Lesson objective

Write an output contract that fixes format, flags, prohibited content and required tie-outs, then refine a failing prompt by diagnosing which layer failed and changing one thing at a time.

Plain-language explanation

Why this matters now

A prompt that says what to do but not what the result must look like produces a draft you have to read in full to check. An output contract changes that. It fixes the shape of the output, the flags the model must use, what it must never include, and the totals the preparer will tie out. Then, when a draft fails, the layered prompt lets you find which part failed and fix that part, rather than rewriting everything and hoping.

The stakes of an unchecked output are rising as AI output feeds other steps. The OWASP Top 10 for LLM Applications 2026, a voluntary community document dated August 4, 2026, lists LLM07 Misinformation. Its project leads note that voters placed Misinformation near the bottom while the incident record placed it near the top, and explain: "When a model's fluent, confident output drives a decision or a tool call, a wrong answer turns into a wrong action." The same list includes LLM10 Improper Output Handling. In accounting terms, a draft figure pasted straight into a schedule, a return or a client email is output handling.

The human side matters as much. Professional guidance (voluntary framework): NIST AI 600-1 (July 2024), Section 2.7, describes automation bias as "excessive deference to automated systems" and says it "can exacerbate other risks of GAI, such as risks of confabulation." A well-formatted draft invites exactly that deference. The output contract is designed to give the preparer something to check rather than something to admire.

Professional guidance: the IRS Office of Professional Responsibility's Alert Issue Number 2026-19 (June 24, 2026), applying existing Circular 230 provisions, lists treating AI-generated text as drafts among its best practices.

A question to sit with

When a draft comes back wrong, do you know which part of your prompt failed, or do you just ask again?

What goes wrong

Constraints with no way to detect a breach. "Do not invent figures" is in the prompt, but nothing in the output would show if the model did. A constraint without a check is a hope.

Tie-outs the model performs and nobody repeats. The prompt asks the model to confirm the schedule totals to the control figure, and the model says it does. That confirmation is more output. The tie-out is done by the preparer, with the spreadsheet or the ledger.

Rewriting the whole prompt after each failure. Three things change at once, the draft changes in five ways, and nobody knows which edit helped or what new problem arrived with it.

Patching a wrong header in the same conversation. The earlier, wrong entity or year stays in the conversation, and later drafts can still draw on it.

Regenerating until the draft looks right. Asking again and again, then keeping the version you like, is choosing among unverified drafts, not refining.

Refining past what the tool can do. In the Dell'Acqua experiment (Organization Science, March 2026), consultants using AI on a task selected to be outside the capability frontier "were 19% less likely to produce correct solutions compared with those without AI." The subjects were consultants using a 2023 model, so the lesson carries only by analogy: more revisions do not move a task inside the frontier.

The practical answer

Write the output contract as six parts.

Part What it fixes Accounting example
1. Format The exact shape of the output "Table: item, packet ID, amount as supplied, explanation, flag"
2. Allowed sources Where content may come from "Only packets A1-A3 and F1-F2"
3. Flags What the model writes instead of guessing [MISSING], [AUTHORITY NOT SUPPLIED], [HEADER CONFLICT], [UNSUPPORTED]
4. Prohibited content What must never appear "No rate, threshold or figure not in a packet; no conclusion on a tax position; no promise of an outcome or date"
5. Required tie-outs Totals the output must restate so the preparer can compare them "Restate the F1 control total and the number of items listed"
6. Draft marker and review list Keeps the output a draft and points the reviewer "Head it DRAFT FOR REVIEW; end with a list of every flag and every figure"

Design tie-outs the preparer can perform outside the model. Suppose a fictional accrual schedule, F1, lists six items with a control total of $48,300, agreed to the ledger. The output contract asks the model to restate the item count and the control total, and to list each amount as supplied. The preparer then ties out with a spreadsheet, not by reading.

Tie-out Expected from F1 Draft says Preparer's check
Item count 6 6 Agreed
Sum of listed amounts $12,500 + $8,750 + $9,900 + $6,200 + $4,450 + $6,500 = $48,300 Listed amounts sum to $48,030 Difference of $270; the $4,450 item was restated as $4,180; corrected from F1
Restated control total $48,300 $48,300 Agreed, but does not match the listed amounts, which is why both are required

Notice what happened. The model restated the control total correctly while listing an amount wrongly. Had the contract asked only for the total, the error would have passed. Requiring both, and summing the list outside the model, caught it.

Refine by diagnosis, one layer at a time.

Symptom in the draft Layer most likely at fault One change to make
Wrong entity, year or basis Engagement header Correct the header and start a fresh session
Figure not in any packet Standing instructions or supplied material Check the figure is truly absent; restate the no-new-figures rule; confirm [MISSING] is required
Authority cited that you did not supply Output contract, allowed sources Tighten allowed sources; confirm [AUTHORITY NOT SUPPLIED] is required
Right content, unusable shape Output contract, format Give one sample row of the format
A reason for a variance not in the notes Supplied material Add the client's notes, or require "unexplained" where no note exists
Task drifts into a conclusion Task block Narrow the task to preparation, and name the decision as the reviewer's

Record each revision as a version: v1, v2, v3, with a one-line note of the change. Keep the prompt that worked, and send the change to the library owner if the template itself was at fault (Module 3).

