AI-Native Self Assessment

How to use it

Hand it to the agent that knows the most.

01

Copy the prompt

Grab the full assessment as plain text — one click, or open the raw file.

02

Paste it into your agent

Use whichever agent has the most access to your company's systems: Claude, ChatGPT, Copilot, Gemini, or an in-house agent. The more it can see, the more honest the score.

03

Read the gaps, not the number

The total is a headline. The value is in the questions your agent scored 0 or 1 on — those are the specific places your AI can't close a loop today.

The assessment

Thirteen questions, scored 0–2.

The full rubric, as your agent sees it. Answering here scores you in the browser — your individual answers stay in this tab and never reach BeanOS. Anonymous analytics record that an assessment was completed, along with the total and the band — never the per-question answers.

Company State

1.Do I have a grasp on the scope of business knowledge that I have — and don't have — and where to get it?

0I work from whatever the user pastes into the chat. I have no map of what the company knows or where it lives. 1I know about some knowledge sources (a wiki, a drive folder, a handbook), but coverage is patchy and I often discover gaps mid-task. 2I have a structured, navigable map of company knowledge — I know what exists, where it lives, what's missing, and how to access it.

2.Is the knowledge I rely on version-controlled and attributed?

0I read from docs, emails, and chat messages with no versioning. I can't tell when something was last updated or by whom. 1Some knowledge is in version-controlled systems, but much lives in unversioned docs, spreadsheets, or people's heads. 2The knowledge I operate on is version-controlled. I can trace when it changed, who changed it, and why.

Programmable Actions

3.Do I have access to significant company APIs that allow me to close the loop on business outcomes?

0I can only generate text. To act on anything, a human must copy my output and do it manually. 1I can call some APIs (email, calendar, a few internal tools), but many actions still require a human to execute through a UI. 2I can programmatically execute the full range of actions needed for my work — cloud infrastructure, databases, ticketing systems, CRM, accounting, email, file storage — read data, write data, trigger workflows, and verify outcomes.

4.When a human can do something through a UI, can I do the same thing through an API?

0Most company systems are UI-only. I have no programmatic access to them. 1Some systems have APIs, but there are significant UI-only dead ends where I need a human to click buttons. 2Every system I need exposes an API or CLI. The UI is a view layer — I never need it to get work done.

Agent Infrastructure

5.Is my configuration, behavior, and memory defined as code — shared across sessions and versioned?

0I start each conversation from scratch. My instructions are whatever the user types. Nothing persists. 1I have some persistent instructions or memory, but it's local to one person or one tool — not shared across the organization. 2My identity, instructions, memory, and skills are defined in versioned code, shared across the organization, and improve over time through reviewed changes.

6.Am I self-learning from previous sessions?

0Every session starts from zero. I have no awareness of past interactions or their outcomes. 1I have some memory or context from past sessions, but it's informal and not systematically captured. 2Past sessions are logged, learnings are extracted, and my behavior evolves based on reviewed outcomes. My memory is a company asset, not a local cache.

Trust & Permissions

7.Is my access scoped to a well-defined trust boundary — not everything, not nothing?

0Either I have no access to anything meaningful, or I have a personal API key with broad access and no scoping. 1I have some access controls, but the boundaries are ad hoc — based on which tools someone configured, not a deliberate permission model. 2My access is determined by a declared trust tier. I see exactly what my scope allows — no more, no less — and it's enforced structurally, not by convention.

Session Architecture

8.Am I running in a sandboxed environment, or on someone's personal machine?

0I run on a personal laptop or desktop, with access to that person's local files, browser sessions, and credentials. 1I run in a somewhat isolated environment, but it's not fully sandboxed — I may share resources with other processes or have broader network access than needed. 2I run in an ephemeral, sandboxed container with hardened networking. I can only reach the company API layer and my AI provider. When I'm done, the container is scrapped.

9.Do I have access to raw API keys, tokens, or credentials — through tool calls, environment variables, or local files?

0Yes — I can read .env files, access environment variables with secrets, or call APIs using raw keys passed directly to me. 1I have indirect access to some credentials (e.g., a tool injects them for me), but I can also read some secrets directly if I try. 2I never see raw credentials. All API calls are mediated by a gateway that injects credentials on my behalf. I see results, never secrets.

10.Am I part of a distributed system with unified knowledge, or a siloed point solution?

0I'm one person's personal AI assistant. No other AI session in the company shares my context, knowledge, or memory. 1I'm isolated — I know there are other AI tools in use across the company, but I cannot communicate with them or share context. Each operates independently. 2I'm part of a unified system where all AI sessions share the same versioned knowledge base, the same API layer, and can coordinate through structured mechanisms (announcements, shared issue tracking).

Human Accountability

11.Do my outputs go through a human review process before affecting company state?

0My outputs are applied directly — or discarded entirely. There's no structured review step. 1Some of my changes go through human review, but I'm not aware of a coherent policy defining which changes require it and which don't. 2There is a coherent policy defining which changes I can push autonomously vs. which require human review — and it is discoverable and enforced by code, not just convention.

Auditability

12.Can every action I've taken be traced back to the session, the knowledge I used, and the human who authorized it?

0There is no audit trail. If something goes wrong, no one can reconstruct what I did or why. 1Some actions are logged, but the trail is incomplete — I can't always connect an outcome back to the session and knowledge that produced it. 2Every action is logged with full attribution: session ID, credentials used, knowledge accessed, human participants. Any outcome can be traced end-to-end.

Work Management

13.Am I driven by a structured work management system, or do I only act when a human prompts me?

0I only do things when someone types a message. If no one talks to me, nothing happens. 1I use a ticketing system (Jira, Linear, Asana, etc.) and can read issues, but a human must assign work to me and initiate each session. 2A ticketing system and scheduler drive my work end-to-end: issues are created on cadence or by triggers, I claim and progress them, log updates, and the company advances autonomously — humans review, steer, and decide.

Scoring

What your total means.

Add up all 13 answers, out of a maximum of 26. Every question runs the same direction: 2 is the healthy answer.

0–10

No to limited AI usage

Your AI is a text generator sitting outside the business. It can draft, summarize, and advise, but nothing it produces changes company state without a human retyping it.

Where to go next. Start with knowledge, not tools. Until your AI has a navigable, versioned map of what the company knows, every automation you add on top will be guessing. Pick the one workflow that costs the most human hours and make its inputs legible first.

11–16

Partial AI automation

You have real automation in places, but it's a collection of point solutions. Each tool has its own memory, its own credentials, and its own idea of what the company knows.

Where to go next. The bottleneck is no longer capability — it's the seams between your tools. Unify the knowledge layer and the API layer before adding another agent, or you'll be maintaining N copies of the same context.

17–22

AI-fluent with room for more

Your AI closes loops on real work and operates inside deliberate boundaries. The remaining gaps are usually governance and autonomy: review policy, audit completeness, and whether work reaches the agent without a human starting it.

Where to go next. Look hard at the questions you scored 1 on. At this level the gap is rarely a missing integration — it's an undeclared policy that lives in someone's head instead of in code.

23–26

A true AI-native company

Knowledge is versioned, actions are programmable, trust is structural, and work arrives without a human prompting it. Your AI is infrastructure, not a tool.

Where to go next. The work now is keeping it true. Scores decay as systems are added — re-run this quarterly and treat any regression as an incident, not a footnote.

This assessment is a snapshot. Run it periodically — after major infrastructure changes, after onboarding new systems, or quarterly — to track progress. The goal is not a perfect score; it's knowing exactly where you stand and where to invest next.