Quick answer
“Proprietary AI” is not a technical claim. It is a label that honestly describes at least six different things, from a prompt library on someone else’s foundation model through to weights the vendor trained. Most sales-AI products sit in the middle tiers: orchestration, retrieval and outcome data. Ours sits at orchestration and outcome data, and we have not pretrained a foundation model. Five questions separate the tiers, and each has a checkable answer.
Why the phrase survives
“Proprietary” is a claim about ownership, not architecture. A system prompt you wrote is proprietary. So is a dispatch table that picks a channel. So is a model you pretrained. True in all three cases, which is why it is the default line on vendor sites: it survives legal review and lets a buyer assume the strongest reading.
That is not automatically dishonest. It becomes a problem when a buyer prices tier five, receives tier two, and finds out during a migration or a security review. The useful move is not to ask “is your AI proprietary” (the answer is always yes), but to ask which layer it sits in, and then to ask for an artefact.
European law already has a vocabulary for this split. Article 3(68) of Regulation (EU) 2024/1689, the EU AI Act, defines a “downstream provider” as “a provider of an AI system, including a general-purpose AI system, which integrates an AI model, regardless of whether the AI model is provided by themselves and vertically integrated or provided by another entity based on contractual relations” (EUR-Lex, Regulation (EU) 2024/1689). That is the category almost every AI sales platform occupies. The regulation writes it in because it expects the model and the system to have different owners.
The six tiers of “proprietary AI”
| Tier | What is proprietary | What is not | Artefact you can ask to see |
|---|---|---|---|
| 1. Prompt library | System prompts, persona and guardrail text | The model, the inference stack, everything below the prompt | A redacted prompt, or a version history showing review |
| 2. Orchestration and routing | Which model runs, which channel fires, when, and the fallback | The models being routed between | An architecture diagram naming routing conditions and fallbacks |
| 3. Retrieval over your data (RAG) | The ingestion, chunking and retrieval pipeline; the corpus is yours | The model weights, which are unchanged by retrieval | Index refresh cadence, chunking strategy, retrieval evals |
| 4. Adapted weights (LoRA / adapters, or full fine-tune) | The trained delta and the dataset behind it | The base model the delta attaches to | Base model and version, dataset description, held-out eval |
| 5. Pretrained model | The weights, the training corpus and the training run | Almost nothing below it | A model card with benchmarked evaluations, plus a hosting story |
| 6. Outcome data | The labelled record of what worked, per segment and channel | Model architecture, which is irrelevant to this asset | How outcomes are labelled, and what changed as a result |
Tiers 1 to 3: real engineering, no model ownership
Retrieval is the tier most often described in a way that implies training. It is not. The paper that named the pattern, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” by Patrick Lewis and colleagues, describes RAG as “models which combine pre-trained parametric and non-parametric memory for language generation” (arXiv:2005.11401). The parametric memory is the frozen model; the non-parametric memory is the index over your documents. Adding a thousand call transcripts to that index changes what the system can retrieve, and nothing about the weights.
That matters commercially. A retrieval pipeline over your CRM is defensible work and yours to lose if you switch vendors. It is just not a model.
Tier 4: adapters are not the same as a fine-tune, and neither is a new model
Hugging Face’s PEFT documentation describes parameter-efficient fine-tuning as methods that “only fine-tune a small number of (extra) model parameters” while “yielding performance comparable to a fully fine-tuned model” (Hugging Face, PEFT documentation). On LoRA it states that “The original pre-trained weights are kept frozen, which means you can have multiple lightweight and portable LoRA models for various downstream tasks built on top of them” (Hugging Face, LoRA conceptual guide).
Google Cloud draws the same split: “Parameter-efficient tuning, also called adapter tuning, enables efficient adaptation of large models to your specific tasks or domain”, against full fine-tuning, where “Full fine-tuning updates all parameters of the model” (Google Cloud, Introduction to tuning, Gemini Enterprise Agent Platform).
A vendor at tier four owns an adapter or a set of updated weights, not the base. The EU AI Act recognises the asymmetry: Recital 109 says that “In the case of a modification or fine-tuning of a model, the obligations for providers of general-purpose AI models should be limited to that modification or fine-tuning …”. The law scopes responsibility to the delta, because the delta is what you made.
Tier 5: rarer than the marketing suggests, partly because the door is narrowing
Pretraining a competitive general-purpose model is a capital exercise, not a product decision. The EU AI Act defines one at Article 3(63) as “an AI model, including where such an AI model is trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks …”. Very few sales vendors have done this, and those that have say so with a model card, not a tagline.
Even the weaker version of “we trained on your data” is getting harder to buy. OpenAI’s own supervised fine-tuning guide states, verbatim, “OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users, but existing users of the fine-tuning platform will be able to create training jobs for the coming months” (OpenAI, supervised fine-tuning guide, checked 28 August 2026). The same page lists supervised fine-tuning base models as gpt-4.1 generation snapshots, not the current flagship family. A vendor claiming a fine-tune now needs a date and a named provider on it.
Tier 6: the one nobody markets
Outcome data is the labelled record of what actually happened: which opener led to a booked meeting in which segment, which follow-up interval recovered a dead lead, which objection handler lost the call. It is expensive to accumulate: you only get it by running real campaigns and recording the losses.
It is also the asset a competitor cannot buy. Base models are commoditising and adapters are portable. A multi-year, correctly labelled record of outcomes in one vertical is not, and it is what makes an evaluation set possible at all. Our view: the most defensible tier on the ladder, and the least talked about, precisely because it does not sound like AI.
