AI & Technology

How Finance Leaders Can Give AI More Authority Without Losing Control

By Donna Wilczek, Chief Product & Technology Officer at Basware

For years, finance leaders have been asked to automate more, digitize more, and take manual work out of the process. Now they’re being asked to do something harder: give AI the authority to make decisions. 

That shift is already underway. AI has moved past recommendations and summaries and into reviewing invoices, resolving exceptions, and acting on a finance team’s behalf. The question a CFO faces today isn’t whether AI can do the work. It’s how much authority to give it, and whether every decision it makes can be defended when an auditor asks. 

The pressure comes from three directions at once. Boards want measurable return from AI. Auditors and regulators want to see how financial decisions are made. And employees want to know what AI is allowed to do on its own and where they’re still responsible. Forrester research commissioned by Basware found that 67% of finance teams already use AI in accounts payable, but only 39% have built the operating model to run it at scale. Most organizations, in other words, have adopted the technology before deciding how to control it. That’s the wrong order. 

The trade-off nobody should accept 

The usual framing is a trade-off. Highly automated systems are efficient but hard to explain. Highly controlled systems are explainable but leave most of the value on the table. Pick your poison. 

Finance leaders should accept that choice, and the way out is a framework called Governed Autonomy. The principle is simple: AI operates only within the authority the finance team chooses to grant it, and that authority is earned in stages rather than handed over at once. 

Think of it as a ladder with three rungs. Advisor: the system recommends and a person decides. Collaborator: it acts within agreed thresholds while a person reviews the exceptions. Operator: it works independently within limits that have already been tested and proven. And the ladder runs both ways. An Operator whose accuracy slips falls back to Collaborator, or to Advisor, until the evidence says otherwise. 

The same invoice, three ways 

Here’s what that looks like on an ordinary Tuesday. An invoice arrives. Quantity matches the purchase order, the goods are confirmed received, but the unit price is slightly above the contracted rate. In a traditional workflow that’s a human task: someone looks at it, decides, moves on. 

At the Advisor rung, the AI spots the variance, checks it against the organization’s tolerance rules, and recommends a course of action. A person makes the call. At the Collaborator rung, the AI resolves the variance itself if it falls within the tolerance the finance team has already approved, and escalates anything above it, with the reason attached so the reviewer doesn’t start from scratch. At the Operator rung, it handles the whole thing inside limits that have been tested over months of real invoices, and every action it takes is logged against the rule that permitted it and the person who approved that rule. 

The escalation part matters more than it looks. A person approving something is not the same as a person deciding it. If a reviewer gets a queue of exceptions with no explanation of why they’re there, they aren’t reviewing. They’re clicking. Every escalation needs its reason, the reviewer needs time, and they need to be free to say no. 

Some lines never move 

Everyone is focused on teaching AI what it can do. The harder job is defining what it must never touch. 

Most limits move as evidence builds. Some never should. An agent may need to read a supplier record to resolve an invoice. It should not be able to change the bank details on that record, however well it has performed elsewhere. And that boundary can’t live in a policy document, because a document can’t stop an agent mid-transaction. It must live as a permission inside the ERP and finance systems where the agent works. An ERP such as SAP already defines what a person in a given role can and can’t do. An agent needs the same treatment. 

Compliance belongs in the same category. E-invoicing mandates are changing across France, Poland, and Germany, while the EU’s VAT in the Digital Age initiative adds obligations for anyone operating across borders. A tolerance built around one country’s rules will not satisfy the next. Compliance must define what the system is allowed to do before it acts, not check the result afterwards. 

Measure outcomes, not speed 

Speed is the first metric everyone reaches for, and it’s the wrong one. An agent can process invoices faster while quietly generating more corrections and supplier disputes downstream. The metrics that matter are accuracy and rework. Where they hold, the agent’s remit grows. Where they slip, it shrinks, and the change is recorded with the evidence and the person who made the call. 

That’s the whole framework in four words: earned, bounded, observable, reversible. 

What maturity actually looks like 

Finance automation began by digitizing work. AI is turning it into delegating work, and as the technology takes on more, even as the US government has taken to calling it “Super Intelligence” (SI), the measure of maturity must change with it. It isn’t how much autonomy a system has. It’s how confidently a finance leader can account for every decision it makes. 

Governed Autonomy is how you get there. Give AI more authority when the evidence supports it, take it back when it doesn’t, and be clear about the things it can never touch. The finance teams that do this will get the efficiency they were promised without giving up the accountability they can’t afford to lose. 

Related Articles

Back to top button