How do we get an AI agent approved in a large organisation?
An agent is approved when someone with authority signs a scope: owner, purpose, data, systems, budget and a review date. The path, and how to shorten it.
- Topic
- Pilots and delivery
- Product
- Novie Control
- Format
- Guide
- Reading time
- 4 min
An AI agent is approved in a large organisation when a person with the authority to accept its risk signs a written scope. The scope says who owns the agent, what it does, which data and systems it reaches, what it may spend, and when the approval is reviewed. Agents stall when nobody knows who that person is. Name the approver and the scope template first, and run the security, privacy and finance reviews in parallel against the same document.
What an approval is
An approval is a statement about a capability at a point in time. It says that this agent, using this model and these tools, may reach this data and take these actions, within this budget, owned by this person, until this date. When any of those change, the statement no longer describes the agent, and the approval should come back for review.
Most approvals fail that test. They are a signature on a demo, given once, with no scope written down. The agent then changes underneath the signature: the model is upgraded, a tool is added, the data it reaches grows. Nobody approved what it does today.
Why it matters now
Agents are arriving faster than the approval paths for them. Deloitte's State of AI in the Enterprise 2026 found 74 percent of organisations expect to be using AI agents by 2027 and 21 percent have a mature governance model for them. In KPMG's Q2 2026 AI Pulse Survey, 53 percent of large US enterprises had agents deployed, and 18 percent were orchestrating several agents across workflows.
The wait is measured. In AvePoint's survey of 750 IT leaders in financial services, healthcare and government, published in June 2026, 86 percent of organisations reported delays to AI agents averaging almost six months, mostly over unresolved data security and data management concerns.
Regulators expect the approval to mean something. APRA's letter to industry of 30 April 2026 asks for human involvement in high-risk decisions, controls over agentic and autonomous workflows, and an inventory of AI tooling and use cases. In New Zealand, the FMA's August 2026 review of AI in financial advice recorded record-keeping and audit trails as a significant barrier to adoption.
Who has to say yes
- The business owner, who answers for the outcome and owns the agent once it runs.
- Security, for the systems the agent can reach and the credentials it holds.
- Privacy, for any personal information in what the agent reads or writes.
- The data owner, for each class of data the agent uses.
- Finance, for the budget the agent draws on and its cap.
- Risk or compliance, in regulated work, for the decisions the agent informs or makes.
One of them has the final say. Write down who, by risk tier, before the first agent arrives. An agent that needs six approvals and has no final approver waits for the slowest of them.
The approval pack
- Purpose: the business outcome, in one sentence, and how it will be measured.
- Owner: the person accountable, by name.
- Model and tools: which model the agent runs on and which tools or APIs it can call.
- Data: the classes of data it can read and write, including personal information.
- Actions: what it can change in which systems, and which actions need a person to confirm.
- Budget: what it may spend per run and per month, and what happens at the cap.
- Record: what is logged, where, and for how long, so the question of what it did last month has an answer.
- Stop: how it is paused, and who can do it.
- Review: the date the approval expires, and the changes that bring it back early.
Why approvals take months
- The reviews run one after another, each waiting for the last.
- There is no template, so every reviewer asks for the same facts in a different form.
- The scope is open, so each reviewer imagines the worst case and asks for controls against it.
- Every agent is treated as new, including the tenth agent of a pattern already approved nine times.
How to make it weeks
- One intake, one pack. Every reviewer reads the same document.
- Parallel review, with a fixed window for each reviewer to object.
- Risk tiers. An agent that reads internal documents and drafts replies for a person to send is a different tier from one that changes customer records.
- Approved patterns. Once a pattern is approved, the next agent built on it inherits the controls and needs a lighter review.
- Approvals with expiry. A dated approval is easier to give than a permanent one.
When the approval should come back
The model behind the agent changes. A tool is added. It reaches new data. It runs over its cap. Its owner leaves. Each of these is a change to the statement the approver signed, and each should reopen it automatically, rather than waiting for someone to notice.
How Novie Control does it
Novie Control is live. Every dollar visible. Every budget governed. Every call optimised. Every agent governed. Every agent carries a named owner, an approval trail and a full audit trail, so the scope that was approved and what the agent did since sit in one place. Caps hold before spend is committed, with a named human on every increase.
Novie OS, launching soon, is where the next agent is described in your own words and approved before it is built, so the approval pack is written as part of building rather than assembled afterwards.
Sources
- Deloitte, Business and IT leaders report AI agents are scaling faster than their guardrails (24 April 2026)
- KPMG, AI Quarterly Pulse Survey Q2 2026 (24 June 2026)
- AvePoint, The State of AI 2026: What 750 Global IT Leaders Revealed About Trust, Control, and the Rise of Agents (29 June 2026)
- APRA, Letter to Industry on Artificial Intelligence (30 April 2026)
- Financial Markets Authority, Thematic review on artificial intelligence in financial advice: engagement overview (August 2026)