The Ops Accountant · Analyze · AI

Analyze · Case study 5 · AI

AI in the review process

AI assistants such as Copilot and Claude are now part of the review toolkit. Used well, they take the first draft off your plate: the query, the formula, the variance write-up. Used carelessly, they hand you a confident number nobody checked. The rule that makes the difference is simple: AI drafts, a person decides, and every number is traced to its source.

Where it helps

Writing the tool

First drafts of SQL, Power Query M, DAX and Excel formulas from a plain description of the check.

  • Faster than starting from a blank editor
  • The draft is tested against data before it's trusted

Explaining the numbers

Drafting variance explanations and review notes from figures you've already confirmed.

  • Consistent wording across wells
  • Lists what still needs confirming

Checking completeness

Comparing a journal entry's support against a checklist, or a procedure against what the team actually does.

  • Catches the missing attachment
  • Turns a process into a checklist

Finding the clause

Pointing to the section of a JOA or accounting procedure that covers a question, such as the overhead rate or the audit window.

  • The reviewer then reads the clause itself

Summarizing threads

Condensing a long partner dispute or email chain into what's been asked, what's been sent and what's open.

  • Useful before a follow-up call

Training material

Turning a review note or a fixed error into a short explainer for new staff.

  • Several pages on this site started that way

The guardrails

AreaRuleWhy
DataOnly company-approved tools see company data. Partner, owner, bank and pay details never go into a tool the company hasn't approved. When asking for help with a query or formula, describe the structure, not the records.JIB and revenue data is confidential to the company and its partners.
NumbersThe figures of record come from the ledger, Excel or SQL, never from the assistant's own arithmetic. Every number in an AI draft is ticked back to its source.A plausible-looking total is the most dangerous kind of error.
CodeEvery AI-drafted query or formula is run against data where the right answer is known, including at least one case that should fail.A query that runs isn't the same as a query that's right.
ContractsAI can point to the clause; the reviewer reads the clause. The JOA and accounting procedure govern, not a summary of them.Forms differ by version and by what was negotiated.
DecisionsAI never approves, posts, bills or sends anything to a partner. A person does, and owns it.Accountability stays with the reviewer.

A worked example: the first draft was wrong

The billed, not paid query started as an AI draft from this request: "List owners billed costs in October who have no October revenue." The draft ran, and returned a clean-looking list. It was wrong in a way that matters in this business:

First draft: joins on the well name

LEFT JOIN paid p
  ON p.well_name = b.well_name
 AND p.owner     = b.owner

Jones Unit 1's revenue was booked to the lookalike cost center CC-2210. Matching on the name alone found that revenue and reported Jones Unit 1 as paid. The error stayed hidden.

Reviewed: joins on well name + cost center

LEFT JOIN paid p
  ON p.well_name   = b.well_name
 AND p.cost_center = b.cost_center
 AND p.owner       = b.owner

Now Jones Unit 1 on CC-2201 shows as billed with no revenue, which leads straight to the lookalike cost center query and the fix.

The assistant wrote a reasonable query. It didn't know that names repeat across cost centers. That knowledge is the reviewer's job, and it's why the test data includes a case that should fail.

Example prompts

Draft a check

Describe the tables and the rule, not the records.

I have tables jib_lines (month, well_name, cost_center, owner, amount) and revenue (month, well_name, cost_center, owner, net_revenue). Write a SQLite query listing owners billed in a given month with no revenue that month on the same well name and cost center. Always join on both well name and cost center. Sort by amount billed, largest first.

Then: run it against data with a known exception, and check the exception appears.

Draft variance explanations

Give it the confirmed figures and what you already know.

Here are this month's accrual variances over $5,000, with the reason Operations gave for each where I have one. Draft a one-sentence explanation per line for the close file, in plain language. Where no reason is given, write "Reason to be confirmed with Operations" instead of guessing.

Then: tick every figure in the draft back to the review table before it goes in the file.

Check a journal entry's support

Here's our checklist for the monthly saltwater disposal and water gathering entry, and the list of attachments in this month's support. Which checklist items have no support attached? Don't evaluate the amounts.

Then: the reviewer still opens the support. The assistant only says what's missing.

Reshape an operator's file

Write a Power Query M function that takes a CSV with columns Period (YYYYMM), Property, Prop No, AFE, Acct Desc, Gross Amt, Your Decimal, Net Amt and returns columns month (YYYY-MM), operator_well, operator_prop_no, afe_no, cost_type, gross_amount, company_wi, net_amount. Treat a blank AFE as LOE.

Then: compare the function's totals to the file's totals. See the Power Query case study.

This page describes a general practice, not any company's policy. Follow your own company's rules on which AI tools are approved and what data they may see. The examples use the site's made-up sample data.