.png)
Key Takeaway
AI agents can genuinely help finance teams, but the label is being stretched far beyond what many products deliver. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027. The teams that avoid that outcome ask hard questions up front and start with narrow, well-governed use cases.
Open any EPM vendor's website right now and you'll find the word "agent" somewhere near the top. Planning agents, forecasting agents, close agents, reporting agents. The promise is a finance function that largely runs itself.
Some of it is real. Some of it is a chatbot or a rules-based automation with a new name. For a CFO or FP&A leader trying to decide where to put budget and attention, telling the two apart has become a skill in itself.
The question isn't whether AI will change how finance works. It's which claims will hold up once the demo ends and your own data goes in.
Gartner has a name for this: "agent washing". It describes vendors relabelling existing products, such as chatbots, robotic process automation and AI assistants, as agentic, even though they lack substantial agentic capability. Gartner estimates that only about 130 of the thousands of agentic AI vendors are real.
Sources: Gartner press release, June 2025 (first two figures; cancellations attributed to escalating costs, unclear business value or inadequate risk controls). Datarails 2026 survey of 270 US CFOs and finance leaders (third figure; vendor-sponsored, self-reported).
That auditability number matters most for finance. Planning and reporting outputs end up in board packs, lender conversations and audited statements. "The AI said so" is not an answer anyone can give an auditor.
A useful way to cut through the noise is to separate tasks where AI supports a person from claims that it can replace one. The table below is a practical lens, not a verdict on any specific product.
Practical Capability vs. Overstated Claims
| Area | Where It Tends to Help | Where to Be Sceptical |
|---|---|---|
| Reporting & Commentary | Drafting variance explanations and summaries for a person to review and edit | Publishing board-ready commentary with no human review |
| Data Preparation | Flagging anomalies, suggesting mappings, matching transactions and surfacing exceptions | Fixing broken source data on its own, or working reliably on ungoverned data |
| Forecasting | Offering a statistical baseline that planners can challenge and adjust | Replacing planner judgement and business context entirely |
| Scenario Analysis | Speeding up the building and comparison of what-if cases | Choosing the right scenario or making the call on the response |
| "Autonomous" Finance | Handling well-defined, rule-bound steps inside a controlled workflow | Any claim that a whole process, such as the close or the plan, now runs unattended |
What Does the Agent Actually Do, End to End?
Ask to see a specific task from trigger to finished output in a realistic environment, not a scripted walkthrough. If it only answers questions or follows a fixed rule, it may be an assistant or a workflow, which can still be useful but isn't what was advertised.
Can I Trace Every Output Back to Its Source?
Explainability is a control requirement in finance. Every figure and every piece of commentary should link back to the data and logic behind it, so a reviewer or auditor can follow it.
Where Does a Human Approve?
The answer shouldn't be "nowhere". Look for clear approval points, role-based permissions and a full log of what the agent did and who signed it off.
What Does It Need From My Data?
An agent is only as reliable as what it reads. If your definitions differ between teams or your hierarchies are inconsistent, expect confident answers built on shaky foundations. Be honest about whether your data is ready.
What Will It Cost at Scale?
Escalating cost is one of the reasons Gartner gives for projects being canceled. Understand how pricing behaves as usage grows, and what the business case looks like once the pilot ends.
A good vendor will welcome these questions. If the answers get vague, that tells you something too.
The common thread in failed AI projects isn't usually the model. It's the foundation underneath: disconnected planning processes, inconsistent definitions and no clear ownership of data. Agents don't remove those problems, they work faster on top of them.
If your planning is already connected, your data governed and your processes documented, agents have something solid to build on. If not, the most valuable AI investment may be the unglamorous work that comes before it.
Pick one narrow, well-defined use case where the cost of an error is low and a person can easily review the output, such as drafting variance commentary or flagging anomalies in a reconciliation. Set a clear measure of success, keep a human in the loop, and expand only when the results hold up. Small proven wins build trust inside the team. Big promises that don't land do the opposite.
Keansa helps finance teams cut through the noise on AI in EPM: assessing where it can add real value, getting the planning and data foundations right, and introducing AI capabilities with the governance and review steps finance needs, so what you adopt is something you can trust and explain.
Want an honest view of where AI fits in your planning process?
Talk to a Keansa Consultant.png){kind=small}
.png){kind=small}
.png){kind=small}