How Widrow Works — and Who Controls the AI

Widrow does not build the intelligence model.

OpenAI, Anthropic and other supported AI providers supply the intelligence.

Widrow is the platform between that intelligence and your business. It connects AI to your people, your company knowledge and the systems you already use, then governs what that AI is allowed to know and do.

Widrow is the platform that puts the people who run the business in control.

The architecture in one minute

There are three distinct layers.

1. The AI provides the intelligence

Your business chooses the AI provider it wants to use.

Widrow currently supports frontier models from providers including OpenAI and Anthropic, either directly or through supported services such as Azure OpenAI.

The AI model does the reasoning, writing, analysis and other intellectual work.

Widrow does not replace that intelligence and does not require you to buy AI usage from Widrow.

Your business brings its own AI account and credentials.

If you change AI provider, the Widrow layer remains: your Digital Employees retain their roles, permissions, budgets and company knowledge.

2. Widrow provides the business layer

Widrow provides the things a general-purpose AI model does not know or control by itself:

• your company knowledge
• user identity
• staff roles
• knowledge access levels
• tools
• business-system connections
• permissions
• token budgets
• approval rules
• Digital Employee configuration
• audit and activity records
• escalation to people
• memory
• governance

This is what turns a frontier AI model into something that can operate safely inside a real business.

3. Your business decides the authority

The AI model does not decide what it is allowed to access or do.

Your business does.

Widrow enforces those decisions.

A Digital Employee can only use the tools its role has been assigned, can only reach the knowledge its access level allows, and can only use connected systems that have been authorised.

Permission is granted. It is not assumed.

What can a Digital Employee know?

Only the information your business makes available to it.

Company Knowledge can be synchronised from sources your business already owns, including websites, SharePoint, Google Drive, OneDrive, Dropbox and uploaded files.

The original documents remain in their existing source.

Widrow maintains a synced copy for use by the business’s Digital Employees.

Knowledge access is then governed by access level.

For example, a staff-level assistant cannot read management-level company knowledge.

A business can therefore make finance material available to authorised finance or management staff without exposing that information to every employee.

Where connected sources themselves have restricted permissions, restricted material can be kept out rather than indexed.

Company Knowledge is separated by tenancy and is not pooled with other Widrow customers.

Company Knowledge is one of the four kinds of memory in Widrow’s digital brain — see how the digital brain works.

What can a Digital Employee do?

That depends on two things:

what the AI model is capable of doing

and

what your business has authorised it to do through Widrow.

A capable frontier model may be able to write an email, analyse a spreadsheet, update a CRM record, research the web or perform many other tasks.

That capability does not automatically give it authority.

Widrow controls the tools and connections through which those actions can occur.

Open a role in Widrow and assign the tools, skills and connected applications that role may use.

A tool that has not been assigned cannot be called.

The permission check is enforced by the platform rather than being left to an instruction inside the AI conversation.

What happens by default?

Widrow starts from restricted authority rather than broad autonomy.

Digital Employees do not begin with permission to do everything.

Tools, access and connections must be granted.

And actions that delete or cancel something pause for a person’s confirmation before they run — the platform stops and asks, every time, enforced by the platform rather than by an instruction in the conversation.

Email provides a simple example.

A personal assistant can read authorised email and prepare replies, but replies are placed into the user’s Drafts folder.

A person reviews the email and presses Send.

The AI does the work.

The person retains the authority.

Can AI operate autonomously?

Yes — where the business deliberately chooses to allow it.

Widrow is designed so that autonomous authority is something the business grants, rather than something the AI assumes.

A business may decide that a low-risk recurring activity can run automatically.

Another action may remain subject to human approval.

The important distinction is that model capability and business authority are separate.

An AI model may be capable of performing an action without being authorised to perform it.

What happens when the AI does not know?

Widrow’s Digital Employees are designed to hand unresolved matters to a person rather than invent an answer or continue beyond their authority.

For customer-facing Digital Employees, a conversation that requires human involvement can be handed over while capturing the person’s details for follow-up.

Unanswered questions can also become useful feedback about gaps in the business’s Company Knowledge.

Who controls access to company information?

The business does.

Widrow supports separate knowledge-access levels and role-based permissions.

A tenant administrator can determine which people or groups receive access to particular levels of company knowledge and which tools their role can use.

Content can also be excluded from Digital Employee knowledge entirely.

If something should not be available to AI, it does not have to be included.

Can access be removed?

Yes, easily and quickly.

Business-system connections are visible in the Widrow portal and can be disconnected.

