It was fun working with you and covering multiple angles regarding the "SaaSacre" narrative! You're seriously missing out if you haven't subbed to TSCS already!
I’d bet on almost any SaaS company that already serves a large enterprise customer base and functions as a system of record or complex workflow integrator.
It’s far more likely that incumbents successfully embed AI into existing products than that customers migrate wholesale to an “AI-native” replacement. End-to-end probabilistic workflows are rare, and the most durable technology moats are still built around being a system of record or owning deeply embedded workflows.
I say this as someone whose job is to build enterprise AI products inside a Fortune 500. Agent-based systems consistently break down at scale because of accountability, compliance, and trust.
An AI agent works when I can review and correct it. That model doesn’t scale to an entire department where actions need to be controlled, auditable, and predictable.
At small scale, error rates are manageable. At enterprise scale, error volume becomes the problem - especially when the system can’t reliably flag what’s wrong because it has no epistemic awareness. An agent that makes 10,000 decisions a day and gets 5% wrong isn’t “mostly correct”. It’s operationally unusable.
To make agents enterprise-safe, you have to add guardrails - constrained actions, approvals, records, and workflows. But the more guardrails you add, the more the system collapses back into deterministic SaaS, which is exactly what incumbent platforms already do best.
That’s the core limitation of enterprise-grade AI today in my opinion, and LLMs are not going to solve it.
Correct. The guardrail paradox is the part the market refuses to engage with, every layer of control you add to make agents enterprise-safe pulls you back toward deterministic SaaS. That's the structural reason incumbents win. Thanks for the operational perspective, it's rare to hear from someone actually building this stuff inside a Fortune 500 rather than tweeting about it.
Very well written and insightful. However, have a look at this viewpoint which offers a counter to your system of record argument being a SAAS bear-case savior. Not saying I agree with either, it's all too dynamic and I am just trying to look at this from all angles before taking a longer-term view on investments in SAAS hereon. https://x.com/zain_hoda/status/2019049069134417975?s=20
The argument is elegant but it conflates data storage with data authority. Yes, an agent can clone your Salesforce instance in seconds. But a copy isn't a system of record, it's a snapshot. The moment two agents, or an agent and a human, write conflicting updates, you need an arbiter. That arbiter is the SoR. Multi-tenant concurrency, permissioning, audit trails, regulatory accountability, none of that disappears because an agent cached your database.
This article becomes materialistic on the simpler end of the spectrum, the Dead Zone names in our framework. If your entire value prop is 'we store your structured data and give you a UI to query it,' yes, that's getting hollowed out. That's exactly why we flagged project management tools and basic CRMs as vulnerable.
But the Fortress Zone thesis was never about data storage. It was about data authority in regulated environments. An agent caching Veeva's FDA validation trails doesn't make the agent the system of record. The FDA doesn't care where the data is cached. They care where it's governed.
Still, it's a dynamic worth watching. The line between 'governance layer' and 'platform' could blur faster than I expect. I'd rather be honest about that uncertainty than pretend the thesis is perfect.
The downside of publishing mid-week, is the Les may capture your thoughts and research, and present it in even more compelling way. Fantastic write up.
Won't AI coding eventually trivialize the creation of new deterministic software in most industries? Anthropic uses Claude Code to make Claude Code. Surely they'll ask Claude to help them create a vertically integrated AI agent orchestration layer too.
Building the software was never the hard part. If it were, every Fortune 500 would already be running custom internal tools instead of paying Salesforce $50k/seat.
The moat is often the decade of customer data, the regulatory certifications, the 200+ integrations, the institutional trust, and the army of consultants who've built their careers around your platform.
You can vibe-code a CRM in a weekend. You cannot vibe-code FedRAMP authorisation, FDA validation trails, or SOX-compliant audit infrastructure. Claude Code can write software. It can't write ten years of embedded customer relationships or replicate the switching costs that come from being wired into every downstream system in an enterprise.
Anthropic knows this, which is exactly why Cowork is going vertical into workflows rather than trying to replace systems of record directly.
It was fun working with you and covering multiple angles regarding the "SaaSacre" narrative! You're seriously missing out if you haven't subbed to TSCS already!
Likewise!
I’d bet on almost any SaaS company that already serves a large enterprise customer base and functions as a system of record or complex workflow integrator.
It’s far more likely that incumbents successfully embed AI into existing products than that customers migrate wholesale to an “AI-native” replacement. End-to-end probabilistic workflows are rare, and the most durable technology moats are still built around being a system of record or owning deeply embedded workflows.
