Rules, LLM or Agent? Choosing the Right Automation for Finance
A practical framework for choosing between deterministic automation, generative AI and controlled agentic workflows in finance.
Not every finance problem needs an AI agent.
The useful starting question is:
Does the process need consistent execution, intelligent interpretation or controlled autonomy?
The answer helps determine whether the right solution is a rule, a workflow, a large language model, an agent or a combination of them.
Use rules where the answer must be consistent
The closer a process gets to accounting, reconciliations, compliance, approvals and financial controls, the more predictable its core logic should become.
The same inputs and approved rules should produce the same result.
Deterministic automation is suited to activities such as:
- Calculations and allocations.
- Transaction matching.
- Threshold and tolerance checks.
- Policy validation.
- Workflow routing.
- Approval controls.
- Reconciliation logic.
A clear rule is easier to test, explain and audit than an AI-generated judgement.
If a rule can resolve the problem reliably, use the rule.
Use an LLM for interpretation
A large language model becomes useful when the facts are established but someone still needs to interpret or communicate them.
Suitable uses include:
- Drafting variance commentary.
- Summarising close issues.
- Preparing management explanations.
- Converting technical findings into CFO-ready language.
- Answering questions from approved finance data.
- Drafting follow-up questions for an exception owner.
The model supports interpretation. It should receive controlled inputs and produce an output for review.
This is intelligence without meaningful autonomy.
Use an agent for a multi-step investigation
An agent becomes relevant when the system must decide how to progress towards an objective.
It may need to determine:
- Which information to collect.
- Which authorised tool to use.
- What step to perform next.
- Whether the available evidence is sufficient.
- Whether to continue investigating or escalate.
- Which person should review the result.
This is a different operating model.
Generating commentary from a prepared trial balance may require an LLM.
Investigating a material variance across the ERP, procurement, inventory and operational systems may require an agent. The system must gather evidence, test possible causes and determine when the case is ready for finance review.
Apply the distinction to variance analysis
A controlled variance-analysis workflow can use all three layers.
Rules
- Calculate actual versus budget or forecast.
- Apply materiality thresholds.
- Identify unusual movements.
- Confirm that the reporting period and source data are complete.
LLM
- Summarise the movement.
- Draft possible explanations from approved evidence.
- Prepare questions for the business owner.
- Convert the analysis into management commentary.
Agent
- Select relevant supporting sources.
- Retrieve operational and financial evidence.
- Test likely drivers.
- Request missing information.
- Assemble an investigation pack.
- Escalate low-confidence or high-risk cases.
Finance remains responsible for the conclusion and action.
Define the control boundary
Autonomy should be designed, not implied.
For every agentic workflow, finance leaders should define:
- Which systems the agent can access.
- Which actions it can perform.
- Which actions require approval.
- What evidence must be recorded.
- What confidence level triggers escalation.
- Who reviews the output.
- How errors and overrides are logged.
An agent that can retrieve evidence and draft a recommendation presents a different risk from an agent that can post a journal or release a payment.
The permission boundary must reflect that difference. The wider AI in finance framework and continuous controls monitoring model show how rules, evidence, ownership and review fit together.
Five questions before selecting the technology
Before approving a Finance AI solution, ask:
- Can a transparent rule solve the problem?
- Are the underlying facts already reliable?
- Is the requirement interpretation or action?
- Does the workflow genuinely need to choose its next step?
- Which decisions must remain with finance?
These questions reduce unnecessary complexity.
They also help prevent teams from labelling every LLM-enabled workflow as agentic.
The objective is not to use the most advanced technology. It is to build a finance process that is useful, controlled and trusted.
Read the original LinkedIn post on intelligence versus autonomy