AI has a context problem

The invisible data architecture behind every great AI experience

Everyone thinks AI has an intelligence problem. It doesn't. It has a context problem. Only 11% of finance workflows run on their own. The bottleneck isn't AI itself. It's the data architecture underneath it.

Unless stated, all statistics referenced below come from "Building An AI-Ready Finance Function", a commissioned study conducted by Forrester Consulting on behalf of Airwallex, June 2026.

In 1900, the looms in a Massachusetts cotton mill still ran off leather belts hanging from one spinning steel shaft in the ceiling, the same rig steam had driven for decades, except an electric motor now turned it. Paul David, the Stanford economic historian, showed that electricity lifted productivity only forty years later, once owners scrapped the shaft, gave every loom its own motor, and rebuilt the mill floor around the new layout.

Finance is repeating the pattern with AI. Most AI strategies mistake software for architecture. Adoption climbs, budgets swell, and boards approve pilots by the quarter. Yet a June 2026 study by Forrester Consulting, commissioned by Airwallex and covering 1,279 finance decision-makers, found that only 11% of finance workflows run without human intervention.

The reason isn't that today's AI models aren't capable enough. It's that most organisations are still running on fragmented data architectures built long before AI arrived. Until AI has complete, trusted context, it will remain an adviser rather than an operator.

Finance AI runs on context, not prompts

The AI people interact with is the smallest part of the system. It appears as a chat window, a recommendation, or an anomaly flagged in amber. Beneath that sits the architecture that determines whether AI can simply make suggestions or safely execute decisions.

For an AI agent to approve a payment, it needs to understand far more than an invoice. It needs to know which entity owns the funds, what currency they're held in, whether another payment is already pending, who has approval authority, and whether the transaction complies with local regulations.

None of that comes from the model. It comes from the data architecture creating context beneath it.

Three layers determine whether AI can act

Data architecture comes first: one consistent view of balances, transactions, approvals, and entities across every market and business.

Workflow architecture comes second: approvals and controls that allow an agent to operate safely within established governance rather than constantly asking for permission.

Infrastructure comes third: licences, payment rails, and scheme connections that allow transactions to execute in the real world.

Miss any one of these layers and AI can generate recommendations. Get all three right and it can execute decisions safely.

Fragmented stacks are stopping AI at the data layer

Forrester asked finance leaders what prevents AI from scaling, and 65% pointed to fragmented, inconsistent finance data. Integration complexity across regions and entities followed closely behind at 61%. Both answers describe the same underlying problem.

Inside a fragmented stack, AI fails in predictable ways.

• An agent that sees spend but not cash position approves purchases the balance can't cover.

• An agent that sees cash position but not pending settlements blocks payments unnecessarily.

The problem wasn't the model. The problem was the missing context. Each mistake teaches the finance team to add another guardrail or approval step.

A unified data architecture changes that equation. When every payment, balance, approval and settlement exists within the same trusted context, AI can make decisions with confidence rather than partial information.

The foundation nobody notices

Every autonomous financial decision depends on a chain of infrastructure working together. Licences determine where money can legally move. Direct connections to local payment rails determine how efficiently it moves. Card scheme memberships, transaction data and workflow controls provide the context that tells AI what should happen next.

Individually, these are infrastructure investments. Together, they create something far more valuable: a single, trusted view of financial activity that AI can reason over.

That foundation takes years to build.

Over the past decade, we've invested in the infrastructure required to operate globally. Today, Airwallex holds more than 85 financial licences worldwide, operates as a licensed money transmitter in 45 US states, and connects directly to local clearing networks in more than 120 countries, keeping over 90% of transactions off SWIFT.

None of these investments were made with AI in mind. They were made because moving money globally demands trust, regulatory depth, and clean operational architecture. As AI has matured, that same foundation has become the prerequisite for autonomous finance.

The commercial benefits are already visible. Minor Hotels replaced a fragmented payments stack with a single Airwallex platform, saving more than US$7million annually through like-for-like settlement and local acquiring. The same unified architecture that reduces operational complexity also creates the consistent data foundation and context that AI needs to automate financial workflows safely.

Security as a load-bearing wall

As AI takes on greater authority across finance workflows, every decision it makes inherits the controls beneath it. An agent that can approve payments or move money is only as secure as the identity controls, access permissions, and audit trails supporting those actions. Weak security doesn't just increase risk, it limits autonomy. Finance teams simply won't delegate meaningful decisions to systems they can't fully trust.

That's why security architecture has become part of AI architecture. The same unified foundation that gives AI complete context must also enforce who can access data, what actions they can take, and how every decision is recorded.

For finance leaders evaluating AI platforms, the question isn't simply whether a vendor has AI capabilities. It's whether the security architecture beneath those capabilities is robust enough to support autonomous financial operations at scale.

AI rewards the architecture beneath it

The mills that prospered during electrification weren't the ones that bought electric motors first. They were the ones that rebuilt their factories around them.

Finance is reaching the same moment.

The organisations that see the greatest gains from AI won't simply adopt better models. They'll build the data architecture those models depend on: connected systems, consistent transaction data and infrastructure that gives AI complete context for every decision.

Licences, payment rails and security matter because they strengthen that foundation. AI may be the visible innovation, but data architecture is the invisible advantage that makes autonomous finance possible.