What we do

General-purpose AI assistants are capable, but they forget context between sessions, they don’t stay inside defined boundaries, and they can’t account for what they did. For someone processing a logistics manifest or reviewing a compliance filing, any one of those gaps rules the tool out.

Our agents are built to close all three gaps, in every session and for every user.

Every product we build runs on a single engine that we designed and own. Customers don’t buy it separately; every product inherits it. It gives each agent four properties in every deployment.

Identity

Each agent receives a unique, verifiable identity when it starts working. Every action it takes is attributable to that agent, working for a specific user, in a specific organization, at a specific moment.

Scope

Each agent works within a defined set of actions, systems and data. An agent set up for logistics dispatch can’t reach the finance system, and an agent working for one organization can’t reach another organization’s data. The architecture enforces these limits, so no instruction can override them.

Policy enforcement

Every action passes a governance check against the organization’s policies before it runs. Rules that leave no room for interpretation are enforced deterministically, without an AI model making the call. Rules that need context get a level of review matched to the risk of the action. Agents have no way around the check.

Audit

Every governance decision, whether permitted, denied or flagged, becomes a permanent record. At any time an operator can see what was attempted, by which agent, for which user, under which policy, and with what result.

Each agent works in its own isolated workspace and can’t see anything outside it. Files cross the boundary only when a person moves them.

Inside the workspace, work is organized by project, status and workflow instead of folders. The agent keeps it in order, so people can ask for things the way they think of them, such as “the contracts we reviewed last month”.

The workspace keeps an encrypted, versioned history. A user can return the whole workspace, or a single document, to any earlier state. Files an agent deletes go to a trash the user can recover from, and permanent deletion needs the user’s confirmation.

Our agents keep three kinds of memory. Operational memory holds what is in progress now. Conversational memory keeps the full record of what was said and agreed, in the original words. Document memory covers every document imported or produced, searchable down to a single clause or figure.

Each session starts with the agent already aware of where the work stands. The user doesn’t have to brief it again.

When a recurring situation comes up, the agent follows the organization’s own documented procedure. It also uses the organization’s own definitions, so “revenue” or “a completed dispatch” means what that organization says it means.

Within each industry, role-based configurations set what a logistics operator, a compliance officer or a commercial manager may do and see. We reach that level of fit by capturing the organization’s own knowledge, without fine-tuning a model for each customer.

People direct the agent in conversation and do the work in documents, in the same application, without switching tools. Someone who likes to delegate by chat can do that. Someone who prefers to work in documents can keep the agent in the background. The balance can shift as their trust in the agent grows.

Our products work in assistance mode. The agent supports, executes and advises, and the professional makes the decisions.

Our products are designed to run in the cloud, in a private cloud, on-premises or in air-gapped environments, with the same governance and capabilities in each. For organizations that need full data sovereignty, agents can run on language models hosted inside the customer’s own infrastructure.

Private deployment is supported by the architecture today. Its commercial packaging, support model and procurement documentation are still in development.

Each customer works in an isolated tenant. No data, query, agent action or memory crosses from one customer to another.

We don’t monetize customer data. Product telemetry stays within each customer’s tenant, and any use of customer data to improve the product requires the customer’s explicit consent.

Customers can ask us to delete their data. Learned memory can keep residual traces that can’t be removed one by one; where that happens, we tell the customer exactly what was and wasn’t deleted.

We apply the GDPR as our baseline in every market. In Mexico we map our practices to the federal law on personal data held by private parties (LFPDPPP) and honor the rights of access, rectification, cancellation and opposition. We are working toward SOC 2 Type II and ISO 27001 certification and don’t hold either yet.

Our research serves the product. We work in four areas.

Governed agentic systems

Governance that the architecture enforces, so an agent can’t reason its way around an instruction. Open questions include how to govern agents that hand work to other agents across trust boundaries, and how to make any agent action within a set window reversible.

Specialization by field

Training for professional-grade reliability within one field. In regulated work, a model that is right 95% of the time isn’t good enough, because the errors tend to land on the high-stakes cases.

Private and sovereign deployment

Running governed agents, language models included, entirely inside a customer’s infrastructure, with no external logging service, cloud-hosted policy engine or network connection for review.

Human and agent collaboration

How professionals' trust and delegation change over months of real work. We study it with customer consent and under strict privacy controls.

Over the longer term we are building a family of language models trained for the fields we serve. It is a multi-year investment and is not yet part of any product.

We publish findings on human and agent collaboration, governance and evaluation when they help the field without disclosing proprietary detail, and we take part in open standards work on agent governance.

• Logistics and supply chain
• Legal and compliance
• Financial operations
• Broadcast and media

See RE:SOURCE Contact us