Products / Under explorationAn open product workbench

From a good method
to a useful tool.

We are exploring product directions around recurring needs in secure AI workflows. There are no released products yet. Each direction needs evidence of usefulness, a defined user and a viable operating model.

Walk through the concept
F / 02Proposed direction
01 / Intent02 / Authority03 / Evidence
Between intent and action, a decision.
Directions under exploration. No released products yet.
01 / Interactive illustration

Same task. Different responsibility.

Imagine a support assistant with access to one conversation. Choose an action to explore how its authority and decision record could change.

A concept illustration using a fictional example. It does not run an agent, send messages or test security controls.

What should the assistant do?
01 / Read

One conversation. Read access.

Summarize the selected customer conversation.

A / Proposed boundary

Read only, within one case

Access covers one conversation. No changing records, browsing other cases or contacting the customer.

Why it matters

Being able to read something should not, by itself, confer permission to act on it.

B / Proposed decision record

What a record could include

  1. Source reference and version
  2. Scope of access at the time of reading
  3. System version and summary outcome
Less data. A clearer record.

References and versions may be enough instead of copies of entire conversations. Scope and retention would need a separate decision.

02 / Propose

Ready for review. Not for sending.

Prepare a reply based on this conversation.

A / Proposed boundary

Prepare a draft

The assistant could prepare a draft. It could not send it or treat it as approved.

Why it matters

Preparing a proposal and taking an action are two different permissions.

B / Proposed decision record

What a record could include

  1. Sources used to prepare the draft
  2. Version of the proposed reply
  3. Status of human review
Less data. A clearer record.

References and versions may be enough instead of copies of entire conversations. Scope and retention would need a separate decision.

03 / Act

Sending needs its own approval.

Send the prepared reply to the customer.

A / Proposed boundary

An exact message and recipient

Sending would require approval for the exact message version and recipient. Without it, the action would stay a proposal.

Why it matters

Approval for an earlier draft should not authorize sending changed content.

B / Proposed decision record

What a record could include

  1. Approved version and recipient
  2. Approver and scope of approval
  3. Send outcome or reason for not proceeding
Less data. A clearer record.

References and versions may be enough instead of copies of entire conversations. Scope and retention would need a separate decision.

02 / Proposed direction

Two directions. A shared starting point.

Agent control tools

What should this agent be allowed to do, right now?

A possible tool layer for scoped permissions, action-bound approvals and explicit revocation. The starting point is a repeatable action contract that keeps a useful workflow within its authorized boundaries.

Explore the underlying research

Decision evidence tools

Could someone else reconstruct what happened?

A possible evidence layer connecting sources, system versions, permissions, approvals and outcomes. The aim is a reviewable record with a proportionate sensitive-data footprint.

Explore the underlying research
03 / Before it becomes a product

A product starts with a repeated need.

Consultancy reveals practical questions. Research tests whether an approach works. A product becomes worthwhile when that approach solves a recurring problem for a clearly identified user.

The next step is to test these directions against specific workflows and users before deciding what to build.

Explore the experience behind these questions

What we want to learn first.

We need to understand the workflow, the failure that matters and what a process owner would need to adopt a tool. Useful feedback can begin with a synthetic example.

  1. Which task recurs often enough to justify a dedicated tool?
  2. What evidence would demonstrate an improvement?
  3. Who owns access, operation and review?
  4. Can it be maintained within realistic cost and support limits?
Begin with a worthwhile question.

Put a real workflow on the table.

A recurring task, a difficult decision, a missing record. That is enough to begin a conversation.