What controls do you need before automating approvals?

Why do controls have to come first?

Automating an approval does not make it safe. It makes it consistent. If the rules are good, every request follows them. If the rules are vague, the system either blocks work or waves it through, and nobody notices until year-end or an audit.

The same principle applies when AI enters the process. In February 2026 COSO published guidance on internal control over generative AI that builds on its widely used Internal Control: Integrated Framework (COSO). In the US, NIST's AI Agent Standards Initiative invited public input on AI agent security through a request for information with responses due on 9 March 2026, alongside a concept paper on agent identity and authorisation (NIST). Whether approvals are routed by a workflow tool or proposed by an AI agent, the questions are the same: who is allowed to decide, on what evidence, and how do you prove it afterwards?

What is an approval matrix?

An approval matrix, sometimes called a delegation of authority matrix, is a table that sets who may approve each type of transaction and up to what value. It is the core rule set an automated workflow will enforce. Here is an illustrative example; the roles and limits are placeholders to adapt, not recommendations.

Transaction	Up to 5,000	5,000 to 50,000	Above 50,000
Purchase order (budgeted)	Department head	Department head + finance manager	General manager + finance director
Purchase order (not budgeted)	Finance manager	General manager	Board or managing director
Supplier payment run	Prepared by accounts payable; released by two authorised signatories
New supplier or change of bank details	Verified by call-back to a known number; approved by finance manager, not the requester
Expense claim	Line manager	Line manager + finance manager	General manager

Write the matrix in business language first and agree it with management. Only then configure it in a system.

Which controls should be in place before you automate?

• Approval limits by role, not by person. Link limits to roles so the matrix survives staff changes.
• Segregation of duties. The person who raises a request cannot approve it, and the person who changes supplier details cannot release payments. See segregation of duties in digital workflows.
• No splitting. The system should flag several requests to the same supplier just under a limit.
• An audit trail. Every request, approval, rejection and edit is logged with who, when and what changed. Approvals given by email or chat outside the system do not count.
• Supporting evidence. Quotes, contracts or receipts are attached before a request can be approved.
• Exceptions with a path. Urgent or unusual requests have a defined route, with after-the-fact review, rather than a phone call that bypasses everything.
• Absence and delegation rules. When an approver is away, the request goes to a named delegate with equal or higher authority, and the delegation is time-limited and logged.
• Protected master data. Changes to supplier bank details, approval limits and user roles need their own approval.
• Regular review. Someone independent looks at overrides, exceptions and user rights at least quarterly.

What changes when AI is part of the approval?

An AI tool might check an invoice against an order, summarise a request or recommend approval. Treat its output as evidence, not as a decision. Keep a person approving anything above a low limit, give the AI its own restricted account, and log what it recommended next to what the human decided. Our guide on which AI agent actions need human approval covers this in detail.

Update, 3 April 2026: Deloitte's summary of the COSO guidance recommends applying a use-case-based framework to right-size the level of human involvement, "including review, approval, and segregation of duties", based on each use case's risk and how far AI outputs drive business decisions or automated processes (Deloitte, 3 April 2026).

How do you test that the controls work?

1. Try to approve your own request. The system should refuse.
2. Submit a request just above a limit. It should route to the next level.
3. Change a supplier's bank details. A second person should have to approve before any payment.
4. Mark an approver as absent. The request should go to the named delegate, and the log should show it.
5. Export the audit trail for one transaction. An outsider should be able to follow it without asking anyone.

Run these tests before go-live, and again after any change to roles or limits.

Update, 25 September 2026: the Mauritius Electronic Transactions (Amendment) Act 2026 came into operation. It adopts a reliability-based, technology-neutral standard for electronic signatures (Bill No. V of 2026, explanatory memorandum; Laws of Mauritius). An approval workflow that identifies the approver and records their intent is the kind of evidence this approach relies on.

Linkgrow's controls and cyber review checks these controls before approvals are automated, and the process redesign service builds the workflow around them. Start with the free gap finder.

Frequently asked questions

What is the difference between an approval matrix and a delegation of authority?

They are often the same thing. A delegation of authority sets who may commit the business and up to what value; an approval matrix is the table that applies it to each type of transaction.

Can a small team automate approvals without full segregation of duties?

Yes, with compensating controls: a second approval above a low limit, independent review of exceptions and overrides, and protected supplier bank details.

Are email approvals enough for an audit trail?

Rarely. Email approvals are hard to link to the transaction, easy to lose and easy to bypass. An approval recorded in the system against the request is far stronger evidence.

Can AI approve purchases automatically?

It can, but whether it should depends on the risk. A sensible default is to keep a person approving anything above a low limit and to treat the AI's check as evidence for that person.