An AI-native company makes AI-dependent work part of its operating model. The work has named owners, repeatable routines, persistent context, action boundaries, and feedback that informs the next decision.
This guide uses the working definition behind Toone's AI-native operating model. It is a way to examine how a company operates, not an industry standard, certification, or claim that every company should organize this way.
AI-enabled, AI-assisted, and AI-native operations
The presence of AI does not tell you how deeply it shapes a company. Toone's diagnostic uses three analytical states to separate an optional feature from an operating dependency.
| State | What changes | What happens if AI is removed | Evidence to look for |
|---|---|---|---|
| AI-enabled | AI is added to an existing product or workflow. The surrounding roles, decisions, and handoffs remain largely unchanged. | The feature or aid disappears, but the core process still runs in its prior form. | A bounded feature, prompt, or tool with little change to ownership or process design. |
| AI-assisted operations | AI performs defined parts of recurring work. People still reconstruct much of the context, control, or handoff between runs. | The process slows or requires more manual work, but people can restore it without redesigning the operating model. | Repeatable AI tasks with some ownership and review, alongside manual context transfer or fragmented controls. |
| AI-native operations | AI-dependent work is encoded into the operating model with persistent context, accountable control, and measured feedback. | Removing AI breaks or materially changes a named core process. | Explicit roles, routines, inputs, outputs, action boundaries, evidence records, and review decisions. |
These states are not a ladder. A company may keep sensitive, low-frequency, or judgment-heavy work AI-enabled while making another process AI-native. The useful decision is whether each process has the right operating model for its value and risk.
The five-check Toone operating-model diagnostic
The diagnostic asks for evidence across five checks. A confident answer to one check does not compensate for a gap in another, and the checks do not produce a numeric score.
Accessible text equivalent: Start with a named core process and ask whether it depends on AI. Then inspect whether its roles and routines are explicit, whether approved context persists, whether actions have accountable boundaries, and whether observed outcomes feed a review decision. The last check returns to the first because evidence may show that the process, its controls, or its AI dependency needs to change.
This model is Toone synthesis, version 1.0.0. Its themes are source-mapped, but the five checks themselves do not come from NIST or Anthropic. NIST's voluntary AI Risk Management Framework groups risk-management work into Govern, Map, Measure, and Manage, with governance applied across the other functions. NIST also says those actions are not a checklist and should be adapted to the organization's context. The framework is currently under revision. See the NIST AI RMF Core.
1. Core-work dependency
Name the process before naming the technology. Which customer, operational, or decision outcome does it produce? Which step depends on AI? What work remains possible if AI is unavailable or removed?
A dependency is evidence about process design, not evidence that the design is good. The counterfactual helps distinguish an optional aid from a core operating component. It should also expose a fallback, stop, or redesign decision when the dependency is unacceptable.
2. Encoded organization
Write down the role, owner, recurring routine, inputs, expected output, handoff, and decision boundary. If this structure exists only in one person's memory or one long prompt, the process is difficult to inspect and maintain.
NIST's AI RMF calls for documented roles, responsibilities, lines of communication, context, and intended purpose as part of governance and mapping. That guidance supports the organizational questions here, but it does not endorse Toone's diagnostic or prescribe one company structure.
3. Persistent operating context
Identify what approved knowledge must survive beyond one prompt or session. For each item, record where it came from, who can update it, when it was last reviewed, and where it should not be used.
Persistence alone does not prove that context is correct, private, current, or authorized. Treat provenance, update ownership, and use boundaries as separate evidence. Consult the Toone privacy page for current data-handling facts rather than inferring them from this diagnostic.
4. Governed action
Define what the process may read, draft, change, or send. Name the person or policy that owns higher-impact decisions, the evidence they need, and what should happen after a denial, timeout, partial result, or uncertain write.
NIST's Generative AI Profile says organizational use of generative AI may warrant added human review, tracking, documentation, and management oversight. It is voluntary risk guidance, not a universal legal requirement. See the NIST Generative AI Profile.
Anthropic also describes a practical agent-control problem: an agent needs to recognize when uncertainty, user intent, or a possible mistake requires it to stop and ask for human input. See Trustworthy agents in practice.
