Finance spent the first wave of generative AI asking whether the technology could produce a useful answer. Agentic AI raises a harder question: What happens when AI can act on that answer?
Agents can gather data, apply rules, investigate exceptions, and take the next step in a workflow with limited human direction: updating a forecast in FP&A, or classifying a transaction and prepping a workpaper in tax. The appeal is obvious. So is the pressure to move fast. Deloitte's Finance Trends 2026 survey found 63% of finance teams have fully deployed AI, and 14% are already running fully integrated AI agents. A separate Deloitte survey found 74% of business leaders expect to deploy agentic AI within two years.
Autonomy is outpacing confidence
The control environment isn't keeping pace. In Deloitte's Q2 2026 CFO Signals survey, only 43% of CFOs said they felt fully confident in their AI governance, and 59% named balancing speed against risk as their top challenge in building an enterprise-wide framework.
That gap matters more with agents, because the risk is no longer a questionable answer on a screen. It's a decision an agent already acted on. An agent can draw from multiple systems, apply business rules, make intermediate choices, and pass an output into the next process before anyone reviews it.
For FP&A, that raises a traceability question. Which data, drivers, and assumptions produced this recommendation? For tax, the bar is even higher. Recent IRS guidance made clear that due diligence, competence, and confidentiality remain the practitioner's responsibility — AI can assist the work, but it doesn't own the outcome. Some tax teams are already building toward that standard. Ahead of an Oct. 14 session, tax and transformation leaders from Crowe Advisory LLC plan to walk through how they're building AI-enabled tax workflows on a governed foundation, rather than layering AI onto data and rules no one has validated.
Build guardrails around the work
A general AI policy isn't enough. Finance needs guardrails built around the actual work an agent performs, including governed, version-controlled data and business logic, clear limits on which systems an agent can touch and which decisions require a human, testing against known outcomes, and activity records complete enough to reconstruct what happened and why. A recent Financial Executives International framework for AI in financial reporting points to the same combination of human review, performance testing, independent comparison, and data analytics.
Guardrails are the path to scale, not the price of caution
It's tempting to treat governance as the cost of moving carefully. The evidence says the opposite. KPMG's 2026 AI in Finance research found organizations that can produce AI-related audit evidence efficiently report three to six times the rate of significant performance improvement compared with those that can't.
INVESTBANK is a live example. Facing 46 new mandatory regulatory reports, the bank's risk team built governed, auditable reporting workflows in Alteryx — cutting report preparation time by 90% while making version control and auditability easier to manage. The governance wasn't traded for speed, it's what made the speed possible.
The lesson is to earn autonomy:
- Start with clear goals and define success
- Measure it
- Keep human judgment where it matters
- Expand an agent's authority only as the evidence supports it.
Speed alone won't determine whether agents create durable value. Finance organizations that build visible, understandable, repeatable and auditable workflows can give agents room to act without losing the ability to explain and defend the result. The strongest guardrails may turn out to be the fastest route to scale.
