GPT-5 is here: Why finance still can’t “let it post”

GPT-5 is here. But it still can’t close your books. It’s faster, cheaper, and better at following instructions. It handles longer chains of logic. It integrates cleanly with external tools. It is, without question, a technical leap. But yep, it still can’t close your books. Not because it lacks intelligence but because it lacks integrity; the paradox we now live in.

The GPT-5 paradox: What “finance-grade” really means

Trust in AI starts with verifiable outputs, enforceable constraints, and auditable systems.

GPT-5 may make it easier to build systems, but it doesn’t make those systems safe. Until an AI can prove its work, it’s still just a very smart intern — eager, articulate, and unaccountable. That part is still on us.

To close the trust gap, we need more than intelligence and infrastructure. And that means four important things:

1. Control. We need outputs we can trace, constraints we can enforce, and systems we can audit. This isn’t just explainability — it’s verifiability. Debits must always equal credits. Logs must be immutable. Logic must be replayable.
2. Integration. The data mess has to go if AI in finance is going to deliver value, and if your ERP is your backbone, then that starts with direct ERP integrations. Too many enterprises juggle fragmented ERPs, disconnected BI tools, and overlapping spreadsheets, none designed for machine reasoning. Giving AI raw access to that chaos is like handing a blindfolded genius the keys to a glass warehouse. You need clean, connected systems that feed clear, explainable data through finance products AI can trust — and Audit can verify.
3. Reliability. Models break when the world changes. Interest rates spike. Policies shift. A forecast that looked solid last week can collapse overnight. We need stop rules, uncertainty flags, and humans ready to step in — not after the fact, but when it matters.
4. Accountability. Who owns the decision? Scaling AI in finance means redefining finance roles so humans own the judgment, not just the outputs. Controllers become control architects. Auditors shift focus from outputs to systems. Reviewers become ‘exception experts’. Everyone moves up the ladder of judgment.

We’ve done this before. When autopilot arrived, we didn’t take pilots out of planes; we redesigned the cockpit. We redefined roles. And we created trust not through the tool, but around it. Finance is at that moment now.

What hasn’t arrived yet and why AI in finance still needs the right systems

Finance doesn’t need smarter AI ideas — it needs systems, guardrails, and standards to make AI safe.

We don’t have verifiable finance reasoning graphs that trace every number, link it to policy, and show the logic that produced it. That’s the difference between a fluent draft and a signed, auditable entry. And we don’t have finance-native guardrails that recognise when a journal entry violates a cash flow identity, or when a set of accounts quietly stops balancing. In today’s LLMs, those principles aren’t hardcoded.

So while a model might draft a plausible entry or assemble a sharp-looking report, it has no inherent sense of whether it just broke accounting logic. It doesn’t know that debits must equal credits, or that cash flow from operations must tie back to net income and working capital. Without that embedded logic and without guardrails, AI remains fragile. It can sound right, and still be wrong.

We’re still teaching models the difference between correlation and causation, between “what happened” and “why it matters.” Auditors can’t yet sign off on AI-generated narratives because the standards don’t exist — and we don’t have control libraries that can be subscribed to like software.

But we will — and fast. This isn’t science fiction. The capabilities are here. All that’s missing is the infrastructure to catch them.

Your two-track plan for the year ahead

So what should a CFO do? Run two plays at once. Pre-design escalation flows and get guardrails approved now, before you need them.

One path is steady and governance-first. Assume that AI will roll out like other enterprise tech. And start now: design your controls, document them, and bring Audit and Risk in early so that when automation scales, your trust scales with it.

The other path is faster and riskier. Assume AI reliability could leap ahead of your controls. Prepare now: build kill switches, map escalation flows, and get guardrails approved before you need them — so when the tech is ready, you are too.

This isn’t about picking one future. It’s about being ready for both.

What you can do this quarter — simply, credibly, and with impact

If you want to prove AI belongs in finance, don't start with risk — start with friction. Target the pain your team feels every month and deliver relief that scales.

Start with the chase: Receipts, approvals, last-minute clarifications — not complex, just relentless. Perfect work for intelligent agents that understand urgency, context, and policy. 

Today, AI can follow up directly with employees in Slack, nudge approvers as deadlines approach, and escalate when it matters. It doesn't guess. It works from your close calendar, your approval flows, and your risk thresholds. It doesn't ask "Is this right?" It asks, "Is this overdue, out of policy, or unresolved?" And if the answer is yes, it acts.

This isn't the future of AI in finance: This is now. Variance explanations, missing receipts, compliance flags — these are all areas where an agent can propose before a human approves. Logs can track every message, every action, and every check. And you don't need to sacrifice auditability for speed.

If your team spends more time chasing than challenging, piloting an agent here will not only save time but also reset expectations and show what good feels like.

And if you're tired of seeing simple purchases trigger complicated workflows, look at procurement. Most teams don't have a purchasing problem, they have a context problem. Employees don't know the process, whether it needs a PO or a card, the right GL code, or the right cost centre. So, they guess, wait, or buy out of policy.

Now imagine an intelligent assistant that starts where the employee is — with a natural-language request — and guides them through the rest. It asks the right questions, knows the policy, figures out whether to raise a PO or fund a card, collects approvals based on pre-set logic, gives finance visibility before the money moves, and does all of this without switching tabs or opening another tool.

That's more than a procurement system; it's a teammate who knows the rules and makes them easy to follow.

Or take business travel. Booking is fragmented, policy is invisible, and reconciliation is manual. But what if employees could request a trip in natural language, get compliant options instantly, and receive confirmations, payments, and receipts automatically mapped to their trip?

That's not overreach, that's orchestration. And AI can already handle the handoffs. Pick the edge where repetition meets rules, where people are most fatigued by admin and most ready for help. That's how you get ROI in faster processes and faster belief. Your team begins to see AI not as a threat or a toy but as a colleague: tireless, trustworthy, and on time. That's how you build momentum and (well-deserved) trust. 

You don’t scale AI, you scale trust

GPT-5 doesn’t remove the need for trust. It raises the bar for how we build it.

When you invest in provenance, constraints, clean data, and clear roles, you’re not just making AI work for you - you’re turning it into operating leverage.

It won’t just help you shorten your close. You’ll also protect your policies and give your team back the hours to lead, not chase. And five years from now, when intelligent agents power AI in finance and close happens in near real time, you’ll look back on today and ask, “Why did we ever hesitate?” 

It’s not a technology problem. It’s a trust problem. And that’s a problem we can solve. 

Learn how our intelligent agents handle the chase, close gaps, and free your finance team to focus on strategy.