# Seven takeaways from SaaSpocalypse Now

> We hosted enterprise IT and business technology leaders in San Francisco for an honest conversation about what the SaaS apocalypse actually looks like from the inside. Here are seven things that came out of the room.

**Author:** Paul Turner  
**Read time:** 12 min read  
**Published:** Apr 22, 2026

---

The SaaSpocalypse conversation happening [in the press](https://www.forbes.com/sites/donmuir/2026/02/04/300-billion-evaporated-the-saaspocalypse-has-begun/) is mostly about stocks and survival. The conversation happening inside IT teams is about something harder: how do you actually make AI work when your systems are messy, your stakeholders are impatient, and the framework you're supposed to follow was designed for a world that no longer exists.

Last week, Tray hosted SaaSpocalypse Now — an invite-only practitioner panel in San Francisco with heads of business technology from [Notion](https://www.notion.com/), [Zuora](https://www.zuora.com/), and [Yext](https://www.yext.com/index.html). No slides or vendor pitches. Just practitioners comparing survival notes on what blew up, what held, and what they're building on the other side of it.

Here are seven things that came out of that room.

## 1. Speed and velocity are not the same thing

_"We're all moving fast. But there's a difference between speed and velocity. A lot of teams are wondering if they’re heading in the right direction, or if they’re just doing a lot of show-and-tell and ending up back where they started." — Jalal Iftikhar, Head of Business Technology, Notion_

Every organization in the room had a version of this problem. Prototypes shipping daily. Pilots running in parallel. Slack channels full of demos. The companies making meaningful headway have stopped treating speed as the metric and started asking whether AI initiatives are actually changing how the business operates, and not just adding activity to it.

## 2. "Agentic hellscape" is a real risk

_"Instead of cleaning up SaaS applications, organizations are creating an agentic hellscape where they don't know where things are at." — Jalal Iftikhar, Head of Business Technology, Notion_

You beat SaaS sprawl. It took a decade, cost a fortune, and required an entire apparatus of iPaaS, SSO, and procurement governance to get there. Congratulations. Now do it again — faster, with higher stakes, because agents don't just hoard licenses. They take actions.

Business teams empowered with AI tools and a mandate to move fast are building, but often without a clear owner, a production path, or visibility into what those agents are connected to or what data they're touching. Shadow IT took years to create governance problems. The agentic version is doing it in months.

## 3. Agents should earn write access

_"Nothing that's AI-enabled can write. Nothing can write to our data. Not yet." — Mark Gill, Head of Infrastructure and Security, Zuora_

Every vendor will tell you their agent has guardrails. Ask them to prove it. The practitioners who have thought hardest about this drew a hard line: nothing AI-enabled writes to production data. Not yet. That's a deliberate governance decision made while they build out the observability, identity, and risk-tiering infrastructure that write access requires.

The ROI case is clear even without write access: significant value is available from agents that gather data, assemble context, and surface recommendations without ever touching a production record. Moving carefully means you don't spend six months untangling what an over-privileged agent did to your transactional systems at 3am.

The write-path question is the one worth pressure-testing with every vendor: not whether they have an approval step, but how write actions are scoped, attributed to a specific agent identity, audited, and reversed when something goes wrong.

## 4. The best AI wins came from connecting data that had never talked to each other

_"It was the first time we actually went — oh, we can use this insight engine to correlate data across systems to help provide a view that really wasn't available before." — Mark Gill, Head of Infrastructure and Security, Zuora_

The biggest wins described in the room weren't complex reasoning or sophisticated automation. They were agents that pulled data from systems that had always lived in separate silos and assembled it into something a human could act on in minutes instead of days. Three patterns surfaced independently across all three companies.

**The first: **a sales-prep agent that pulls from a dozen unrelated systems via Slack and delivers a customer brief that used to take a team of four an entire week. The agent didn't automate that work. It made the work obsolete.

**The second:** natural-language access to data warehouses with role-aware access control built in — meaning users get the data they need without IT having to give everyone open access to Snowflake or Tableau. The agent enforces the access policy.

**The third:** design-spec generation, combining an MCP server with recorded customer calls to auto-generate implementation specs in about ten minutes. The bottleneck was never writing the spec. It was the lag between the customer conversation and someone sitting down to produce it.

All three patterns share two things: they're read-heavy or have a narrowly bounded write surface, and they sit on top of an integration layer that reaches across many systems cleanly. That combination is not accidental.

