Is Your Finance Process Ready for AI? Use the Smart New Joiner Test
A finance process is ready for AI when its inputs, rules, exceptions and escalation paths are clear enough for someone new to operate it with limited supervision.
I have started using a simple test to judge whether a finance process is ready for AI:
Could a smart new joiner run the process without constantly asking someone what to do?
If the answer is no, an AI agent will probably struggle too.
This test shifts the discussion away from tools and prompts. It forces us to examine whether the process itself is clear.
What the test reveals
Many finance processes appear structured from a distance. The detailed reality is different.
Data comes from several systems. Someone cleans it in Excel. Some rules are documented. Others are understood only by experienced team members. Exceptions are spotted because someone looks at a number and says, “This does not look right.”
The process works because experienced people compensate for gaps in its design.
An AI system cannot reliably compensate for rules that have never been made visible.
Before introducing AI, the process needs five things:
- Defined inputs and source systems.
- Visible calculation and policy rules.
- A clear distinction between normal cases and exceptions.
- Named owners for investigation and resolution.
- An escalation path when evidence is incomplete or judgement is required.
These are process-design requirements. They are also the foundation of reliable finance automation.
Start with process clarity
Finance teams should be able to answer a few direct questions:
- What event starts the process?
- What information is required?
- Which checks must always be performed?
- What makes a transaction or balance unusual?
- Which decisions can follow an approved rule?
- Which decisions require finance judgement?
- What evidence must be retained?
- When should the process stop and escalate?
If the team cannot answer these questions consistently, adding AI may increase speed without improving control.
The first task is to make the process explicit.
Move beyond isolated AI outputs
Consider variance analysis.
A narrow use case asks AI to explain a variance. The system receives a number and prepares commentary.
A more useful workflow looks different:
- The variance is calculated using approved data.
- Material movements are flagged using defined thresholds.
- Supporting information is collected.
- Likely drivers are identified.
- The responsible owner is notified.
- A proposed action is drafted.
- Resolution is tracked.
- Finance reviews and closes the exception.
AI may help investigate the cause, organise evidence and draft the explanation. The workflow around that intelligence creates the operating value. The applied variance-analysis example shows this boundary in practice.
Most of the design effort sits in understanding what should happen next.
Process judgement is an AI skill
Finance professionals already use many of the capabilities required to supervise AI-enabled workflows.
They define processes. Set control points. Review evidence. Handle exceptions. Challenge unusual outcomes. Improve the process after something fails.
This is process judgement.
Prompting and technical knowledge matter. So does knowing:
- What should happen.
- What can go wrong.
- Which exceptions matter.
- Where judgement should remain human.
- What evidence is sufficient to proceed.
These are familiar finance disciplines applied to a new operating environment.
A practical readiness test
Before approving an AI use case, ask the process owner to explain how a capable new joiner would run it.
If the explanation depends heavily on phrases such as “usually,” “it depends,” “the team knows,” or “someone checks it,” the process needs more design work.
Document the missing rules. Identify the hidden decisions. Define the exception paths. Establish the evidence required for review.
Then decide where AI adds value. The related guide on choosing between rules, an LLM and an agent provides the next decision step.
The technology is new. The discipline of designing processes that work without constant supervision is not.