Removing a connection removes that route of access.

Roles and permissions can also be changed.

Configuration changes are versioned, allowing administrators to see changes and roll back to an earlier configuration if required.

Is activity recorded?

Widrow logs the use of connected systems and maintains an audit trail of activity.

Configuration changes are versioned.

Requests across the platform carry correlation identifiers so actions can be traced through the system.

The purpose is simple:

if AI is being allowed to operate inside a business, the business should be able to determine what happened afterwards.

Who owns the AI account?

You do.

Widrow does not require your business to use a Widrow-owned OpenAI or Anthropic account.

Your business provides its own supported AI credentials and is billed directly by the AI vendor.

Widrow therefore acts as the governance and connection layer rather than as an AI-token reseller.

It also means the AI provider is replaceable.

The business can change the intelligence provider without rebuilding the business knowledge, roles and governance that sit above it.

Where does company knowledge live?

Company Knowledge is stored within the business’s Widrow tenancy.

Company Knowledge is used to ground that business’s own Digital Employees and is never used to train any AI model — not Widrow’s, not anyone’s. That commitment is written into Widrow’s Data Processing Agreement.

The original documents remain in the systems the business already owns.

Widrow’s document connectors are used to maintain the synced knowledge copy rather than replacing the original source.

If the business leaves Widrow, those originals remain untouched.

What about employee privacy?

Company Knowledge and personal assistant memory are different things.

Company Knowledge belongs to the business and is governed by business access controls.

An individual’s assistant memory is private to that individual.

Tenant and partner administrators cannot open and read another person’s personal assistant memories.

Each person can review and remove their own stored memories.

A tenant administrator can export everything held for a person as a file, or erase it, from Tenant Portal → Settings → Privacy Requests — with no browse view of anyone’s memories. The administrator’s view shows counts and dates, never content.

What protects the platform itself?

Widrow runs on Microsoft Azure in the Australia East region.

Widrow states that the platform uses:

• TLS 1.2 or later for data in transit
• AES-256 encryption for data at rest
• Azure Key Vault for secrets
• automated database backups and point-in-time recovery
• role-based access
• isolation between customer tenancies
• correlation identifiers for tracing requests
• two-factor authentication for portal access

The WordPress bridge component, def-core, is open source, allowing that part of the system to be inspected publicly.

The main Digital Employee Framework repositories remain private.

A Data Processing Agreement is available from Widrow.

What Widrow is — and what it is not

Widrow is not a new LLM competing with GPT or Claude.

It is not simply a chat interface placed over an LLM.

And it does not make the AI model itself trustworthy by declaration.

Widrow’s job is different.

It provides the controlled environment in which frontier AI is allowed to operate inside a business.

The AI supplies intelligence.

Widrow supplies company context, identity, tools, connection, permissions and governance.

The business determines the authority.

That is the architecture.

What is a platform fact, and what is independent assurance?

Not every statement about a software platform belongs in the same category.

Some statements describe how Widrow is built and how it operates.

Examples include:

• tools must be assigned before they can be used
• actions that delete or cancel something pause for a person’s confirmation
• knowledge access is controlled by role and access level
• email remains in Drafts until a person sends it
• autonomy must be explicitly granted
• activity is logged
• access and connections can be revoked
• the business supplies its own AI credentials

These are concrete platform behaviours and architectural controls.

Other statements concern the underlying infrastructure, security design and operating environment.

Widrow publishes those technical details so customers and advisers can understand how the platform is intended to operate.

That is different from independent certification.

If Widrow has not obtained a particular certification, audit or external assurance, we do not describe it as though it has.

A prospective customer with specific regulatory, security or procurement requirements should, in conjunction with their implementing Partner, evaluate Widrow against the standards that apply to their organisation.

Why this matters

Giving staff access to powerful AI is easy.

Governed business AI is harder.

A frontier model may be capable of reading, writing, analysing, researching and acting across many systems.

Capability alone does not determine whether it should be allowed to do those things.

Widrow separates intelligence from authority.

The AI model provides the capability.

Widrow controls the knowledge, tools, connections and permissions through which that capability reaches the business.

Management decides how much authority to grant.

That means the business can start with tightly controlled AI assistance and expand authority over time where it makes sense to do so.

The short version

OpenAI, Anthropic and other supported providers supply the intelligence.

Widrow connects that intelligence to your people, company knowledge and business systems.

Widrow governs what the AI can know and do.

Your business controls the authority.

Widrow is delivered by a Widrow Approved Partner. Why that matters, and what your Partner does.