## 5. The build vs. buy question has a new prior question

_"Should we buy, should we build — the first question we're actually asking is, should a human actually even do this? What is the craft that you bring to this?" — Jalal Iftikhar, Head of Business Technology, Notion_

Most AI strategies are built on the wrong question. "How do we do this faster?" is yesterday's ask. The teams ahead of the curve are asking something more uncomfortable: "Should anyone be doing this at all?" Before asking whether to buy a tool or build one, some teams are now asking whether the work needs to happen at all. Anyone selling "we'll automate your existing process" is selling a diminishing return.

For the work that does need to happen, the pattern that generated the most consistent ROI was the same across every deployment described: narrow scope, deep integration, clear output, delivered where users already work. An IT support agent handling L1 requests entirely through Slack — processing screenshots, identifying issues, creating tickets — with no portal login and no form fields. A finance bot answering contract questions on demand instead of routing them through an analyst. The agents that get used are the ones that don't ask people to change how they work.

## 6. The path to production is still being invented

_"All big POCs will die. You need a clear owner who can take responsibility and accountability to take that POC to production — like how we do in any project." — Tulasi Donthireddy, Senior Director of IT and Business Systems, Yext_

The ownership problem is real — but it sits inside a bigger one. The old frameworks don't quite fit anymore. Agile was built for a world where requirements were knowable upfront. The SDLC assumed a relatively stable definition of what "done" looks like. AI projects don't behave that way. The goalposts move, the technology improves mid-build, and leadership priorities shift faster than delivery cycles.

The teams making progress aren't trying to retrofit old frameworks. They're building new ones from scratch by defining clear criteria for what deserves to graduate from POC, assigning owners before anything ships, and establishing what production-grade actually means for an AI system: observability, a cost center, an incident rotation, and someone whose name is on it.

One practical starting point from the panel: define the input, the AI task, and the output before you write a line of code. Know who owns it. Integrate the result where users already live.

## 7. Change management is the most underrated blocker

_"There is a specificity gap. What is that specific solution you want to identify so that you can help the middle layer of your organization move fast towards adoption?" — Tulasi Donthireddy, Senior Director of IT and Business Systems, Yext_

The change management problem in AI adoption has a specific shape that's different from previous waves of enterprise technology. Previous transformations such as cloud migration, SaaS adoption, or mobile-first were largely contained to IT and product teams. AI is being pushed into every function simultaneously, often by people who have never had to think about systems architecture, data governance, or what a production environment actually means.

Finance teams are being handed GitHub access. Legal teams are being asked to vibe-code contract workflows. Operations teams are building agents on laptops and wondering why IT keeps slowing them down.

But the pressure is on both sides. The practitioners in the room are being asked to govern something they didn't build, and the business teams building it are being asked to move fast on tooling they don't fully understand. The organizations that get this right will be the ones that treat the middle layer — the people who aren't IT but aren't pure business either — as the most important investment they can make right now.

## A diligence checklist for your next vendor conversation

The integration platform sitting underneath your agents is becoming the real control plane. Most IT teams are still evaluating integration vendors on connector count and ease of setup. Here are the questions that actually matter now.

**Ask your vendor to demonstrate an agent attempting a write to a system of record.** Not a happy-path demo. Show what gets logged, what can be rolled back, how the action is attributed to a specific agent identity, and who owns the audit trail. If they need to schedule a follow-up to answer that, the governance model isn't real.

**Ask what happens when you have 500 agents in production instead of five.** Registry, lifecycle, ownership, deprecation. If the answer lives in a slide deck, you are buying a tool that will become part of the sprawl problem rather than the solution to it.

**Ask how the orchestration layer is called by agents rather than by humans.** MCP server coverage, agent-callable workflow APIs, how the platform handles the transition from human-initiated to agent-initiated workflows. This is where most legacy iPaaS platforms fall apart.

**Ask who is accountable for each agent.** No named owner means no governance. No governance means a short production life or an irreversible accident waiting to happen.

**Ask about pricing.** Per-seat SaaS pricing is on borrowed time. If your vendor's commercial model depends on seats, your renewal conversation in eighteen months is going to be uncomfortable. Renegotiate now — or plan to.

**Ask whether you've rationalized your SaaS portfolio first.** The companies making the most durable AI investments spent the twelve months before they started consolidating overlapping tools and redirecting the savings. That's a more sustainable funding model than a one-time AI budget line — and it forces the architectural discipline that good AI deployment requires anyway. Replacing a legacy integration platform before building agents on top isn't glamorous. It's why the agents work.

_The SaaSpocalypse isn't coming for the companies that admit they're still figuring this out. It's coming for the ones that think they already have. [Reach out if you’re ready to talk survival tips.](https://tray.ai/contact/talk-to-sales)_