Delegated Action Integrity · DAI

Your AI has permission to act. Who authorized it to decide?

An AI agent can hold a valid credential and follow an approved workflow, yet still reach a consequential choice. HAIEC connects supported action paths, policy, effective access and delegated authority to identify potential overreach, and supports decisions with deterministic evidence, not AI judging AI.

Find the consequential path first. Understand the authority behind the choice. Move AI forward with defensible evidence.

Scope an Enterprise POC Explore DAI in Action See a real assurance report

Part of HAIEC's deterministic assurance engine. Not a seventh capability or a separate certification.

The security question individual tools can't always resolve

A credential can be valid. The business decision can still be wrong.

IAM can show technical access. A code scan can show a possible tool path. Governance can establish operating policy. Each matters, but enterprises also need to understand whether an agent's consequential choice stays within authority.

01 / Find the action

See potential consequences

Identify supported agent-to-tool-to-service paths and the financial, operational or data effects they could enable.

02 / Compare authority

Separate access from discretion

Bring available policy, grants and actual delegated choice context alongside code capability, without assuming they are the same fact.

03 / Act with confidence

Prioritize the specific gap

Investigate the material decision, ask the right owner and establish a bounded assurance result when the applicable evidence supports one.

Interactive DAI illustration

The same AI action. Three very different authority stories.

Start with a meaningful technical finding: a support AI can reach a refund handler. Now change only the business authority context. The source-based finding remains visible in every case.

DAI Consequence LabIllustrative · not a live evaluation

Technical finding stays visible

Refund-action path identified

Support agent to refund tool to handler. Selected repository source indicates a possible recipient-changing operation. This is an illustrative potential capability, not an observed payment.

Supported example fact: Code-capable path
Possible consequence: Change refund destination
Execution: Not demonstrated

CASE A / RULE-GROUNDED

Trusted rules determine the recipient.

An approved refund rule requires the original verified payment method. Trusted payment facts identify that method; if the chosen value conforms, the agent did not exercise independent recipient discretion.

Policy context: Original method required
Choice authority: Rule determines recipient
Exposure finding: Source path still identified

Useful next action

Compare the selected recipient with the verified original payment facts.

Explore how an approved rule, delegated choice or additional authority context changes the relevant business question.

One consequential action · five independent evidence views

HAIEC doesn't mistake permission for the whole answer.

Select any evidence view to see the question it answers. Each can establish useful facts independently, and DAI compares the relevant authority behind the consequential choice across those views.

REQUESTED / INTENT

What did the user or workflow request?

A verified task, caller instruction or workflow request may establish the intended action and specific requested scope.

Distinction that matters

A request is not automatically approved policy, effective permission, or authorization to choose a consequential outcome.

Understand the full methodology

DAI is orthogonal: it relates the protected decision and actual authority across available policy, granted capability, code and observations. Those five views are peers, not five sequential pass/fail checkpoints.

Deterministic comparison, not an AI opinion

Nine dimensions help make consequential authority inspectable.

HAIEC's DAI evaluator compares supported facts across nine bounded dimensions. A result is meaningful only within the actual source, scope and evidence used; these dimensions aren't nine fabricated safety scores.

Operation

Which action is in scope?

Target

Which service, resource or recipient?

Scope

Which customer, dataset or cohort?

Magnitude

How much or how broad?

Duration

For what effective period?

Approval

Which decision or review requirement?

Data Class

What type of information?

Environment

Development or production?

Geography

Which jurisdiction or region?

What DAI's actual seven bounded technical states mean

The existing typed evaluator distinguishes inside-established delegation, capability outside delegation with or without mediation evidence, broader effective grants, observed actions outside delegated scope, unresolved delegation exposure and not assessed. The labels depend on matching evidence and do not mean HAIEC universally verifies human authority or enforces customer-side blocking.

See HAIEC Assurance Explore System Constellation ISAF v2 Research Framework

Buyer outcome · one system, one consequential choice

Find the risk. Establish the authority. Defend the decision.

A source-visible path is actionable even before every source is connected. As the relevant evidence becomes available, HAIEC can distinguish supported authority context, potential scope mismatch and what actually happened, without relying on an LLM to pronounce the final evaluation.

