Skip to content
Audit & compliance

Compliance
enables AI.

Operationalising AI immediately introduces complex regulatory demands. Compliance managers have to satisfy the EU AI Act, GDPR and works councils simultaneously, using legacy tracking software designed strictly for human activity. Traditional platforms fail because they cannot prove what an autonomous agent actually executed.

Devctrl Console
console.devctrl.ai/policies

New policy

Describe it in a sentence, or build it in the editor.

Save policy
Customer data must never go to an external model|

Devctrl turns that into this rule:

  • IfContent contains customer data or account numbers
  • AndThe destination is a model outside your environment
  • ThenBlock it and note it in the record

Applies to

Every team, as the organisation's floor

Watch first

Runs alongside for two weeks before it bites

Ask Devctrl⌘J

This is how a policy comes about: one sentence from the person who owns the rule. No compulsory editor, no ticket to IT.

The problem

AI agents do not behave the way people do.

Autonomous AI agents operate at speeds and volumes that human audit frameworks were never designed to manage, yet ultimate responsibility stays with the compliance department. Auditing an agent that can reach a single file is easy. Granting broad folder permissions opens up serious risk of unintended data processing. Devctrl replaces static folder permissions with granular, task-level governance, giving enterprise teams full operational oversight without holding the workforce back.

  • 01

    The risk of overprivileged agents

    Static role assignments grant agents persistent, broad access far beyond the need of a single task. One compromised credential or one malicious prompt injection can exploit those excessive rights, exposing every system and file within reach.

    An agent fetches one number from accounting

    • It can reachevery system the role has
    • It would needone of them

    Within Devctrl

    Access is limited strictly to pre-approved system connections, and it runs under each person's own account. Corporate security policies add further restrictions on top, so an agent can never exceed the organisation's own rules.

  • 02

    Beyond generic permission checks

    Traditional permission models evaluate raw action types while remaining entirely blind to operational context. Without deep inspection of the request content, a static gateway grants access to irrelevant or unsafe operations simply because it can.

    An email going outside, with the customer list attached

    • The classic permission check asks

      May it send email?

      Yes
    • Devctrl asks

      May this email go out?

      No

    Within Devctrl

    The action itself is checked, with everything in it: the tool, the target system, the contents, and whether the data leaves the company.

  • 03

    Policy drift

    Manually maintained security policies create dangerous governance gaps as an enterprise grows. Independent team exceptions obscure administrative oversight, turning unmonitored policy drift into severe regulatory exposure.

    One question, three answers in the same company

    • SalesExternal mail needs approval
    • FinanceAllowed, exception since March
    • SupportNothing on file

    Nobody sees the whole picture

    Within Devctrl

    The organisation sets the floor, each team inherits it and narrows further. Where two rules disagree the stricter one wins, and every change stands in the dashboard with a name and a time.

Audit levels

Four flexible audit tiers.

Retention schedules operate independently from logging tiers. Enterprise teams can enforce short activity retention windows for general operations while keeping permanent compliance archives for high-risk business units.

  • Everything

    Recorded

    Full capture of prompts, model output, tool inputs and tool outputs

  • Inputs only

    Recorded

    Capture prompts and tool inputs without storing tool outputs

  • On policy events

    Activity logged

    Buffer content during the run and retain it only when a policy rule triggers

  • Activity only

    Activity logged

    Track user identities, timestamps and tool selections without saving payload content

Flexible administrative settings deliver complete visibility across every workflow. The badge in the app explicitly signals “Recorded” whenever content is saved and “Activity logged” when tracking is limited to activity, so the status is always clear, with the full detail on hover.

The record

Verifiable audit ledger.

Every execution step and every published output flows through a single processing pipeline into an append-only audit ledger. Standardised event schemas across all operations make it straightforward to filter, quick to search and immediate to verify.

Separation of operational and audit stores
Operational messaging data stays entirely separate from the append-only audit record. Keeping less inside active team workflows reduces operational exposure without compromising core application functionality or end-user safety.
Strict role-based access control
Comprehensive logs are highly sensitive security assets and need explicit protection. Access is restricted to designated audit and security roles, with auditor access configured as read-only by design to protect data integrity.
Audited log access
Every request to view recorded content becomes an immutable part of the ledger itself. Viewer identities, timestamps and session details are logged automatically, so administrative teams are accountable too.
Pre-execution logging integrity
Audit entries are written before an action executes, not after it completes. If the write fails, the operation halts immediately, so no unrecorded action can take place.

One run, in order

  1. Anika Schwarz started a task

    Prep the Q3 churn deck · Sales · full capture

  2. Model callAllowed

    Planned the task · 4 steps · attributed to Anika

  3. Retrieved from NotionAllowed

    As Anika, under her own workspace access · 6 pages

  4. Screened tool outputRedacted

    2 customer names matched the PII detector

  5. Attempted an external sendBlocked

    Sales may not email outside the org unattended

  6. Paused for approvalNeeds approval

    Routed to Anika · she allowed it

  7. Published artifactAllowed

    Q3-churn.pdf · post-screened, shared with Revenue

Overview

Complete operational visibility.

Governance dashboards answer the questions an audit really asks about departmental policy variations. The grid maps capture tiers, retention windows and policy triggers by team, while the change history tracks every policy modification with a name and a timestamp.

ScopeCaptureRetentionChains · 30dWith contentPolicy events
Organisation defaultInputs only180 days1,0461,0469
FinanceEverythingKept indefinitely1,2841,28431
SalesActivity only90 days3,902064
EngineeringInputs only365 days2,1402,14012
PeopleOn policy events180 days41277

Recent changes

  • M. WeberSales · inputs only → metadata only12 Jun
  • D. OkaforPeople · retention 90 → 180 days3 May
  • M. WeberLegal · retention set to indefinite21 Apr

Integrity

A record no call gets past.

Every audit entry is written before an action happens, not after. A tool call only goes out once its record is safely stored. If that first write fails, the whole action is cancelled immediately.

Every action is captured
Every tool call passes through a single gateway that creates a permanent entry before anything takes effect. An agent has no way around it.
Records are final
Once an entry enters the chain it stays untouched. Search and dashboard views read that history without changing the original.
Incomplete runs stay visible
If a run stops midway, nothing erases it. The console marks it as never finalised, so a task that only half ran stays visible.

One call, from the inside

  1. Entry written

    Who, with which tool, on what. Durable, before anything takes effect.

  2. Call to Notion

    The one step that cannot be taken back.

  3. Entry written

    Result, effect, timing. The chain is complete from here.

If step 1 fails, nothing is called. What was not recorded did not happen.

Ready?

Take control of your AI.

We walk you through a real use case, one that is realistic for you. Including what happens when a policy denies something.