Practical guide

How to create an AI-assisted control assessment

A control assessment should let a reviewer answer three separate questions: is the control designed to address the stated risk, was it put into operation, and did the selected evidence show it operating during the period tested? An AI-generated narrative cannot answer those questions by itself. The useful deliverable is a traceable test record in which every conclusion is tied to a frozen.

By Two Prune

A control assessment should let a reviewer answer three separate questions: is the control designed to address the stated risk, was it put into operation, and did the selected evidence show it operating during the period tested? An AI-generated narrative cannot answer those questions by itself. The useful deliverable is a traceable test record in which every conclusion is tied to a frozen source, a stated procedure, and a named reviewer. This guide uses the DESIGN method to build that record. It is an educational operating method for an internal assessment, not an audit opinion or professional attestation.

Start with a testable control statement

Write the control as an observable action, not an aspiration. “Payments are appropriately approved” is too vague. A testable version is: “For supplier payments above $10,000, a finance manager who did not prepare the payment approves the payment in the banking system before release.”

Freeze these inputs before using AI:

The U.S. Government Accountability Office describes internal control as a process used by management to help achieve objectives. Its 2025 Green Book distinguishes control design, implementation, and operating effectiveness and emphasizes quality information and evidence. Those concepts provide a useful structure here, even when the organization is not applying the Green Book as its governing standard. GAO Green Book

  • the objective and risk the control is meant to address;
  • the control owner, operator, frequency, threshold, and system;
  • the assessment period and complete population;
  • the required evidence and retention location;
  • the sampling or full-population procedure;
  • the person authorized to evaluate exceptions and accept remediation.

Sources: [1]

Apply the DESIGN method

### D — Define the decision and risk

State what the assessment must decide and what it will not decide. Name the risk in concrete terms: an unauthorized payment could be released, a terminated user could retain access, or an unsupported model output could enter a customer communication.

Record the assessment criterion before examining results. Otherwise a reviewer can unconsciously move the threshold after seeing the evidence. AI can help turn a policy paragraph into proposed criteria, but a control owner must confirm that the wording matches the actual process and authority.

### E — Establish the population and evidence map

Create a source manifest with file name, system, owner, extraction time, period covered, row count, checksum when available, and known exclusions. Reconcile the population to a system total or another independent source.

Map each test attribute to evidence. For the payment control, the attributes might be amount, preparer, approver, approval timestamp, release timestamp, and segregation between preparer and approver. A policy document establishes the rule; it does not prove that a particular transaction followed it.

AI is useful for proposing missing fields, normalizing evidence labels, and flagging inconsistent records. It should not fill missing approvals, infer an identity match from a similar name, or silently treat an absent timestamp as a pass.

### S — Specify and execute the procedure

Write a procedure another reviewer could repeat. For example:

Keep the AI instruction, model or service identifier when policy permits, input manifest, output identifier, and reviewer changes. If the tool extracts fields from screenshots or PDFs, verify the rendered source rather than trusting optical recognition.

### I — Inspect exceptions without collapsing categories

Use at least four results: pass, exception, unable to test, and not applicable. “Unable to test” is not a pass. Separate a control failure from an evidence failure: a missing approval may mean the action did not occur, or it may mean the evidence repository is incomplete. Both matter, but they lead to different remediation.

For every exception, record the affected item, criterion, observed evidence, owner response, root-cause hypothesis, consequence, interim protection, due date, and planned retest. AI may cluster similar exceptions, but a human reviewer should challenge the cluster and retain the item-level record.

### G — Ground the conclusion in results

Conclude separately on design, implementation, and observed operation. A control can be sensibly designed but not implemented, or implemented but unsupported for part of the test period. State the tested population, sample, exceptions, evidence limitations, and period. Avoid extending a sample result to a broader period or business unit without a defensible basis.

The Institute of Internal Auditors’ Global Internal Audit Standards address engagement planning, evidence, findings, conclusions, action plans, communication, and monitoring. They are a useful primary professional reference for review discipline, while the applicable assurance standard and reviewer qualifications must be determined for the real engagement. IIA Global Internal Audit Standards

### N — Name remediation, retest, and residual uncertainty

Do not close an exception when a plan is written. Record the completed action, new evidence, retest procedure, retest date, reviewer, and result. If management accepts residual risk, identify who had that authority, what was accepted, until when, and what monitoring remains.

  • reconcile the payment extract to the period total;
  • identify every payment above the threshold;
  • select the stated sample or test the full population;
  • inspect the approval and release event for each item;
  • compare preparer and approver identities;
  • record the exact evidence anchor and result;
  • route anomalies to the exception register.

Sources: [2]

Copyable control test sheet

| Field | Record | | --- | --- | | Objective and risk | what the control supports and the failure it addresses | | Control statement | actor, action, object, threshold, timing, system | | Period and population | dates, source, count, total, exclusions, reconciliation | | Test procedure | repeatable steps and sample basis | | Expected evidence | exact records and attributes required | | Actual evidence | source anchor, value, timestamp, reviewer | | Result | pass / exception / unable to test / not applicable | | Exception analysis | consequence, cause, affected scope, interim protection | | Action and authority | owner, due date, approver or risk authority | | Retest | procedure, evidence, date, result, remaining uncertainty |

Worked example: duplicate supplier payment prevention

A finance team assesses a control that blocks a supplier invoice when supplier ID, invoice number, and amount duplicate an earlier item. The period contains 8,420 invoices. The team reconciles the extract to the ledger total, then runs a full-population duplicate comparison. AI proposes 31 possible pairs after normalizing punctuation in invoice numbers.

The reviewer does not label all 31 as failures. Twelve are credit-and-rebill pairs, nine use recurring reference numbers allowed by policy, seven are genuine duplicate attempts blocked before posting, and three posted twice. The test sheet links each pair to the invoice image, posting event, rule log, and disposition. The three posted duplicates become exceptions; one also lacks a usable rule-log record and is marked with both an operating exception and an evidence limitation.

The owner changes the matching rule and documents a temporary daily review. Two weeks later, the reviewer reruns the original three cases plus the new population. The record shows what changed and what the retest observed. It does not claim that a short retest proves future performance.

Acceptance checks

Before the assessment is issued, confirm that:

Two Prune can assess how a participant frames a control problem, navigates evidence, uses AI, verifies results, exercises judgment, and communicates a test record in a realistic work simulation. It does not perform the organization’s control assessment, certify a control environment, or replace the qualified people responsible for the real decision.

  • the population reconciles and exclusions are visible;
  • criteria existed before results were evaluated;
  • every tested item has a source anchor and reviewer result;
  • AI suggestions remain distinguishable from source facts and human judgments;
  • exceptions, evidence gaps, and risk acceptances are not merged into “resolved”;
  • conclusions match the actual period and scope;
  • remediation is closed only after evidence and retest; and
  • the final assessment names its limitations and accountable reviewer.

Sources

  1. 1.GAO Green Book · U.S. Government Accountability Office
  2. 2.IIA Global Internal Audit Standards · www.theiia.org

Explore this topic