Redwood Data: AI Agents and Governance in Finance

Enterprises are deploying AI agents faster than they can govern, connect or monitor them, exposing businesses to unexpected costs and operational risk, according to new research from Redwood Software.

The orchestration platform provider surveyed 300 senior technology, risk and AI leaders via research firm Leger Opinion. 

The resulting report highlights a clear divide between adoption speed and control. 

While 89% of organisations increased agentic deployment over the past year, existing systems and fragmented ownership continue to impede progress.

Only 22% of respondents fully trust AI agents for an autonomous financial transaction. 

Speaking to Finance Chief exclusively, Billy Newman, CFO of Redwood, says: 

!It doesn't surprise me that financial transactions are where trust in AI agents sits lowest. Finance is one of those industries where being roughly right just isn't good enough. 

“If a marketing model misses, you adjust the campaign. If a payment posts to the wrong entity, you're explaining it to your auditor and your board. That caution is healthy, but it shouldn't become paralysis. 

“The way you earn trust is by starting with the rules-based work, building audit trails into it from day one and proving the output holds up month after month. Trust follows evidence, not enthusiasm.”

Uncontrolled agents drive up business costs

The financial and operational fallout of unmonitored AI is already hitting balance sheets. 

The survey found that 77% of respondents have experienced a negative consequence from an ungoverned AI agent.

Unexpected cloud costs were the most common issue, followed by operational disruption and compliance problems.

Billy Newman says: “77% reporting negative consequences from an ungoverned AI agent tells me governance didn't keep pace with deployment. Most finance teams I talk to didn't get burned by the technology itself. 

“They got burned because nobody defined who owns the process. You need segregation of duties. 

“You need to know which automation touched which entry. And somebody has to review those rules as the business changes, because the rules you wrote last year may not fit the company you are today.”

Despite these risks, adoption continues to accelerate. Overall, 54% of respondents report scaling or mature agentic adoption. 

However, control mechanisms lag behind deployment. Just under half (42%) of organisations maintain a centralised registry of deployed agents, and just 51% have real-time monitoring of agent activity in place.

Legacy infrastructure blocks scaling efforts

Strategy is no longer the main bottleneck for AI expansion. Instead, 82% of leaders agree that existing systems present bigger obstacles to scaling AI than a lack of strategy. 

Only 42% report a strong ability to connect AI agents to legacy or core applications.

Ownership of agentic initiatives remains equally fractured across corporate structures.

Responsibility is currently split between IT Operations, dedicated AI or automation teams and software engineering, leaving no single function accountable for overall governance.

Over half (54%) of respondents say they feel fully prepared to orchestrate AI across existing systems.

Observability key to future expansion

To expand agentic AI with confidence, 62% of respondents pointed to better observability or real-time visibility, while 58% cited a proven ability to integrate AI agents with legacy systems.

Roughly one in nine businesses (11.3%) has started pulling back AI spending due to return on investment issues, while an additional 21% report that although automation provides value, expenditure remains unnecessarily high.

Billy concludes: “You've got agentic deployment expanding at nearly 90% of organisations, while roughly one in nine is already pulling back AI investment. Both things are true at once. 

“Companies scaled usage before their data was clean or their processes were stable, and the results didn't hold. 

“My advice to finance leaders is unglamorous. Pick the highest-volume, most repetitive work. Fix the data feeding it, set KPIs before you flip anything on, then measure honestly. The teams that slow down at the start are the ones that don't have to retreat eighteen months later.”