GS interviews Toffer Grant, founder of PEX

AI is gaining ground in financial operations, but many finance leaders remain reluctant to entrust it with financial decisions. A recent PEX study revealed a persistent gap between AI adoption and confidence in its capabilities, even among large organizations. Green Sheet spoke with Toffer Grant, founder of PEX, about what it takes to build that trust, how businesses can measure AI's value and where human judgment should remain central as financial processes become increasingly automated.

Green Sheet: If lack of trust is the biggest barrier to AI adoption in finance, what specifically must an AI system demonstrate before finance leaders should allow it to make decisions rather than simply recommend them?

Toffer Grant: Trust has to be earned through repetition and adopted in a Crawl, Walk, Run phased approach. Start at a crawl by letting AI handle a low-risk workflow, like receipt tracking, and see what kind of time it can save your team. Can it meaningfully remove a small part of the grind and free up an employee for more intensive work? That kind of measurable impact is a critical hurdle any AI system must overcome before a finance leader can trust it with more critical work.

From there, you can walk by expanding AI into workflows where the stakes are higher, but a human is still in the loop. Let it flag unusual transactions, suggest how expenses should be categorized or identify potential policy violations, while leaving the final decision with an employee. Finance leaders can then see how consistently the system gets it right and where human judgment is still necessary.

The run stage is when AI begins taking action on its own, such as enforcing certain policies or handling routine financial decisions within clearly defined parameters. But autonomy should be earned, not assumed. Before getting there, the system should have a track record of making accurate recommendations, operating within established controls and knowing when a decision needs to be escalated to a person.

GS: AI use rises sharply with company size, yet discomfort remains high at the largest organizations. What does that tell you about the relationship between experience with AI and trust in it?

TG: It tells us that AI usage does not automatically equate to trust. The largest organizations in our study use AI at more than twice the rate of the smallest organizations, yet 42 percent are still uncomfortable letting it make financial decisions. Across businesses of all sizes that we surveyed, discomfort was relatively uniform. Together, that tells us that just because a business is using AI doesn’t mean it’s realizing the returns that actually build trust.

There’s also a big difference between using AI and giving it authority over money. Before a business is willing to let AI autonomously pay a $15,000 invoice, it needs proof that the system will reliably pay $15,000 and not accidentally pay $150,000. Those consequences only get bigger at larger organizations, where the volume and value of transactions can be much higher. More experience with AI can help build trust, but finance leaders still need evidence that it can operate accurately and consistently before they hand over decision-making authority.

GS: PEX recommends starting with lower-risk tasks such as receipt matching. What should finance teams measure in those early deployments to determine whether AI is actually delivering a return?

TG:The strongest gains to be had from deploying AI address a specific operational challenge and can be tracked with measurable outcomes. For something like receipt matching, I’d keep the scorecard pretty boring: How much manual review disappeared? Are receipts getting into finance faster? Is it cutting back time spent chasing down employees who didn’t submit that receipt?

That’s where the early value is showing up too. Among teams that are piloting or using AI, 69 percent have reduced manual review and 51% have shortened close time. Those are tangible outcomes a finance leader can point to, rather than simply saying their team is using AI.

The other thing I’d watch is whether those gains hold up over time. Saving a few minutes on one task isn’t enough if employees constantly have to check or correct the AI’s work. The real test is whether it can consistently take work off the team’s plate without creating new work somewhere else. If it can do that, you have a much stronger case for expanding AI into more complex workflows.

GS: Only 18 percent of respondents enforce spending policy at the point of transaction. How could AI change real-time spend controls without creating false declines or interfering with legitimate purchases?

TG: The bigger issue here is that finance still manages too much spending after the money has already left the building. Forty percent of teams review and flag issues after the fact, while, to your point, just 18% enforce policy when the transaction happens.

AI creates an opportunity to move more of that control upstream without turning every policy into a blunt yes-or-no rule. Instead of looking at a purchase in isolation, AI can help evaluate the context around it: Is this a normal purchase for this employee? Does it fit the purpose of the card and the company’s policy? Is there something unusual about the amount, merchant or timing that warrants another look?

But the goal shouldn’t be to have AI decline anything that looks slightly unusual. It should be to identify the transactions that truly need intervention and know when to bring a human into the loop. That’s particularly important when just 28% of finance leaders are comfortable letting AI make routine financial decisions today. The best real-time controls should make legitimate spending easier while giving finance a chance to catch genuine problems before the money is spent.

GS: As AI moves from flagging anomalies to taking autonomous action, where should the line remain between machine decision-making and human judgment in financial operations?

TG:The line should be between automating repetitive tasks within boundaries the business has already set and making a new judgment on behalf of the business. If a company has clearly defined that employees can spend up to a certain amount with certain merchants for certain purposes, AI should eventually be able to apply those rules without finance reviewing every transaction.

Human judgment becomes more important when the answer depends on context the system may not have. An unusual purchase isn’t necessarily an inappropriate one, and an exception to policy isn’t necessarily a problem. AI can surface those situations and provide the information needed to evaluate them, but people should still be the final decision-makers.

That’s ultimately where I see the division. People establish the boundaries and make the judgment calls; AI handles more of the repetitive decisions that fall cleanly within those boundaries. The goal isn’t autonomous finance for its own sake. It’s using autonomy where it actually removes work without removing judgment where it matters.

GS: Middle market companies often lack the technology and staffing resources of large enterprises. What should they be doing now to avoid either falling behind on AI or adopting it before they have adequate controls in place?

TG: Middle market companies don’t need the AI strategy or resources of a $10 billion enterprise to make meaningful progress. The most important thing is to just start somewhere. Look at the technology and providers you already have, identify one low-risk, repetitive process where AI can solve a real problem, and establish clear boundaries around what it can and can’t do.

From there, test, learn and expand incrementally. Measure whether AI is actually saving time or reducing manual work before bringing it into more complex workflows, keeping people involved as the stakes increase. Once the technology has proven itself, businesses can start giving AI greater autonomy over routine decisions and eventually use it to support higher-order work like forecasting, identifying financial risks or recommending where to allocate resources, with the right controls in place.

That incremental approach is how middle market companies can avoid both falling behind and getting too far ahead of their skis. You don’t need a massive AI budget or dedicated team to get started, but you also shouldn’t give AI more responsibility than the results and controls you’ve built can support.

Notice to readers: These are archived articles. Contact information, links and other details may be out of date. We regret any inconvenience.