EzraGroup

AI tool classification & evaluation

Describe one capability as it works today.

Your answers give Ezra Group enough evidence to classify one specific AI capability by system type, then evaluate its autonomy, power level, governance, orchestration and verification status. You do not need to decide whether your product is an AI agent — Ezra makes that assessment from the behavior you describe.

  • Answer for one capability as deployed by default today, not a roadmap capability.
  • If your product has multiple capabilities that behave differently, submit the form once per capability.
  • Describe what the system actually does, not the marketing label you use for it.
  • Anything sensitive can be answered in a follow-up under NDA.

Submitter

Section A

Context — what this capability does

Answer for one capability as it is deployed by default today, not a roadmap capability. Be concrete: describe what the system actually does, not the marketing label.

In one or two sentences, what is this capability supposed to do, and who uses it?

Example: “Helps advisors prepare for client review meetings by analyzing CRM notes, emails, and portfolio data.”

Describe what happens from start to finish when this capability is used.

Include what starts the work, the major steps, and what output or action ends it.

What specific outcome does the capability deliver?

Example: “Creates a meeting agenda and proposed follow-up tasks for advisor review.”

What normally starts the capability?

Which wealth-management area(s) does it operate in?

Select all that apply.

Which systems or data sources does this capability read from or write to?

CRMs, custodians, portfolio systems, email/calendar, document repositories, client portals, trading systems, etc.

Section B

System type

These questions let Ezra classify the capability as Automation, AI Assistant, AI Workflow, or AI Agent. Please answer behaviorally; do not self-classify.

Is an AI/LLM model used at runtime to interpret, generate, classify, reason, or choose based on the current input?

Is the capability mainly responding to a human request, with no multi-step process to manage?

During a multi-step run, who determines what happens next?

Can the AI change its planned sequence mid-task because of new information, without a human re-prompting it?

If yes, give one real example of what it planned, what changed, and what it did differently.

How does the capability know the job is finished?

Give one concrete example where the same capability took a different path for two different cases because of what the AI discovered.

If no such example exists, write “No dynamic path change.”

Section C

Autonomy

Autonomy is scored separately from system type. Describe how much human participation is required in normal operation.

As deployed by default today, what role does the human normally play?

Before the capability's output or action takes effect, what human review is required by default?

After the work starts, can the capability complete a normal case successfully without a person participating again?

Section D

Power level

Power level asks what the capability is actually allowed to do or change. Higher is not better; it means broader or more consequential authority.

What is the highest-impact action this capability is permitted to take?

What permissions does it hold in each connected system, and where are those limits enforced?

Example: read-only CRM access, write access to named fields, email-send permission, trading API scope.

Section E

Governance

Governance measures the controls around the capability. Ezra evaluates these controls separately from the autonomy operating mode.

Escalation — does it pause or flag rather than proceed when confidence is low or data looks wrong?

What triggers the escalation?

Optional

Audit trail — is every action logged and traceable after the fact?

Can the firm access/export the log? What fields are captured?

Optional

Permission scope — is there a bounded set of actions it cannot exceed, with least-privilege access?

How is the boundary enforced?

Optional

Rollback / reversibility — can actions be undone after they happen?

Describe the mechanism and time window.

Optional

Approval capability — can a firm require human sign-off before consequential actions?

How granular is the configuration?

Optional

Section F

Orchestration

A listing attribute describing architecture. It does not by itself make a system more autonomous or more agentic.

Is this capability one AI component, or several coordinated components?

If multiple components are used, what does each one do and who/what decides which component runs next?

Optional

Note whether routing is hard-coded, classifier-driven, or chosen dynamically by an LLM.

Section G

Evidence & verification

Written answers establish a Claimed rating. Demos, traces, and production evidence can move a capability to Demoed or Verified.

Can you provide run-level evidence showing the capability's behavior?

Select all that apply.

Can you provide two examples of the same capability handling materially different cases, including the actions/tools it chose and the outcome?

Technical contact for follow-up verification

Name, title, email.