Own the logic, own the data: the CTO's case against point-SaaS AI

You survived SaaS sprawl. Now the same pattern is back with agents inside — and this time the vendor holds your workflows, not just your billing.

Every CTO of the last decade has lived through the same movie. A team has a problem; a SaaS point solution exists; the card gets swiped. Multiply by five years and forty teams, and you're running a portfolio of 200 subscriptions, an SSO spreadsheet, and a renewals calendar that functions as a hostage schedule.

That was annoying but survivable, because those tools mostly held your records. The AI wave is replaying the exact same pattern — an AI tool for support, another for sales emails, another for meeting notes, another for code — with one difference that changes the risk math entirely:

This generation of point-SaaS doesn't just hold your data. It holds your labor.

An AI vendor that runs your support conversations, drafts your customer emails, or automates your workflows isn't storing what your company did. It's doing what your company does. The prompts, the conversation policies, the escalation logic, the tribal knowledge you fed it — that's operational IP, accumulating inside a product you can't read, on infrastructure you can't inspect, under terms you don't control.

Rent a system of record and switching is painful. Rent your labor and switching means re-hiring your own company.

The four bills that come due

Lock-in, compounding. Classic SaaS lock-in was about data gravity — exporting records is hard. Agentic lock-in is worse because what accumulates is behavior: months of tuned prompts, corrections, and workflow wiring that exist only inside the vendor's black box. There's no export button for "how our agent handles angry enterprise customers." Two years in, the switching cost isn't migration; it's reconstruction from memory.

Pricing power you handed over. Per-seat, per-resolution, per-token-with-markup — AI point tools price against the value of the labor they've absorbed, and they reprice as you depend on them. Once an agent handles a workflow nobody on your team remembers how to run, your negotiating position at renewal is a formality. You taught the vendor's product your business, and now you're paying rent on your own knowledge.

The audit you can't pass. When a customer, a regulator, or your own board asks how an AI decision was made — what the agent saw, what it did, why — "we'd have to ask the vendor" is not an answer that survives contact with a serious review. Invisible logic on someone else's infrastructure means your compliance posture is whatever the vendor's roadmap says it is.

Data exhaust you don't control. Every conversation, ticket, and internal document that flows through a point tool lives under that vendor's retention policies, subprocessor list, and security posture — multiplied by however many AI tools your teams have adopted. Each one is a new perimeter you don't operate and mostly can't see.

None of this requires a villainous vendor. It's just the structure of the deal: their product gets smarter about your business; you get a subscription.

What ownership actually means

"Own your AI" gets thrown around loosely, so let's be precise. Ownership is four concrete properties:

1. The logic is readable and versioned. Your agent's behavior lives in files — plain text your engineers can read, diff, review, and roll back. Changing how the support agent escalates should be a pull request, not a vendor ticket. In rysh, an agent is a markdown skill file in your git repo. That's the whole trick: agent behavior as source code, with everything source code gives you — history, review, blame, portability.

2. The keys are yours. The model access belongs to you, not resold through a markup layer. Rysh runs on Claude with your Anthropic API key. Your usage relationship, your rate limits, your bill, your data-handling terms with the model provider — direct.

3. The runtime is yours to host. Self-hostable, on your infrastructure, inside your perimeter. (To be precise about what we are: rysh is self-hostable, not open source — the distinction matters and we won't blur it.) The agents that run your workflows should be processes you can point at, on machines you control.

4. The trace is yours. Every action every agent takes is observable and stored where you can audit it — not summarized in a vendor dashboard with a 90-day window. When the question is "what happened," the evidence is in your house.

Test any AI vendor — including us — against those four. The answers separate tools you're adopting from dependencies you're incubating.

"But point tools are faster"

The honest counterargument: the point-SaaS tool is live this week, and the owned platform is a project. Sometimes that trade is right — a disposable tool for a disposable workflow is fine.

But notice what the speed actually costs across a portfolio. Ten AI point tools means ten black boxes, ten data perimeters, ten renewal negotiations, ten different answers to "can we see what it did" — and zero compounding, because nothing learned in one tool transfers to the next. The eleventh workflow starts from zero, with an eleventh vendor.

A platform inverts that curve. One engine, one governance model, one audit story — and each new agent is a markdown file, not a procurement cycle. The first workflow takes longer than swiping a card. Every workflow after it is faster, because the foundation already exists and you own it. That's the difference between spending on AI and accumulating AI.

The portfolio question

You don't have to rip anything out this quarter. But run the audit:

1. List every AI tool doing real work in your company today.
2. For each: can you read the logic? Export the behavior? See the full trace? Leave without losing the capability?
3. Count how many answer "no" to all four — and ask what that list looks like at renewal time, two years from now, when the workflows are load-bearing.

Then pick the next workflow you were about to solve with a point tool, and build it on something you own instead. The logic in your repo. The keys in your name. The trace on your infrastructure. The leverage — finally — accruing to you.

We're taking on a handful of design partners — companies that want their agent stack owned, observable, and governed from workflow one. If that's you: rysh.ai/design-partner.

Related: Self-hostable, your keys, your data · Rysh vs building it yourself