AI Accuracy & Bias Disclosure

How Germinate.ai approaches the accuracy and fairness of the agents we build and run.

Version 1.0 · Effective January 12, 2026

1. Purpose

This disclosure explains, in plain terms, how Germinate.ai approaches the accuracy and fairness of the AI agents we build and run for our customers. It tells you what our agents can and cannot do, where the limits are, and what we do to manage them.

We would rather be specific about limitations than make claims we cannot stand behind. AI is useful, not infallible. This document is how we keep that honest.

2. Scope

This disclosure applies to all AI agents and features Germinate builds, deploys, or manages on behalf of customers on the Germinate platform, including the context layer that connects customer data, the agents that act on it, and the orchestration that runs them. It does not govern third-party tools a customer operates independently of Germinate.

3. How Our Agents Produce Results

Germinate agents combine large language models with your own connected data — databases, documents, and the systems your team already uses. An agent retrieves relevant information from your data, reasons over it with a model, and produces an output: a draft, an answer, a flag, a routed task. The quality of that output depends on the quality and coverage of the connected data, the model selected for the job, and the way the task is defined.

Because these systems are probabilistic, the same question can produce slightly different wording on different runs, and an agent can be confident and wrong at the same time. We design for this rather than pretend it away.

4. What We Say About Accuracy

Our commitments

• No guarantee of correctness. Our agents are decision-support and work-acceleration tools. They can make mistakes, omit relevant information, or misinterpret a request. We do not warrant that any output is complete, current, or correct.
• Grounded in your data, not invented. Wherever possible, agents are built to draw from your connected sources and to cite or surface what they used, so outputs can be checked against the record rather than taken on faith.
• Human accountability stays with you. For any consequential decision — financial, legal, contractual, safety-related, or affecting an individual — a qualified person should review the output before it is relied upon. Agents inform decisions; they do not own them.
• Fit for the job they were built for. Each agent is scoped to a specific task. Using an agent outside that scope is outside the conditions we tested for, and accuracy expectations no longer apply.

5. What We Say About Bias and Fairness

AI models learn from data, and data carries the patterns — including the biases — of the world and the organizations that produced it. An agent built on biased inputs can reproduce or amplify that bias. We treat this as a real risk to manage, especially anywhere an agent touches people: hiring, screening, evaluation, or access to a service or benefit.

Our commitments

• We name the higher-risk uses. Agents that influence decisions about people get extra scrutiny, additional testing, and mandatory human review. We will tell a customer when a proposed use falls into this category.
• We do not deploy agents to make autonomous decisions about individuals. Where an agent touches hiring, screening, or similar, it surfaces and ranks information for a human decision-maker. The decision remains a person's.
• We test for disparate results, within the limits of the data. Where it is feasible and lawful, we evaluate whether an agent produces materially different outcomes across groups, and we adjust, constrain, or decline to ship when we find a problem we cannot mitigate.
• We are honest about residual risk. No testing removes bias entirely. We tell customers what we checked, what we found, and what remains their responsibility to monitor in use.

6. Your Responsibilities as a Customer

Accuracy and fairness are shared work. To get reliable results and keep risk in check, customers are responsible for:

• Providing accurate, reasonably current, and lawfully obtained data for the agent to work from.
• Keeping a qualified human in the loop for consequential decisions, as described above.
• Using each agent for its intended purpose and telling us when the business changes in ways that affect it.
• Meeting the legal and regulatory obligations that apply to your industry and your use of the outputs.

7. Alignment with the NIST AI Risk Management Framework

Germinate's accuracy and bias practices are organized around the four functions of the NIST AI Risk Management Framework (AI RMF 1.0), and informed by the NIST Generative AI Profile (NIST-AI-600-1, July 2024). We reference the framework as a discipline; we do not claim certification, which NIST does not offer.

NIST AI RMF function	How Germinate addresses it
Govern	Accuracy and fairness commitments, roles, and review cadence are set in policy (this document) and owned by a named accountable role.
Map	Each agent's intended use, data sources, affected people, and out-of-scope conditions are documented before build. People-affecting uses are flagged here.
Measure	Agents are tested against representative cases, accuracy bars are agreed per task, and outcome/fairness checks are run and recorded where feasible.
Manage	Human-in-the-loop checkpoints, monitoring, failure triage, and scheduled re-review keep risk managed across the agent's life, not just at launch.

8. Review and Updates

We review this disclosure at least annually, and sooner when our models, methods, or obligations change materially. The version and effective date at the top of this document reflect the current edition.

9. Contact

Questions about this disclosure, or a specific agent's accuracy and fairness profile: policy@germinate.ai — or speak with your Germinate contact.

This disclosure describes our practices and commitments. It is not a warranty, and nothing here overrides the terms of a signed agreement between Germinate and a customer.

Related: Responsible AI · AI Ethics Policy