Does using Pushary make our AI system compliant?

No, and any vendor telling you otherwise is selling you something. Compliance depends on your classification, your risk management, your documentation, your conformity assessment, and how you actually operate the system. Pushary supplies a mechanism for the intervention and the event record. The obligation stays with you.

Is a Slack message enough to satisfy an oversight requirement?

That is a question for your counsel, and the practical problem is easier to describe. A chat message is delivered whether or not the right person reads it, there is no return value the system waits on, and nothing about it fails closed. If the point of oversight is that a person can effectively intervene, an oversight step nobody received is difficult to evidence afterwards.

What is stored, and where?

The default stored asset is the decision and policy event: the question text you composed, the answer, the person, and the timestamp. It is not a transcript of your system's inner workings and it is not a copy of your records. Keep sensitive content in your own environment and pass a reference and a short summary in the question.

Can an approver override the system rather than only approving it?

Yes, and for an oversight step it is usually the more useful shape. Use the select type with an override option and the person's choice comes back as the answer, recorded as an override in the ledger rather than as a rejection. Article 14 speaks about deciding not to use the output, which is exactly that distinction.

Does this work for a product sold into the EU from outside it?

The mechanism does, because it is addressed by your own user id and reaches whatever device that person carries. Whether the regulation applies to you as a provider or deployer placing a system on the EU market is a legal question, and it is one worth answering early rather than late.

EU AI Act

Human oversight, as a thing you can actually ship.

The AI Act asks that a person can oversee a high-risk system and that what happened is recorded. Most of that is a product question about reaching the right person and keeping the evidence. This page separates the parts you can build from the parts you need counsel for.

What does the EU AI Act require for human oversight?

Article 14 of Regulation (EU) 2024/1689 requires high-risk AI systems to be designed so natural persons can effectively oversee them, including the ability to intervene or interrupt. Article 12 requires automatic logging of events over the system's lifetime. Pushary covers the operational half: it reaches a named person, blocks the agent until they decide, fails closed if they do not, and writes each decision to a durable ledger.

Read the regulation itself: Regulation (EU) 2024/1689 on EUR-Lex.

Article by article

What the text asks for, and which part is buildable

Article 14

Human oversight

High-risk systems must be designed and developed so that natural persons can effectively oversee them during use, including understanding the system's capacity and limits, staying alert to automation bias, correctly interpreting output, and being able to decide not to use the output or to intervene and interrupt.

Interrupting is a call your agent makes and waits on. Pushary blocks the run, puts one narrowed question in front of a named person on the device they carry, and returns their choice. Overriding output is a first-class answer rather than an exception path.

Article 12

Record-keeping

High-risk systems must technically allow for the automatic recording of events over the lifetime of the system, to a degree appropriate to the intended purpose.

Each decision is written to a durable ledger holding the question as the person saw it, the option they chose, their identity, and the timestamp. It is queryable and exportable, and resolved decisions post a signed webhook into your own evidence store.

Article 26

Deployer obligations

Deployers of high-risk systems must assign human oversight to natural persons who have the necessary competence, training, and authority, and must ensure they have the support to carry it out.

Oversight is addressed to a specific person by the id you already hold for them, not to a shared channel. Two approvers on the same action are two independent calls with two separately attributable ledger rows.

Annex III

High-risk use cases

The annex lists the use cases treated as high-risk, including areas such as employment, essential private and public services, creditworthiness, and risk assessment and pricing in life and health insurance.

Whether your product is in scope is a legal question about your use case, not a technical one, and it is the question to settle before anything on this page is relevant to you.

What this page is not

This is not legal advice, and Pushary does not make anyone compliant. Human oversight is a property of how your system is designed and operated. A tool can supply part of the mechanism and cannot supply the obligation.
Dates in the AI Act timetable have been the subject of active political discussion, including proposals affecting when parts of the high-risk regime apply. Check the current position on EUR-Lex or with your counsel rather than relying on a date on a vendor page. We deliberately do not print a countdown.
Pushary is GDPR-aligned and self-assessed. It holds no ISO certification, no SOC 2 report, and no AI Act conformity assessment, and we will not claim any of those.
Whether your product falls inside Annex III at all is the first question, and it is one for your own counsel. Nothing here should be read as an opinion on your classification.

The oversight step in code

Override is an answer rather than an escape hatch, and a person who does not respond produces a denial rather than a default action.

import { createPusharyServer } from "@pushary/server"

const px = createPusharyServer({ apiKey: process.env.PUSHARY_API_KEY! })

// The oversight step: a named natural person, not a role or a queue.
const decision = await px.decisions.ask({
  externalId: reviewer.id,
  type: "select",
  question: `Case ${caseRef}: system proposes ${proposal}. ${topReason}`,
  options: ["Accept", "Override", "Refer upward"],
})

// Nothing proceeds unless a person chose it. A timeout returns denied.
if (decision.answer === "Override") return recordOverride(caseRef, decision)
if (!decision.approved) return holdForReview(caseRef)

// The ledger row is the event record: who saw what, what they chose, when.
return applyProposal(caseRef, proposal)

Questions, answered

Read on

Auf Deutsch lesen Embed it for your users Audit trail Human-in-the-loop API Insurance agents

Build the oversight step, keep the record.

One call reaches a named person and waits for them. Nothing proceeds on silence, and every decision is written down. What that means for your obligations is a conversation with your counsel, and this is the part you can ship.