Skip to content

Solutions

Control AI that changes systems

Ticket automation, record updates and infrastructure actions. Give IT a single place to see what AI is changing and to constrain when it may change it.

The change surface

Systems reachable

11

Changing things unseen

3

The second number is the one that matters to IT. Those are systems that can change state and emit nothing anyone can read.

Why this page is not about the AI you run

IT's exposure is the AI everybody else runs against your systems.

Other functions govern their own agents. IT is accountable for the systems those agents change — and for the ones nobody registered, which arrived with a seat licence and have write access to a system of record.

In the sample estate, 11 registered AI systems attempt 10 kinds of action across 10 categories of system. Alongside them, 5 unregistered systems have been found by more than one detection source, and 3 of the systems that can change state are still unresolved.

That last number is the whole problem in one figure. It is not a list of things behaving badly — it is a list of things whose behaviour nobody can see, which is a different and worse position.

Categories AI currently reaches

  • Accounting
  • CRM
  • Database
  • Documents
  • ERP
  • Email
  • Payments
  • Payouts
  • Storage
  • Support

2 connected systems are currently reached by nothing: Slack, Workday. Worth seeing rather than assuming, because a system nothing reaches today is a system whose exposure is a decision rather than an accident.

Registered systems

11

under governance

Action kinds

10

across the estate

Found, corroborated

5

seen by two sources

Can act, unresolved

3

changing state, unobserved

IllustrativeIllustrative figures for the OpsAI sample estate. How discovery finds them.

The control IT already understands

A write window is change control expressed where it can be enforced.

Every IT organization already has an opinion about when production may be changed. The problem is that the opinion lives in a calendar, a runbook and a shared understanding — none of which an agent consults.

write.window

Writes to production happen inside the working window, when a person is awake to notice them.

Owned by Ops · v2 · in force since 8 Apr 2026

The useful detail is what happens outside the window: the action is not discarded and not silently retried. It is held until the window opens, which means a batch job running at the wrong hour becomes a delay rather than an incident.

Applies to the action, not the actor
So it covers every AI system reaching that target, including ones registered after the rule was written and ones nobody has registered yet.
Held, not dropped
Queued until the window opens, which is the one place in the product where waiting rather than refusing is correct — the request is fine, the timing is not.
Versioned like any other rule
So a change made during a freeze can be replayed against the window that was actually in force, rather than argued about.
What this replaces
A calendar invite, a Slack reminder, and the hope that whoever wrote the scheduler knew about the freeze. None of those produce a record.

What can be changed

One list of every system AI can reach, and exactly what it may do there.

This is the artefact IT usually cannot produce. Not because nobody wrote it down, but because it is spread across a dozen service accounts, some integration configuration and whatever an agent's prompt happens to say today.

Systems AI can reach

12 systems · 11 connected
Every system in the sample estate, its category, the actions AI may perform against it, and whether it is connected.
SystemCategoryWhat AI may changeStatus
RazorpayPaymentsissue.refund · capture.paymentConnected
RazorpayXPayoutsinitiate.payoutConnected
Zoho BooksAccountingpost.journal_entryConnected
NetSuiteERPcreate.vendor · update.vendorConnected
SalesforceCRMupdate.recordConnected
PostgresDatabaseadjust.stock · update.rowConnected
ZendeskSupportclose.ticketConnected
GmailEmailsend.emailConnected
SlackMessagingnotify · request_cosignConnected
Amazon S3Storageexport.datasetConnected
BoxDocumentsread.contractConnected
WorkdayHRNot connected

IllustrativeThe OpsAI sample estate, a fictional company. An unconnected system stays in the register rather than being omitted.

The database case is the one worth looking at twice

A write to an application is bounded by that application’s own logic. A write directly to a database is bounded by nothing except the grant, which is why the one action reaching a database here carry their own policies.

adjust.stockPostgres · inventory
Governed by write.window, evaluated before the statement reaches the database.
Why not just restrict the grant
A narrow grant is the right first step and it cannot express timing, amount or approval. A bound can, and it produces a record either way.

Granting less, reversibly

Autonomy is a rung, and breaking a rule moves it down automatically.

The usual choice offered to IT is binary: give an integration write access or do not. A ladder is more useful, because it makes granting less than everything a normal state rather than an admission that you do not trust the thing.

  1. 01

    Observe

    Reads. Never writes.

    Light, targeted checks

  2. 02

    Advise

    Proposes. A person commits.

    Output quality checks

  3. 03

    Act with approval

    Acts inside a bound. Above the bound, a person co-signs.

    Meaningful review · audit trail

  4. 04

    Act autonomously

    Acts inside a bound, watched continuously, circuit breaker armed.

    Circuit breaker armed

Where to start

Find out what already has write access, then put a window on it.

Discovery reads the systems you already run — the identity provider's OAuth grants, the egress proxy, the SaaS admin console. Nothing is installed on a laptop and no browsing history is inspected.