Does SIGNLD train on your data

No. SIGNLD is never trained on your data. Inference runs on a single-tenant AWS Bedrock instance, and AWS Bedrock contractually guarantees that customer inputs and outputs are never used to train or improve foundation models, and are not shared with model providers.

By Christopher Rafter · · 8 min read · Product explained
Does SIGNLD train on your data

In this article

What does "never trained on your data" actually mean

"Never trained on your data" means the content of your questions, your connected records, and the answers SIGNLD returns are never used as training examples to adjust the weights of any underlying model. The model behaves the same way after answering your questions as it did before.

For the wider context, see our overview of what SIGNLD is.

This is different from a system quietly improving itself using customer conversations, which is a pattern some consumer AI products use. SIGNLD's inference layer does not do that. What one customer asks, and what data that customer connects, does not change how the model responds to any other customer, or to that same customer later.

What is the mechanism behind that guarantee

The guarantee is not just a policy statement, it is built into where inference runs. SIGNLD's AI inference happens on a single-tenant AWS Bedrock instance, a private LLM powered by AWS Bedrock and dedicated to one customer's account rather than shared infrastructure.

AWS Bedrock's terms state that customer inputs and outputs processed through Bedrock are never used to train or improve the underlying foundation models, and are not shared with the model providers that built those foundation models. That contractual guarantee is what backs the claim, not just SIGNLD's own configuration. Because the instance is single-tenant, there is also no cross-account exposure at the infrastructure level. Read the full technical breakdown at /security, and see SIGNLD's broader position on responsible AI use at /legal/responsible-ai.

Does SIGNLD send data to third-party AI APIs

No. All AI processing for a Decision Brief happens inside the single-tenant AWS Bedrock instance. SIGNLD does not route customer data through a separate, external AI API from another vendor as part of generating an answer.

This is a meaningful distinction for companies with data governance rules about where financial, customer, or operational information can be processed. The inference step stays inside one dedicated, contracted deployment rather than passing through an additional vendor with its own separate terms.

How does the read-only connection fit in

Every source connection SIGNLD makes, whether to a CRM, an accounting platform, or a spreadsheet, is read-only. SIGNLD reads records to build the Knowledge Graph and cannot write back to any connected system.

This read-only design also limits what ever reaches the model. Only the data needed to answer a specific question is drawn into a prompt at query time, not an unrestricted feed of every connected system into a training pipeline. There is no training pipeline at all in this picture, only a read step and an answer step.

Is retrieval at query time the same as training

No, and this distinction matters. When SIGNLD answers a question, the Knowledge Graph supplies relevant context, records, and matched entities to the model as part of that specific request. This is retrieval, not training: the model uses the context to generate an answer, then the context is gone from the model's weights.

Training would mean the model's underlying parameters are permanently adjusted using your data, so that your data influences how the model responds to other customers or to unrelated future questions. That does not happen. Each question gets fresh context from the graph, and the model itself is not modified by having answered it.

The practical difference shows up if you ask the same question twice, a week apart, after connecting a new system in between. The second answer can differ because the graph now has more context to retrieve, not because the model learned anything new about how to answer in general. Retrieval changes what the model sees for one request; training would change how the model behaves for every request, and SIGNLD's inference layer does not do that.

Closing a Topic follows the same logic. It ends the active thread of questions and Decision Briefs, but triggers no additional processing of the data behind it. The context retrieved to answer questions inside a Topic was already discarded from the model's working context after each individual answer, well before the Topic itself was closed. If a Topic is reopened later, the underlying Knowledge Graph still reflects whatever systems remain connected, so new questions are answered against current data rather than a frozen snapshot.

Is my activity logged

Yes. Every query run through SIGNLD is logged and timestamped, including the question asked, the Topic it relates to, and the Decision Brief returned. This logging exists for traceability and account administration, not as an input to model training. An administrator can review the full history of questions asked against any Topic, and this log sits separately from the inference layer and is not used to retrain or fine-tune the model.

Each query generates a log entry at the moment it is submitted, recording the question text, the account and Topic it belongs to, and the time it was asked, with the Decision Brief logged against the same entry along with its evidence links and confidence score. This log is stored as account activity data, separate from the Bedrock instance that performed the inference. Because logs are timestamped, they support after-the-fact review: if a decision is questioned months later, the log shows precisely what was asked and what evidence the answer relied on.

What should you ask any AI vendor to confirm the same guarantee

If you are evaluating any AI product on this dimension, not just SIGNLD, there are a few specific questions worth asking rather than accepting a general privacy statement. Ask whether inference runs on shared infrastructure or a dedicated, single-tenant deployment, since shared infrastructure is where cross-customer exposure risk usually lives. Ask directly whether customer inputs and outputs are ever used to train or fine-tune the underlying models, and for the contractual language that backs that answer rather than a marketing claim. Ask whether any third-party API sits in the path of a typical request, since each additional vendor is a separate set of terms to verify. Finally, ask what is logged, where it is stored, and who can access it. A vendor that can answer all four with specifics, rather than general assurances, is one where you can verify the guarantee instead of trusting it.

How does this map to an internal data governance review

Most internal data governance reviews ask some version of the same questions above, along with where data physically resides and what compliance certifications apply. SIGNLD's answers map cleanly onto that structure: inference runs on a single-tenant AWS Bedrock instance, not shared infrastructure; no third-party AI API sits in the request path; customer inputs and outputs are never used for training, per Bedrock's own terms; and account activity is logged separately from the inference layer, with SOC 2 Type II controls covering the broader environment.

A governance reviewer comparing SIGNLD against an existing vendor checklist can generally substitute these specifics directly into their standard questions about model training, data residency, and third-party data sharing. For companies that require a completed vendor security questionnaire as part of that review, the same details, along with the certifications, are documented at /security.

Key takeaways

  • \"Never trained on your data\" means the content of your questions, your connected records, and the answers SIGNLD returns are never used as training examples to adjust the weights of any underlying model.
  • Every source connection SIGNLD makes, whether to a CRM, an accounting platform, or a spreadsheet, is read-only.
  • Most internal data governance reviews ask some version of the same questions above, along with where data physically resides and what compliance certifications apply.
  • Improvements come from how the Knowledge Graph resolves entities and how questions are processed, not from adjusting the model with customer data.
  • See the single-tenant inference model and read-only connections work on your own systems.

FAQ

Does AWS Bedrock share my data with the companies that built the foundation models

No. Under AWS Bedrock's terms, customer inputs and outputs are not shared with the underlying foundation model providers, and are not used to train or improve those models.

Is SIGNLD's AI a shared model across all customers

Inference runs on a single-tenant AWS Bedrock instance dedicated to each customer's account, separate from other customers' inference traffic.

Does connecting more systems mean more of my data is used for training

No. Connecting more systems gives the Knowledge Graph more context to retrieve at query time. It does not create any new training exposure, because no training happens.

If SIGNLD doesn't train on my data, how does it get more accurate over time

Improvements come from how the Knowledge Graph resolves entities and how questions are processed, not from adjusting the model with customer data.

Does SIGNLD use any third-party AI APIs at all

No. AI processing happens within the single-tenant AWS Bedrock instance. There is no separate third-party AI API in the path of a Decision Brief.

Where can I read SIGNLD's formal position on responsible AI

The full policy is at /legal/responsible-ai, and the technical security architecture is documented at /security.

Related reading in this series: How does SIGNLD work and How long does SIGNLD take to set up.

Try SIGNLD free

See the single-tenant inference model and read-only connections work on your own systems. Try SIGNLD free.