "AI can do this in a YouTube video" and "AI can do this reliably for our business" is still significant
The counter-hype reality: Running a business is 100x different than generating code. Distribution, marketing, and sales still require human judgment. AI amplifies. Make sure you're amplifying something good.
And expect 2026 to be about cleaning up vibe-coded messes by seasoned AI-assisted developers. That's a real opportunity, and one we're positioned for.
What We Saw
The MCP Explosion
Let's cover the obvious: Model Context Protocol went from "interesting Anthropic side project" to "the standard everyone's talking about."
I wrote about MCP in our very first newsletter this year, saying it would be like having an API: table stakes within a few years. I didn't expect the adoption curve to be quite this steep.
By mid-year, every serious AI-first product was either building an MCP server or asking how to build one. Then in December, it went further: Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation, with OpenAI, Google, Microsoft, and AWS as platinum members. 97 million monthly SDK downloads. The "USB-C for AI."
And on December 17th, OpenAI opened the ChatGPT App Directory to submissions. Real distribution for MCP-based integrations to 800 million weekly users, finally.
What surprised me: how quickly the security conversations started. Early MCP adopters got excited about the capabilities and then immediately hit questions about authentication, rate limiting, and "wait, should AI really have access to this?" Good problems to have, but real problems.
The AI Agent Wave (And Reality Check)
2025 was the year "agentic AI" went mainstream. Not just chatbots that answer questions, but AI that takes actions. Books meetings. Files tickets. Moves data between systems.
But I also had a lot of conversations with people who got burned by vibe-coded agents that worked in demos and broke in production.
The counter-hype reality from research:
- 95% of enterprise AI pilots fail to deliver ROI (MIT)
- Only 16% of enterprise deployments qualify as true agents
- LLM limitations are fundamental, not fixable. Hallucination is a structural consequence, not a bug
My take: agents are real, they're useful, and they're not magic. Treat them like any other tool: test thoroughly, fail gracefully, keep humans in the loop for anything that matters.
Connectivity Infrastructure Attracted Billions
The Salesforce-Informatica deal ($8B) was the headline, though it's worth noting that Informatica is data infrastructure and pipelines, not integration in the traditional SaaS-connector sense. Still, it underscores the same theme: connectivity is critical, and the market is pricing it in.
| Deal | Value | Category |
|---|---|---|
| Salesforce + Informatica | $8B | Data infrastructure & pipelines |
| ServiceNow + Moveworks | $2.85B | AI automation + enterprise service |
| Workday + Pipedream | Undisclosed | 3,000+ connectors to enterprise HR |
| Zapier + Utopian Labs | Undisclosed | AI agent capabilities |
| n8n funding | $180M | Workflow automation (valued at $2.5B) |
| Clay funding | $100M | GTM data enrichment (valued at $3.1B) |
The market is consolidating and the categories are blurring. Connectivity infrastructure is attracting serious investment.
What I keep telling clients: the "build vs. buy" framework for integrations now has a third option: "have AI handle it." That changes the calculus for when you need a dedicated integration platform vs. when you can just let agents coordinate.
HubSpot Ecosystem Maturity
I have to mention HubSpot because we work in their ecosystem a lot and they had a good year.
Their date-based API versioning rollout was handled well. The Provider Program continues to mature. Scott Brinker leaving was news (though he graciously gave me some time to chat before the announcement).
More broadly, HubSpot's ecosystem approach (marketplace, partner tiers, clear documentation, developer experience investment) is becoming the model other platforms copy.
What We Heard
"We've been meaning to build those integrations..."
This one never gets old. Every year, every quarter, someone tells me they've had integrations on the roadmap forever but keep pushing them for other priorities.
I get it. Integrations are complicated, they solve problems for only a subset of users, and you only control half the equation. When there are fires to fight, integrations get deprioritized.
But the backlog is building. And AI is making the stakes clearer: when someone asks "why can't Claude pull this from Salesforce?", it's an integration problem.
"Are we doing this right?"
More strategic conversations this year than ever before. Not "build us this connector" but "how should we think about our integration strategy?"
Common themes:
- Native vs. embedded platform vs. point solutions. There's no one right answer, but there are clearer frameworks now
- AI-first vs. AI-augmented. How central should AI be to the integration architecture?