I say this as someone whose job is to build enterprise AI products inside a Fortune 500. Agent-based systems consistently break down at scale because of accountability, compliance, and trust.
An AI agent works when I can review and correct it. That model doesn’t scale to an entire department where actions need to be controlled, auditable, and predictable.
At small scale, error rates are manageable. At enterprise scale, error volume becomes the problem - especially when the system can’t reliably flag what’s wrong because it has no epistemic awareness. An agent that makes 10,000 decisions a day and gets 5% wrong isn’t “mostly correct”. It’s operationally unusable.
To make agents enterprise-safe, you have to add guardrails - constrained actions, approvals, records, and workflows. But the more guardrails you add, the more the system collapses back into deterministic SaaS, which is exactly what incumbent platforms already do best.
That’s the core limitation of enterprise-grade AI today in my opinion, and LLMs are not going to solve it.
Correct. The guardrail paradox is the part the market refuses to engage with, every layer of control you add to make agents enterprise-safe pulls you back toward deterministic SaaS. That's the structural reason incumbents win. Thanks for the operational perspective, it's rare to hear from someone actually building this stuff inside a Fortune 500 rather than tweeting about it.
Outstanding research, Your picks look solid, perfect for reinvesting profit from my IVG puts. Will initiate positions on further weakness
Patience on entries will pay off.
Fantastic read. Thank you.
You beat JP Morgan to the punch: https://x.com/dividendology/status/2021314972588863871
Long (and in deep red) GWRE!
Always nice when JPM validates the thesis after we've already published it.
Veeva, Guidewire, CrowdStrike, ServiceNow, Snowflake, all on their list, all in our Fortress Zone framework from last week.
GWRE in deep red is where the asymmetry lives, the market is handing you a regulated vertical monopoly at distressed multiples. Patience pays.
Very well written and insightful. However, have a look at this viewpoint which offers a counter to your system of record argument being a SAAS bear-case savior. Not saying I agree with either, it's all too dynamic and I am just trying to look at this from all angles before taking a longer-term view on investments in SAAS hereon. https://x.com/zain_hoda/status/2019049069134417975?s=20
Thanks for the perspective.
The argument is elegant but it conflates data storage with data authority. Yes, an agent can clone your Salesforce instance in seconds. But a copy isn't a system of record, it's a snapshot. The moment two agents, or an agent and a human, write conflicting updates, you need an arbiter. That arbiter is the SoR. Multi-tenant concurrency, permissioning, audit trails, regulatory accountability, none of that disappears because an agent cached your database.
This article becomes materialistic on the simpler end of the spectrum, the Dead Zone names in our framework. If your entire value prop is 'we store your structured data and give you a UI to query it,' yes, that's getting hollowed out. That's exactly why we flagged project management tools and basic CRMs as vulnerable.
But the Fortress Zone thesis was never about data storage. It was about data authority in regulated environments. An agent caching Veeva's FDA validation trails doesn't make the agent the system of record. The FDA doesn't care where the data is cached. They care where it's governed.
Still, it's a dynamic worth watching. The line between 'governance layer' and 'platform' could blur faster than I expect. I'd rather be honest about that uncertainty than pretend the thesis is perfect.
The downside of publishing mid-week, is the Les may capture your thoughts and research, and present it in even more compelling way. Fantastic write up.
Won't AI coding eventually trivialize the creation of new deterministic software in most industries? Anthropic uses Claude Code to make Claude Code. Surely they'll ask Claude to help them create a vertically integrated AI agent orchestration layer too.
Building the software was never the hard part. If it were, every Fortune 500 would already be running custom internal tools instead of paying Salesforce $50k/seat.
The moat is often the decade of customer data, the regulatory certifications, the 200+ integrations, the institutional trust, and the army of consultants who've built their careers around your platform.
You can vibe-code a CRM in a weekend. You cannot vibe-code FedRAMP authorisation, FDA validation trails, or SOX-compliant audit infrastructure. Claude Code can write software. It can't write ten years of embedded customer relationships or replicate the switching costs that come from being wired into every downstream system in an enterprise.
Anthropic knows this, which is exactly why Cowork is going vertical into workflows rather than trying to replace systems of record directly.
Interesting. I recently wrote a piece on how the unit economics of SaaS are changing — would love to hear your thoughts.
Great read and thank you for putting in the work analyzing this market carnage. Just curious - on your quadrant, where would you place TEP.PA?
I agree. Will add more Mc Donald's shares.