Rohit Gupta is the CEO and co-founder of Auditoria.AI, a pioneer in AI-driven automation solutions for corporate finance teams.
getty
Six months ago, someone on your accounts payable team approved an invoice with missing documentation. Your enterprise resource planning (ERP) system knows the date, the amount and who clicked approve. It has no idea why.
Business context shapes routine decisions across finance, but it rarely sits in one system. In accounts receivable, why was one customer given another 30 days while another was escalated to collections? In accounts payable, why did an approver make an exception for this supplier? The answer may sit in an email thread, an approval comment, a customer call, an earlier dispute or the memory of the person who handled the same situation last year.
Deloitte’s Finance Trends 2026 research, based on a survey of 1,326 finance leaders, found that 63% of finance functions had deployed and were actively using AI. Among those already using it, however, only 21% believed the investment had “already delivered clear, measurable value.” Just 14% had fully integrated AI agents into the finance function.
AI can already answer almost any finance question. Whether that answer would survive contact with your policies and your history is another matter. Contextual AI is the difference between the two.
Finance work begins where the rule ends.
In my years working with finance teams, one truth holds everywhere: Exceptions consume all the time. Ten thousand invoices can flow through the system without attracting attention. One that doesn’t match will occupy three people for the rest of the afternoon.
I’ve seen every version of this. The invoice is legitimate, but the supplier used a different legal entity. An approval falls outside policy, although a senior finance leader previously authorized the same exception under similar circumstances.
This is why feeding an AI system more financial data does not make it more useful. Volume and context are different things. A thousand rows containing an internal code such as “05A” still tell the system very little unless it also knows what that code means, which business unit uses it and how the finance team interprets it.
Context gives the numbers their address.
The missing record is often the decision itself.
Most finance systems preserve the outcome of a decision. They rarely preserve a complete account of how the decision was reached.
Now imagine the same supplier submits another invoice under similar circumstances. The AI finds the earlier approval. The question is whether it understands why that exception was made. The final status says “approved.” The history tells the system whether that approval should become a precedent.
Contextual AI should be able to reconstruct that history. It should locate the transaction, retrieve the relevant communications and approvals, identify the applicable policy and show the sequence of evidence that led to the conclusion. When the same issue returns, the earlier judgment can inform the analysis without silently becoming a universal rule.
A new accounts payable manager may understand accounting and controls on the first day. It takes longer to learn that a particular supplier invoices through several subsidiaries or that a certain class of exception must always be escalated to the controller.
Over time, that person develops institutional memory. Contextual AI gives finance a way to preserve and use that memory more deliberately.
It also makes the AI easier to challenge. A user should be able to ask why a vendor was assigned a high-risk rating and which sources contributed to the assessment. The system should be able to return to its earlier analysis instead of starting again with no recollection of what it previously considered.
Trust comes from showing the work.
A 2025 global study from KPMG and the University of Melbourne surveyed 48,340 people across 47 countries. It found that 66% of employees using AI did not regularly evaluate its output for accuracy, while 56% reported making mistakes in their work because of AI.
Finance cannot build its control environment on the assumption that someone will catch every unsupported conclusion as it appears. The system must show its work.
For a recommended action, that means showing the relevant transaction, applicable policy, supporting communications, earlier cases considered and any external information used. It also means recording what the AI did so the analysis can be examined later.
This creates a more useful form of auditability. A log showing that an AI system accessed five applications is a technical record. A finance leader needs to understand how information from those applications supported the recommendation.
Context can support a measured path toward greater autonomy. Your finance team may permit AI to take action when the evidence is complete, the decision follows established policy and the circumstances fall within a defined level of authority. An unfamiliar exception, missing source or conflicting precedent should send the decision to the appropriate person.
The important boundary is not between human work and machine work. It is between a decision your organization can explain and one it cannot.
Context must accumulate over time.
Every resolved exception has the potential to strengthen your organization’s decision-making. The challenge is knowing which decisions deserve to become precedent and which should remain one-off exceptions.
This does not mean you should allow AI to absorb every historical action as accepted practice. Companies make inconsistent decisions, and people approve exceptions they later regret.
Your finance team must distinguish between an isolated decision and an approved precedent. They must preserve who had the authority to make the decision, when it was made and which version of the policy was in effect. They also need a clear way to retire outdated information so yesterday’s exception does not quietly become tomorrow’s rule.
As you assess the growing number of AI agents entering the market, look beyond whether a system can answer a question or complete a task. Can it explain the context and evidence behind its recommendation? Can it show measurable improvements in areas such as cash performance, exception handling or reduced manual effort? If its decisions cannot be explained and its impact cannot be measured, it is not an effective finance agent.
The information provided here is not investment, tax or financial advice. You should consult with a licensed professional for advice concerning your specific situation.
Forbes Finance Council is an invitation-only organization for executives in successful accounting, financial planning and wealth management firms. Do I qualify?
