We published a piece the other day — “The Wave of AI Agent Churn To Come: Prompts Are Portable” — and the responses were sharp and fierce. Especially:
“OK, I get it. But is MY AI category actually at risk?”
Every AI agent founder hopes they’re the exception. That their AI B2B product has the deep moat. That their switching costs are real.
Some of them are right. Most of them probably aren’t.
And what I can also tell you is so many AI B2B purchases today, the customers are really just buying for a year for now. To see how it goes. And then … look again in 6-9 months. And if they can, just migrate the prompts.
After running 20+ AI agents at SaaStr and watching this play out across our portfolio, I think there are really four distinct levels of prompt portability. And the level you’re at essentially determines whether you’ll hold 95% gross retention or watch it decay to 80% (or lower) over the next year or so.
Level 1: Copy-Paste Portable (The Danger Zone)
Examples: Basic AI SDR agents, outbound sequencing agents, AI email writers, generic chatbots, meeting summarizers
Switching time: Hours. Maybe a day.
What actually transfers: The entire core prompt — tone of voice, ICP definition, objection handling, qualification criteria, persona targeting. All of it. You literally copy it from Vendor A’s dashboard and paste it into Vendor B’s dashboard.
We did exactly this at SaaStr. Took our best-performing AI sales agent prompt — the one we’d spent weeks refining — and dropped it into a competitor. Worked on the first try. Not perfectly. But 80% of the way there in about 20 minutes.
The “moat” these vendors claim is usually some combination of better UI, better analytics, and “our model is fine-tuned.” The UI is nice but not sticky. The analytics are useful but replaceable. And the fine-tuning advantage lasts about six months before the underlying models converge.
And the truth is, for many simpler applications … you only need just so much memory. Just so many emails iterated and sent. Then you catch up to where the last app was.
One real moat in this category? Security, email deliverability and domain reputation. If your AI SDR vendor has spent months warming up sending domains, building IP reputation, and managing deliverability across millions of emails — that’s actually harder to replicate. It’s just not an AI moat. It’s an infrastructure moat. And the founders who understand that distinction are the ones who will survive.
Retention forecast: 80-85% gross retention. Expect annual bake-offs to become the norm. Buyers will test 2-3 vendors every renewal cycle because the switching cost is negligible.
Level 2: Prompt-Plus-Data Portable (The Uncomfortable Middle)
Examples: AI customer support agents, AI sales coaching, AI content generators trained on brand voice, AI recruitment screeners
Switching time: 2-4 weeks.
What transfers: The core prompt and instructions transfer easily. What doesn’t transfer cleanly is the accumulated training data — thousands of resolved tickets, months of call recordings, hundreds of brand-approved content examples. This data has to be re-ingested, re-indexed, and re-validated.
This is the category where vendors overestimate their moat and buyers underestimate the switching cost.
The vendor thinks: “We have 18 months of their ticket data training our model. They’ll never leave.” The buyer thinks: “It’s just a prompt, I’ll switch in a weekend.” The reality is somewhere in between — it’s a real project, but it’s a 2-4 week project, not a 6-month CRM migration.
The critical variable here is data portability. If the vendor makes it easy to export your training data, conversation history, and performance analytics — they’re actually accelerating their own churn risk. If they make it hard, they buy time but breed resentment. Neither option is great.
The winning play for vendors in this category: make the data layer so valuable and so continuously improving that the switching cost isn’t about the data transfer — it’s about the data loss. If your model genuinely gets measurably better every month from accumulated customer interactions, and that improvement resets to zero with a new vendor, you have a real retention story. But you have to prove it with metrics, not marketing.
Retention forecast: 85-90% gross retention. Better than Level 1, but still vulnerable to annual competitive reviews. The vendors who publish clear “your model improved X% this quarter” reports will hold retention. The ones who can’t quantify their data advantage won’t.
Level 3: Workflow-Embedded (The Integration Moat)
Examples: AI coding agents (Cursor, Windsurf, Claude Code, etc), AI agents deeply embedded in CRM/ERP workflows, AI agents that manage multi-step business processes
Switching time: 1-3 months.
What transfers: The prompt is almost irrelevant here. The value isn’t in what you told the AI to do — it’s in how the AI is woven into your existing tools, workflows, and daily habits.
Cursor is a good example. The value is in the agent managment, the IDE integration, the codebase indexing, the way it understands your specific repo structure and coding patterns. “Switching” doesn’t mean copying a prompt — it means relearning an entire development environment. That’s a meaningful cost, even if it’s measured in productivity loss rather than dollars. It’s easy to play with other tools. But harder to move an entire enterprise over.