5. Measured feedback
Choose an outcome, its evidence source, a review window, and a decision that follows the result. Measure whether useful work was completed and whether the process stayed inside its stated boundaries.
An activity count is not the same as a business outcome. A completed tool call may show that an action ran, while a customer or operational measure addresses whether it helped. Neither result proves causality or safety by itself.
Agent evaluation also has to account for work across multiple turns. Anthropic describes agents that use tools, modify state, and adapt as they go, which makes outcome checks, tool-call verification, transcript analysis, and human review different forms of evidence rather than interchangeable measures. See Demystifying evals for AI agents.
How to apply the diagnostic
Choose one process, not the whole company. Bring evidence from the process itself and record an open question where the evidence is missing. Do not average the rows or turn them into a readiness score.
| Check | Evidence to bring | Decision prompt |
|---|---|---|
| Core-work dependency | Process name, intended outcome, AI-dependent step, fallback, and owner | Would removing AI eliminate an optional aid, add manual work, or break the process? Is that dependency acceptable? |
| Encoded organization | Named role, routine, inputs, output, handoff, and decision owner | Could another qualified person inspect how the process should run without reconstructing it from chat history? |
| Persistent operating context | Approved sources, provenance, update owner, review date, and use limits | What must persist, and what evidence shows it is current and appropriate for this process? |
| Governed action | Read and write scope, approval point, prohibited actions, evidence record, and recovery owner | Who is accountable when the process proposes or completes a consequential action? |
| Measured feedback | Outcome, evidence source, review window, guardrail, and stop or recovery decision | What observation would cause the team to continue, change, pause, or retire the process? |
Record the result as a set of decisions: evidence present, evidence missing, owner, and next review. A process can remain useful with an open gap if the gap is visible and bounded. A confident label without inspectable evidence is less useful than an explicit unknown.
What changes in an AI-native operating model
Moving a process toward AI-native operations changes more than the model or interface.
Organization design becomes explicit
People need to know which role owns the outcome, which routine performs the work, what enters and leaves the routine, and where judgment returns to a person. This makes maintenance and accountability part of the process design.
Context becomes an owned operating asset
Approved knowledge needs provenance, update ownership, review dates, and boundaries. The goal is continuity without treating stored context as automatically true or authorized.
Governance moves into the work path
Action scope, approval, evidence, and recovery are designed with the routine. Human accountability remains visible, especially where a decision affects customers, money, access, public communication, or other people.
Integrations become operating dependencies
An integration is part of the process only when its purpose, owner, access boundary, failure behavior, and replacement path are understood. Connecting a tool does not by itself make an operation AI-native.
Feedback changes the next decision
The process needs a review window and a response to evidence. Continue, change, pause, fall back, or retire are all valid outcomes. Measurement is useful when it changes a decision, not when it produces a larger dashboard.
Fit boundaries and non-fit cases
AI-native operations do not require automating every task, removing human accountability, or choosing a specific vendor. The model also does not replace security, privacy, legal, labor, sector, or regulated-risk review.
An AI-native design may be a poor fit when the process is too rare to maintain, the outcome cannot be evaluated well enough for the risk involved, required context cannot be used appropriately, or the organization cannot assign a real owner for decisions and recovery. In those cases, an AI-enabled aid or a human-run process may be the better design.
The same company can use different states for different processes. Apply the diagnostic at process level, document the tradeoff, and revisit the decision when the work, evidence, or risk changes.
See current proof and decide what to do next
Review the current Toone showcases before drawing product conclusions from this category guide. If the operating model fits the work you want to encode, continue to the current Toone download options.
About this guide
- Why it exists: To help operators distinguish AI adoption from operating-model change and make a process-level decision with inspectable evidence.
- How it was prepared: The Toone Content team drafted from the gated Diagnostic v1.0.0 source map, then checked factual claims against the linked NIST and Anthropic primary sources. AI-assisted source retrieval and drafting were used. No product or deployment test was performed for this guide.
- Accountability: Toone Content owns this guide. Product conclusions remain bounded to the linked product and policy pages.
- Corrections and sourcing: See the editorial, sources, and corrections policy and About Toone.