Responsible AI
Effective: April 2026
Last updated: October 2026
On this page
Klivo builds agentic AI for enterprises, including in regulated sectors like banking, insurance, healthcare, and IT services, where the quality and safety of an AI system's behavior matters. This page describes how we approach responsible AI: the principles we hold, and the practices we apply so that the agents we deploy are useful, controllable, and accountable. We describe our approach honestly and at the scale we operate. We calibrate the depth of these practices to the risk of each deployment, and we work through the specifics with each customer rather than applying a one-size-fits-all process.
OUR APPROACH
1. Principles
Five principles guide how we design and deploy agents.
Grounded, not guessing.
Wherever possible, agents draw on the customer's own documents, records, and systems, and can show what they relied on. We treat confident answers with no basis as a failure to design against, not an acceptable outcome.
Clear that it is AI.
End users should understand when they are interacting with an AI agent and should have a path to reach a person. We build for that transparency and configure it with each customer.
Human oversight where it matters.
For actions with legal, financial, clinical, or other significant consequences, we recommend and support keeping a person in the loop. The exact level of human oversight is configured per deployment, based on the customer's risk and requirements.
Controllable by design.
Agents operate within guardrails set with the customer, including what they can access, what actions they can take, and where they must stop and defer.
Privacy by default.
We collect the minimum personal data needed, keep customer data isolated, and support redaction or tokenization of sensitive fields where appropriate. Our data handling is described in our Privacy Policy and DPA.
GOVERNANCE
2. How we govern AI across its lifecycle
We organize our responsible-AI work around four functions, drawn from the widely used NIST AI Risk Management Framework: govern, map, measure, and manage. This gives us a consistent way to think about risk from first scoping through live operation.
Our responsible‑AI practices follow the NIST AI Risk Management Framework functions: govern, map, measure, and manage.
Govern.
We set clear ownership and guardrails for each deployment, agree with the customer what the agent may and may not do, and document intended use and known limitations.
Map.
Before building, we work with the customer to understand the use case, the people it affects, the data involved, and where things could go wrong, so risk is considered in scope, not discovered in production.
Measure.
We test agent behavior proportionate to the risk of the deployment, checking that the agent stays grounded, stays within its guardrails, and handles realistic cases and edge cases before it goes live.
Manage.
Once live, we monitor behavior, keep a path to adjust or roll back configurations, and respond to issues with the customer. Higher-risk deployments get closer oversight.
IN PRACTICE
3. Keeping agents accurate and grounded
To reduce incorrect or fabricated output, we lean on the customer's approved sources rather than a model's open-ended recall. Where it fits the use case, agents retrieve from customer-approved documents and data, use structured tool calls for factual lookups, and follow constrained formats for sensitive actions. Where an agent is not confident, we design for it to say so, defer, or hand off to a person, rather than present a guess as fact. Because AI output can still be wrong, we are clear with customers that human review is appropriate for high-consequence use.
4. Model selection
We are not tied to a single model provider. For each customer and use case, we select the model we consider best suited to the requirement, balancing accuracy, latency, cost, privacy, and deployment constraints. Customers can require a specific model or a specific deployment region where they have their own obligations, and we document the choice and its known limitations.
5. Transparency and user choice
We build so that end users are told when they are dealing with an AI agent and are given a clear route to a person. Where an agent provides substantive answers from source material, we support showing what it relied on. Customers can configure default paths for users to opt out of AI-assisted handling and reach human service, set per channel.
6. Human accountability
AI supports people at Klivo's customers, it does not replace their accountability. Customers remain responsible for how they use the agents and their output, and we design deployments so that meaningful human control is available for consequential decisions. We describe the limitations of AI output plainly, including in our Terms of Service.
7. Raising a concern
If you believe an AI agent built or deployed by Klivo behaved in a way that was unsafe, unfair, or otherwise concerning, tell us at support@getklivo.com. We take these seriously, review every concern raised, and work with the affected customer to understand and address it. As our practices mature, we expect to formalize and expand these commitments.