Same with AI agents that sit inside Salesforce, HubSpot, or your ERP. The prompt might be portable, but the 47 workflow automations, the custom field mappings, the approval chains, the reporting dashboards built on top of the agent’s output — none of that transfers. You’re not switching an AI agent. You’re switching a chunk of your operating system.
This is the classic SaaS playbook applied to AI: the product itself isn’t the moat, the integration surface area is. And it still works — but with a caveat. The integration moat erodes faster than it used to because AI agents can increasingly self-configure integrations. What took a solutions engineer three weeks to set up in 2024 might take an AI-assisted setup wizard 30 minutes in 2026/2027 (and then a bunch of human QA after).
Retention forecast: 90-93% gross retention. Close to traditional SaaS levels, but don’t get complacent. The integration advantage is real today but has a shorter half-life than historical SaaS integrations.
Level 4: Domain-Locked (The True Moat)
Examples: AI medical coding and billing agents, AI legal contract analysis, AI accounting close and audit agents, AI drug discovery, AI insurance underwriting
Switching time: 6-12+ months. Sometimes effectively impossible for legacy matters.
What transfers: Almost nothing meaningful.
This is the only category where I’d tell a founder: you can probably stop worrying about prompt portability at the moment. Not because your prompts aren’t portable — they are. But because the prompt is maybe 5-10% of the value. The other 90% is:
Proprietary training data that took years to accumulate. A medical coding agent trained on millions of actual claims with known outcomes. A legal analysis agent trained on tens of thousands of contracts with real-world dispute resolutions. This data doesn’t exist in a prompt. It exists in specialized fine-tuned models, RAG databases, and validation layers that were built over years.
Compliance and certification infrastructure. If your AI agent is FDA-cleared, SOC 2 compliant, HIPAA-validated, or has gone through a formal audit process — that’s a moat that no amount of prompt copying can replicate. The competitor doesn’t just need to match your AI quality. They need to match your compliance posture. That’s a 12-18 month process minimum.
Domain expert validation loops. The best vertical AI agents have built feedback systems with actual domain experts — doctors reviewing medical coding suggestions, lawyers reviewing contract clauses, CPAs reviewing journal entries. These human-in-the-loop systems are expensive to build, slow to scale, and represent genuine accumulated intellectual property.
Retention forecast: 93-97% gross retention. This is where AI agent economics start to look like traditional enterprise SaaS. The companies building here will command premium multiples, and rightfully so.
Be Honest Which You’re Building. You May Have to Run Even Faster.
The honest self-assessment: look at your product and figure out which level you’re actually at. Not which level you want to be at. Not which level your pitch deck claims. Which level a smart buyer would put you at after using your product for six months.
- If you’re at Level 1, your entire company strategy needs to be about moving to Level 2 or 3 as fast as possible. That means building proprietary data flywheels and deep integrations, not shipping more prompt templates.
- If you’re at Level 2, the clock is ticking. You probably have 9-12 months before buyers start running annual bake-offs against you. Use that time to build measurable data advantages and prove them to customers quarterly.
- If you’re at Level 3, you’re in decent shape but not bulletproof. Invest in making your integrations deeper and harder to replicate, but watch out for AI-assisted migration tools that could compress your switching cost advantage.
- If you’re at Level 4, congratulations — you’re building the kind of AI company that will actually hold value long-term. Just make sure you’re not deluding yourself about being Level 4 when you’re really Level 2 with a compliance wrapper.
What This Means If You’re Buying
Know your leverage. If you’re buying a Level 1 agent, negotiate accordingly — short contracts, aggressive pricing, and explicit data portability clauses. You have all the power.
If you’re buying Level 3 or 4, the traditional enterprise buying playbook still applies. Longer commitments in exchange for better pricing make sense because your switching costs are real.
The worst mistake you can make is treating a Level 1 purchase like a Level 4 commitment. Three-year contracts for AI SDR tools at 2026 prices? That’s leaving money and optionality on the table.
Maintain your prompt library regardless. Even at Level 4, having your core instructions, brand guidelines, and domain knowledge documented outside your vendor’s platform is just good hygiene. It won’t make switching easy, but it will make it possible.
The AI agent market is sorting itself into these four levels whether vendors want it to or not. The founders who understand which level they’re actually at — and build accordingly — are the ones who will still be here in three years.
The ones who think a system prompt is a moat won’t be.