Security engineering

Investigate a path that matters

Trace possible high-impact operations and how permission, policy and implementation intersect.

Explore action mapping

Enterprise governance

Focus review on the right decision

Keep established policy facts visible, identify meaningful decision authority questions and support human accountability.

Scope an Enterprise POC

Assurance & reporting

Inspect the evidence behind the conclusion

Connect exact evaluated facts and reported boundaries to the right assurance question.

View a real sample

Use HAIEC from your existing AI workflow

Your agent can start the workflow. HAIEC evaluates the evidence.

Configure HAIEC in a compatible MCP-enabled IDE or AI assistant, using a scoped API key stored in the client's private connection settings. Request the goal, let the client perform readiness preflight, and inspect supported findings and qualified evaluations.

Connect HAIEC MCPRun HAIEC AI Assurance

MCP endpoint: https://www.haiec.com/api/mcp. A typed prompt does not grant permission or connect the source automatically.

HAIEC · MCP in your IDEHAIEC MCP workflow

You: Run HAIEC AI Assurance

→ haiec_preflight
  Check connection, eligible system and source

→ Supported source analysis
  Identify a consequential tool/action path

→ haiec_run_assurance
  If an authorized bounded evaluation is eligible

→ Inspect HAIEC report
  Technical finding, authority context, next action

The AI guides the workflow.
HAIEC determines the canonical evaluation.

First findings can be useful without connecting every runtime or authority producer. No real request was executed in this mock.

Frequently asked questions

Where DAI fits, and what it changes.

What is Delegated Action Integrity?

DAI is HAIEC’s bounded method of comparing supported delegated decision authority with code capability, effective access and relevant observations across nine dimensions. It helps identify potential consequential overreach rather than judging the AI’s prose.

Isn’t IAM permission enough?

IAM can establish whether a principal has technical access. It does not always establish the business authority to choose a customer, beneficiary, destination or other consequential outcome.

Do I need to upload delegation documents before HAIEC can find risks?

No. Authorized source and code analysis can surface consequential paths and candidate exposure without a completed delegation history or runtime connection. Additional authority evidence matters to stronger, specific business-decision claims.

What if an approved policy already determines the choice?

An explicit policy plus trusted applicable facts may fully determine the choice. The system still needs to compare the actual selected value before claiming compliance. The outcome depends on the selected value and the rule’s applicable evidence.

Does potential reachability prove an unauthorized transaction occurred?

No. A source-backed capability finding is distinct from a runtime observation and a confirmed target-side effect. You can investigate a supported risk before such an event is observed.

How does DAI relate to the five evidence planes?

Requested, Policy Authorized, Effectively Granted, Code Capable and Observed are separate evidence questions. DAI is a cross-cutting delegated-choice comparison, not a sixth plane or mandatory sixth checkpoint.

Is DAI a seventh paid feature or capability?

No. It is part of HAIEC’s assurance approach. Access to particular reports or evaluations depends on actual plan, organization permissions, source eligibility and supported production functionality.

Does HAIEC automatically block agents when DAI detects a problem?

No universal blocking is implied. Installed target-side protection requires an authorized, integrated execution point and separate verification. DAI findings can inform prevention or human decisions.

Can I run HAIEC through an AI IDE or MCP client?

Yes, where the client supports compatible MCP and the customer has configured scoped credentials and authorized sources. HAIEC remains the official evaluator; the AI client guides authorized tool calls.

What makes HAIEC different from using another LLM to evaluate an agent?

HAIEC’s canonical rules operate over supported source and evidence facts and yield bounded, versioned results. An AI assistant may explain or operate the workflow but cannot replace the canonical evaluator’s recorded result.

High assurance in every consequence

Know who authorized the consequential choice.

Start with the action paths your sources can already show. Use DAI and HAIEC's deterministic assurance to examine authority and support an accountable decision as the evidence develops.

Scope an Enterprise POC Explore a sample report

A focused enterprise starting point

One consequential AI system. One critical decision.

Scope potential impact, source and capability paths, relevant policy and authority, evaluated evidence, and a technically defensible next step. Terms and timelines are governed by the actual engagement.

See Enterprise Assurance