The five questions that separate the tiers
| Question | A tier 4-5 answer sounds like | A tier 1-3 answer dressed up sounds like |
|---|---|---|
| What base model is under this, and what version? | A named model and version, and the change process | “A blend of models”, or “that’s proprietary” |
| Who hosts inference, and in which jurisdiction? | A named provider and region, or a private deployment | “Our secure cloud” with no provider named |
| What happens if that provider retires the model? | A migration path, a re-evaluation plan, a notice window | “That won’t affect you” |
| What exactly was trained, on what data, and when? | Base model, dataset, method (full fine-tune or LoRA), date | “Trained on billions of sales conversations” |
| Is the “learning” retraining, or prompt and state updating? | A clear statement of which, plus a change log | “It gets smarter with every call” |
Add a sixth: ask for an artefact. The idea comes from “Model Cards for Model Reporting” by Margaret Mitchell and colleagues, which proposed that “Model cards are short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions” (arXiv:1810.03993). A tier five vendor has one. A tier two vendor has an architecture diagram, which is a fine answer if they say so.
The deprecation question is the one buyers skip
If a vendor is a downstream provider, their roadmap has a dependency they do not control, and the terms of that dependency are published. OpenAI’s deprecations page sets minimum notice periods before model retirement — at least 6 months for generally available models, at least 3 months for specialised variants of them — and states that these apply unless “safety or compliance concerns require a faster timeline”. Preview models sit outside that floor: they “may be retired with much shorter notice, such as 2 weeks” (OpenAI, Deprecations). Anthropic says it provides “at least 60 days’ notice before model retirement for publicly released models” (Anthropic, Model deprecations).
This bites hardest at tier four, because a fine-tune is pinned to the base model it was trained from. OpenAI’s fine-tuning guide states that “All fine-tuned models will remain available for inference until their base models are deprecated” (OpenAI, supervised fine-tuning guide). The base model can be retired underneath the adapter, and the adapter goes with it.
All three provider pages were checked on 28 August 2026. Your vendor’s notice window is bounded by their provider’s, and prompts tuned against one model version do not automatically hold on the next. Ask what their re-evaluation process looks like when a model is retired. That tells you more about engineering maturity than any architecture claim, and pairs with our voice AI vendor security questionnaire.
Where Zian sits, exactly
We have been running outbound acquisition since 2017. The learning engine tracks around 420,000 data points across more than 10,000 leads a day. PrecisionPitch AI continuously split-tests scripts and approaches; SmartReach AI orchestrates message, channel and timing.
We have not pretrained a foundation model and we do not claim to have one. On the ladder above we are at tier two and tier six: orchestration, and the outcome data behind it. When you read “proprietary” on our site, that is what it refers to — not weights we trained.
We also offer private model deployment on customer infrastructure. That is a hosting and data-residency choice, not a claim about who trained the weights, and the two are worth keeping separate when you evaluate anyone, us included. More in private AI deployment for sales agents.
Zian is in waitlist beta: no public pricing, no free trial and no self-serve signup. Any comparison you build should account for that.
How we sourced this
Every technical distinction above is sourced to a primary document we opened on 28 August 2026: Hugging Face PEFT and Google Cloud model tuning documentation for adapters against full fine-tuning, Lewis et al. for parametric against non-parametric memory, Mitchell et al. for the model card artefact, Regulation (EU) 2024/1689 via EUR-Lex for the definitions, and OpenAI and Anthropic documentation for fine-tuning availability and deprecation notice.
What we could not verify and therefore did not claim: any figure for how many AI sales vendors use the phrase “proprietary AI”, because no census exists; and the internal architecture of any competitor, because it is not published. We have named no vendor but ourselves and the model providers we quoted. Zian’s tier assignment is our own, stated so you can hold us to it.
Frequently asked questions
Does a sales vendor need its own foundation model to be good?
No. In outbound sales, outcomes are usually decided by orchestration, retrieval quality, latency, compliance handling and the accumulated record of what worked. A vendor with a frozen frontier model and good feedback engineering will generally beat one with mediocre custom weights.
Can a vendor still fine-tune a frontier model on my data?
It depends on the provider and the date, and access is narrowing. OpenAI’s own documentation currently states, verbatim, “OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users, but existing users of the fine-tuning platform will be able to create training jobs for the coming months” (OpenAI, supervised fine-tuning guide, checked 28 August 2026). If a vendor claims a fine-tune, ask which provider, which base model and when the job last ran.
What happens to my AI agent when the underlying model is retired?
Your vendor has to migrate, and prompts tuned against one version do not always hold on the next. Provider notice windows bound the warning they get: OpenAI publishes minimum notice of at least 6 months for generally available models and at least 3 months for specialised variants, unless safety or compliance concerns require a faster timeline, and preview models can be retired with much shorter notice, such as 2 weeks. Anthropic says it gives at least 60 days for publicly released models. Ask what their re-evaluation process is, not whether it will happen.
Is a private or on-premise deployment the same as a proprietary model?
No. Private deployment answers where inference runs and who can see the data. It says nothing about who trained the weights. A vendor can deploy an open-weights model into your VPC without having trained anything, and that may be exactly what you need if your requirement is data residency.
Does “our AI learns from every call” mean the model is being retrained?
Usually not. Most of the time it means prompts, retrieval content, routing rules or conversation state are being updated by engineers based on outcome data. That is real improvement, but it is a different mechanism with a different failure mode. We covered the mechanics in do AI sales agents actually learn, which is the companion piece to this one.
How should I compare vendors once I know their tier?
Compare on outcomes and operating constraints rather than architecture, because architecture is an input, not a result. Our notes on AI SDR benchmarks and vendor evaluation cover which numbers are worth asking for and which are meaningless without a denominator.
Where to go next
Run the five questions across every vendor on your shortlist, us included, and write the answers verbatim. Who answers precisely and who answers with adjectives sorts the list faster than any feature matrix. If you want Zian on that list, Apply For Partnership and we will answer all five in writing before you ask.