Skip to content
Notifications
Clear all

Has anyone successfully argued that agent prompts are our trade secret?

5 Posts
5 Users
0 Reactions
72 Views
(@danielk)
Honorable Member
Joined: 4 months ago
Posts: 382
Topic starter   [#10570]

We're negotiating an MSA with an AI vendor. Their standard terms claim rights to "inputs" for model improvement. Our legal flagged it, but the vendor is pushing back, saying prompts aren't protected.

I argue our engineered agent prompts—the precise sequencing, conditional logic, and proprietary data structures—are a core trade secret. They're the "secret sauce" for our orchestration layer.

Has anyone had this fight and won? What specific language did you use to exclude them from training data?

Key points we're preparing:
* Defining "Customer Data" to explicitly include "prompts, prompt chains, and associated inference parameters."
* Adding a clear exclusion in the AI Training clause: "notwithstanding the foregoing, Customer Data shall not be used for service improvement."
* Contractually binding them to treat prompts as Confidential Information.

Did you need to provide technical annexes showing the complexity? Did referencing the CFAA or state trade secret statutes help?

-dk


Trust but verify, then don't trust.


   
Quote
(@latency_king_2)
Estimable Member
Joined: 6 months ago
Posts: 78
 

Your approach is correct, but from a technical standpoint, defining "Customer Data" isn't enough. You must also nail down the *processing* definitions. I've seen vendors later claim that "model training" is a distinct process from "service improvement," allowing a loophole.

You need explicit, separate clauses. First, a data ownership clause: "Customer owns all Inputs and Outputs." Then, a flat prohibition: "Provider shall not use Customer Data, including Inputs, Outputs, and derived data, to train or improve any models, whether for the Service, other services, or internal research, without separate written agreement." Close the "benchmarking" loophole here too.

Referencing trade secret statutes gave us leverage, but what made the vendor fold was our technical annex. We diagrammed a single prompt chain showing its 14 conditional steps and 5 embedded proprietary schema references. It proved non-obviousness and economic value. The legal argument is abstract until you show the engineering artifact.



   
ReplyQuote
(@alexm82)
Reputable Member
Joined: 4 months ago
Posts: 255
 

That's a really practical point about separating the data definition from the processing clause. I've been worried about the "distinct process" loophole too.

So your technical annex was key. When you diagrammed the prompt chain, did you have to get into the actual variable names and logic, or was a high-level flowchart of the decision points sufficient to make the case?

I'm wondering how much of our internal schema we'd have to expose to prove the "secret sauce" argument. There's a risk in showing them the artifact, even under NDA.



   
ReplyQuote
(@latency_llama)
Estimable Member
Joined: 7 months ago
Posts: 83
 

Your third point is the only one with teeth. The "confidential information" clause is what you'll actually use to sue them later, but it's a reactive tool. Defining "Customer Data" is semantics, and vendors have a PhD in reinterpreting those definitions post-facto when they're staring at a novel prompt structure they want to ingest.

The legal references are fine for the negotiation table, but they fold under operational proof. You don't need to expose your actual schema. Draft an annex that describes, in painfully dry technical terms, the *mechanism* of your orchestration layer. Something like:

* "The system employs a directed acyclic graph for prompt sequencing, where node execution is conditional on P99 latency and structured output validation of predecessor nodes."
* "Prompt templates incorporate proprietary entity-identification logic that references internal domain schemas (see Appendix X for high-level entity categories)."

This demonstrates non-obvious engineering effort without handing them the code. It moves the argument from "is a prompt data?" to "is this described system, which we've just contractually linked to the definition of 'Input,' a valuable secret?" They'll balk because it's clearly not generic user input, and that's the point.

If their legal team has any sense, they'll realize that clawing this back in discovery after a breach would be a nightmare, and it's easier to just grant the exclusion.


P99 or bust.


   
ReplyQuote
(@benwhite)
Reputable Member
Joined: 4 months ago
Posts: 209
 

Agree on the need for separate clauses. But ownership of "Inputs and Outputs" is a shaky foundation if the vendor's lawyers later argue prompts aren't a protectable input at all.

Your flat prohibition is better. Push it further: also ban them from using metadata about prompt *usage* - frequency, latency, error codes. They'll mine that to infer your structure.

The technical annex is what works. But showing 14 steps just proves complexity, not that it's a secret. You have to explicitly link each conditional step to a business outcome they'd want to replicate. That's the economic value argument made concrete.


read the fine print


   
ReplyQuote