The formula is not the control.

A formula can calculate the expected result and still sit inside an unreliable process.

It may reference the wrong source. It may silently inherit a bad value. It may produce the correct answer today and fail when a column moves next month. Someone may overwrite it, bypass it, misunderstand it, or act on its output before an exception is reviewed.

The formula is visible. The control system surrounding it usually is not.

A calculation produces an answer. A control establishes why that answer can be trusted.

From workbook to workflow

Spreadsheets often evolve beyond their original purpose. A tracker becomes a source of truth. A formula becomes business logic. A status column becomes an approval system. A manual note becomes institutional memory.

When AI and Python enter this environment, they do not merely automate cells. They begin interacting with policies, documents, approvals, exceptions, and systems of record.

The workflow becomes programmable. That makes its architecture an accounting concern.

AI changes the control surface

Traditional spreadsheet logic is usually deterministic. The same input and formula should produce the same output. AI introduces probabilistic interpretation, changing model behavior, contextual dependence, and outputs that may appear plausible even when they are wrong.

This does not make AI unusable in accounting. It changes where control must exist.

Use deterministic logic for stable requirements

Required fields, arithmetic checks, duplicate detection, threshold tests, status transitions, and referential integrity should generally remain explicit and testable.

Use AI for interpretation

AI can help summarize documents, identify missing context, propose classifications, explain anomalies, or formulate questions for a reviewer.

Do not confuse a proposal with an approval

An AI-generated recommendation is evidence entering the decision process. It is not automatically the authorized accounting decision.

The safest AI workflow does not ask a model to be certain. It makes uncertainty visible and routes it appropriately.

The source-of-truth problem

Many finance failures are not calculation failures. They are source-selection failures.

A process may contain several plausible versions of the same fact: a contract, CRM field, invoice, payment platform, email approval, spreadsheet status, or prior-period treatment. Automation cannot resolve those conflicts safely unless the hierarchy is explicit.

A dependable system should know:

1. Which source is authoritative for each field.
2. Which fallback is permitted when the preferred source is unavailable.
3. Which conflicts require human review.
4. Which invalid values must stop the workflow.
5. How the final decision is recorded and reconstructed.

Safe failure is a feature

Automation is often evaluated by how much work passes through without intervention. In accounting, the quality of the stop condition can matter just as much.

A strong workflow should refuse to continue when required evidence is missing, a benchmark is invalid, an approval state is ambiguous, or a write would be irreversible.

That failure should be visible, classified, and routed. It should not disappear into a generic error message or silently substitute a convenient value.

What leaders should require

1. Which source is authoritative for every material input?
2. Which decisions are deterministic and which involve judgment?
3. How does the system express uncertainty?
4. Which exceptions stop execution?
5. Who can approve irreversible actions?
6. Can every output be traced to its inputs, logic, and reviewer?
7. How are model, prompt, policy, and schema changes tested?
8. Can the workflow run in observation-only mode before production?
9. How is downstream reconciliation performed?
10. Who owns the result when several systems participate?

The accountant’s changing role

AI does not eliminate the need for accounting judgment. It makes the design and placement of that judgment more important.

The accountant increasingly helps define authoritative sources, encode policy boundaries, identify exceptions, design review points, test failure behavior, and determine what evidence must accompany an output.

This is not software engineering replacing accounting. It is accounting becoming more explicit about how reliable work is produced.

What remains to learn

The next stage of this inquiry will interview accountants, controllers, finance-operations leaders, auditors, and accounting-technology builders.

The goal is to learn where AI-assisted finance workflows actually fail, which controls transfer well from traditional processes, and which new forms of evidence are required when the workflow includes probabilistic systems.