- Build team vs. outsource. Integration expertise is rare; should you hire it or rent it?
The 10x Rule
Multiple conversations touched on this: an integration needs to be 10x better than the alternative (usually copy-paste or manual process) to justify existing.
With AI getting better at handling manual processes, that bar is rising. The integrations that survive are the ones that deliver clear, measurable, significant value.
"What should we do about MCP?"
This question went from zero to constant over the course of the year. By Q4, basically every product-focused conversation included some version of "do we need an MCP server?"
My standard answer: if AI assistants are a plausible channel for your users, yes. If you're B2B SaaS, assume they are. And with the ChatGPT App Directory now open, the distribution question just got answered.
What We Did
Launched the Unlimited Integrations Cohort
Our biggest experiment of the year. Instead of fixed-fee per-integration pricing, we offered a cohort model: dedicated team, steady pace, iterate over time.
The hypothesis was that integration work benefits from relationship and context-building more than discrete project handoffs. So far, the hypothesis is holding.
We learned a lot about:
- Onboarding. We need a few days to really understand a product before we can build well
- Value metrics. Defining success upfront makes everything easier
- GTM materials. A great integration nobody knows about is a tree falling in an empty forest
Frigg 2.0 Development
Our open-source integration framework got a major overhaul. Serverless-first, CloudFormation-based, infrastructure-as-code from the start.
The goal: make it easier to build integrations that are production-ready from day one, not prototypes that need to be rebuilt later.
Internal Automation Experiments
We practice what we preach. This year we got serious about automating our own operations.
Some wins:
- Automated proposal generation that pulls context from CRM and previous conversations
- Meeting prep docs that synthesize relevant background before calls
- Fathom transcripts feeding into knowledge management
Some failures:
- AI-generated follow-up emails that were technically accurate but tonally off
- Automated scheduling that created more confusion than it saved
- Various "clever" automations that were more clever than useful
The lesson, as always: automation amplifies. Make sure you're amplifying something good.
What's Next
MCP Everywhere
If 2025 was "MCP exists and is interesting," 2026 will be "MCP is expected." The platforms that have MCP servers will have an advantage with AI-native users. The platforms that don't will feel the gap.
We're ready: We've been building MCP servers since the early days. If you need to get an MCP app into the ChatGPT directory, we can help.
The Vibe-Coded Hangover Continues
All those integrations and apps built by prompting Cursor without understanding the underlying systems? They're going to keep breaking. "Integration debt" becomes a real category.
We're ready: We specialize in cleaning up and professionalizing integrations. AI-assisted, human-supervised.
Integration as Strategy
The conversations we started having in 2025 ("how should we think about integrations strategically?") will become more common. Integration stops being a technical topic and becomes a product and business topic.
We're ready: Our strategic assessments help executives understand their integration landscape and make informed decisions.
The "Integration Surgeon" Role
This is maybe wishful thinking, but I think there's a real specialty emerging: people who aren't full-stack developers, aren't no-code builders, but are specifically expert at connecting systems. We've been doing this for 10 years; maybe 2026 is when it becomes a recognized thing.
The Takeaway
Integrations are the critical infrastructure that makes AI work. For all of AI's promise (agentic workflows, autonomous assistants, intelligent automation), none of it delivers without connectivity.
This shows up in three layers:
- Data integrations remain foundational, syncing information between systems
- MCP Apps let AI assistants use external tools through a standard protocol
- MCP UI (emerging) brings interactive experiences directly into chat interfaces
AI is the catalyst driving investment. MCP is the standard enabling it. Market consolidation is the context: big players making big bets means more resources flowing into connectivity infrastructure.
The vibe-coded hangover is the opportunity. Complex integrations that AI can't handle dynamically are becoming MORE valuable.
If any of this resonates, or if you're staring at an integration challenge you're not sure how to solve, let's talk.
Living the vibe-coded hangover right now? Drop into the chat and tell us your war story. And if this made you feel seen, share it with a friend who's been through it too.
Read next: 25 Integration Insights from 2025, 26 Predictions for 2026, and The 3 Layers Shaping Integration