Remember this

  • An output contract fixes format, allowed sources, flags, prohibited content, required tie-outs and a review list.
  • Tie-outs are performed by the preparer with the spreadsheet or ledger; the model's own confirmation is only more output.
  • Require both a restated total and the listed amounts, so that an error in either shows.
  • When a draft fails, diagnose the layer, change one thing, record the version, and start fresh when the header changes.
  • After three revisions that do not pass, stop and triage again.

Accounting example

Abbott Finch LLP, a fictional firm, prepares month-end accrual narratives for the fictional client Sorrel Hotels Inc. Staff accountant Jonah Price uses the library prompt ACCRUAL-NARR v2. Its output contract requires a table of items with packet IDs and amounts as supplied, the restated item count and control total, the flags, and a closing list for the reviewer. F1, the accrual schedule, has six items totalling $48,300, agreed to the ledger. Draft v1 comes back with a count of 6 and a control total of $48,300. Jonah pastes the listed amounts into a spreadsheet: they sum to $48,030. The utilities accrual of $4,450 has been restated as $4,180. He corrects it from F1 in the working copy and notes the error. The draft also explains the laundry accrual as "reflecting higher occupancy", but the controller's notes, C1, say nothing about occupancy. Jonah diagnoses the layer: the prompt does not require "unexplained" when no note exists. He makes one change, adding that rule to the task block, records v2-local with a one-line note, and reruns. Draft v2 marks the laundry accrual [UNSUPPORTED] and lists it in the reviewer's flags. Jonah emails the controller, who replies that the increase reflects a new linen contract; he adds the email as C2 and reruns the single section. Manager Priya Desai, a licensed CPA, reviews by reading the flags list, repeating the spreadsheet sum herself and tracing the linen contract explanation to C2. She approves the narrative and asks the library owner to add the "unexplained" rule to the template, citing Jonah's version note.

Responsible-use warning

Do not accept a model's statement that its output ties out. A draft can restate the correct control total while listing an amount wrongly, and the model's claim that the figures agree is just more generated text. Perform every tie-out yourself with the spreadsheet, the ledger or the tax software. And do not refine by asking again until the draft looks right: diagnose the layer, change one thing, record the version, and stop when three revisions have not produced a draft that passes.

Human-review requirement

The reviewer, a qualified professional, starts with the draft's flags list, then repeats the required tie-outs independently, then traces every figure and citation to its packet. They check the version notes to see what was changed and why, and confirm no revision loosened a prohibition. The owner of the output, the licensed CPA, signing preparer or responsible controller as the work requires, decides whether each flag has been resolved from a source before release. A clean flags list is not taken as proof that nothing is missing.

Practical exercise

Use fictional figures only.

  1. Write a fictional schedule of five items with amounts of your choosing, and calculate its control total with a spreadsheet. Write down the total and keep the spreadsheet as your figure packet.
  2. Write an output contract for a narrative on the schedule with all six parts from this lesson, including at least three flags and two prohibited kinds of content.
  3. Write a fictional draft that restates the control total correctly but lists one amount wrongly, as a model might.
  4. Perform the tie-outs outside the model and record the expected figure, the draft figure and the difference for each, in a table like the one in this lesson.
  5. Add to the draft one explanation that no supplied note supports, and write the flag it should have carried.
  6. Using the diagnosis table, name the layer at fault for each of the two errors and write the single change you would make for each.
  7. Record the revisions as versions with one-line notes, and say which change, if any, belongs in the library template rather than only in your working copy.
  8. Write a short note to a reviewer explaining where to start and what the tie-outs showed.
  9. Decide what you would do if a third revision still failed its tie-outs, and write down the reason under the stop rule and who you would tell.

Knowledge check

  1. What are the six parts of an output contract?

    Answer guidance: Format, allowed sources, flags, prohibited content, required tie-outs, and a draft marker with a review list.

  2. Why should an output contract require both a restated control total and the listed amounts?

    Answer guidance: Because a model can restate the total correctly while listing an item wrongly. Summing the listed amounts outside the model and comparing with the control total catches errors that either figure alone would hide.

  3. What does OWASP's 2026 list say about misinformation that is relevant to accounting outputs?

    Answer guidance: The project leads say that when a model's fluent, confident output drives a decision or a tool call, a wrong answer turns into a wrong action. A draft figure copied into a schedule, return or email is such a step, so it must be checked first. OWASP's list is voluntary.

  4. What should you do when a draft uses the wrong tax year?

    Answer guidance: Diagnose it as an engagement header failure, correct the header and start a fresh session, rather than patching it in the same conversation where the wrong year remains in the context. Record the revision as a new version.

Sources and further reading