,

Evaluating AI in Audit Software: What Is Real and What Is Agent-Washing

Every audit management vendor now says something about AI. Optro (formerly AuditBoard), TeamMate, Diligent One Platform, Workiva, Onspring, LogicGate, IBM OpenPages, MetricStream, Archer, ServiceNow, SAI360, Protecht, DataSnipper and MindBridge have all shipped a named AI feature since the start of 2025, and most now call at least one of them an “agent.” Two analyst firms have looked past the marketing at the same time, in the same quarter, and reached the same conclusion. Gartner’s Market Guide for Audit Management Software, published 13 April 2026, tells buyers to be “particularly wary of agent-washing.” Forrester’s Wave for GRC Platforms, Q2 2026, is blunter: its own reference customers told Forrester that AI “is providing minimal value for customers today,” and that vendor pricing for it is confusing enough that “limited consensus exists about how to price AI, making comparison hard.” Those two findings, from the two firms every audit software shortlist eventually cites, are the reason this guide exists.

This guide does not review one product. It gives a buyer what they need to test any vendor’s AI claim before it reaches a contract: a vendor-by-vendor table of what has actually shipped, with launch dates, named models and data-use statements where vendors publish them; twelve questions worth putting in writing before a demo; a five-task script that separates a real feature from a wrapped chatbot; and an AI usage policy outline to have ready before the first vendor call. The evidence is the same base this site uses everywhere: vendor documentation and release notes, the individual reviews linked throughout, and the two analyst reports above. The site’s guide to types of internal audit software covers the platforms themselves; this page is only about testing what each one says its AI does.

How to read this guide

What this guide is for. A method for separating a real AI feature from marketing language, before a demo and before a contract: what fourteen audit and audit-adjacent products have actually shipped and disclosed, the questions to put to any vendor in writing, a demo script that forces a live answer, a usage policy to have ready first, and a worked example applying all three. It does not replace the product-specific detail in each platform’s own review.

Evidence. Research-based: vendor documentation, release notes and security or trust pages, analyst coverage from Gartner and Forrester, and the verified facts used across this site’s own reviews. The site has not used any of these products hands-on for this guide.

Last verified. 27 September 2026.

In this guide

Agent-washing has a name now

Agent-washing is Gartner’s word, not this site’s, and it names something specific: renaming ordinary automation, a single wrapped call to a large language model, or a chatbot bolted onto a search box, as an “agent” or “agentic AI” feature, because the term tests well in a demo and in a renewal conversation. Gartner’s Market Guide for Audit Management Software, published 13 April 2026 and written by analyst James Bourke, tells audit functions to scrutinize vendor claims about AI capabilities and be “particularly wary of agent-washing” before buying on the strength of a features slide. The report does not name which vendors it means; the table further down this page is this site’s attempt at the naming work Gartner leaves to the buyer.

Forrester’s Wave for GRC Platforms, Q2 2026, tested the same claims from the buyer’s side and found less to show for them than the marketing implies. Its own language is direct: AI “is providing minimal value for customers today but must change quickly,” and “much of the current AI functionality boosts current capabilities rather than the promised transformational change.” On price, Forrester found the same confusion this guide’s own pricing section documents case by case below: “limited consensus exists about how to price AI, making comparison hard,” and Forrester’s reference customers reported being confused by vendors’ pricing approaches, repeatedly citing a lack of clarity over what an AI add-on was actually worth. Two independent analyst firms, working from different methods, arrived at the same warning in the same quarter. That is unusual enough for a buyer to take seriously rather than file away as routine analyst caution.

What internal audit teams are actually using: the survey evidence

Adoption and depth are different questions, and the survey data answers them differently. The IIA’s 2025 North American Pulse of Internal Audit, a survey of 405 audit leaders, found that 41 percent of functions already use generative AI in their work. Wolters Kluwer’s own 2025 survey of 4,214 internal audit professionals, fielded 30 April 2025, found a close current figure, 39 percent, with 41 percent more planning to adopt AI within the next twelve months, a combination the survey itself describes as adoption roughly doubling by 2026. The IIA’s separate Vision 2035 survey, with 6,506 respondents, is the number in this set that should give a buyer pause: audit management software itself is implemented at a high level by 40 percent of functions, but AI is used at an advanced level by only 7 percent. Most of what functions call “using AI” today is closer to the first rung of a long ladder than to the agentic workflows a vendor describes on its homepage.

SurveyWhat it measuredFinding
IIA 2025 North American Pulse of Internal Audit (405 respondents)Current use of generative AI in audit work41 percent already use it
Wolters Kluwer 2025 survey (4,214 respondents, fielded 30 Apr 2025)Current and planned AI adoption39 percent already use it; 41 percent more plan to within 12 months
IIA Vision 2035 (6,506 respondents)Depth of technology implementationAudit management software at a high implementation level: 40 percent. AI at an advanced level: 7 percent
Gartner press release, 16 Sep 2026Audit management software adoption, 259 departments surveyed84 percent of audit functions have adopted audit management software

What each platform ships, and what it discloses

The table below covers every product this guide was asked to cover: the eleven platforms and suites reviewed elsewhere on this site, plus three specialist tools bought alongside them. Each row states what is named and dated in the vendor’s own documentation, whether a specific AI model or provider is disclosed, and what the vendor’s data-use statement says, or that none was found. Full detail on each product beyond what fits in one row is in its own review, linked from the product name.

ProductNamed AI feature(s) and launch dateModel or provider disclosedData-use statement
Optro (formerly AuditBoard)Accelerate: Audit Agent and Document Intelligence (22 Oct 2025); Midship, a SOX-automation AI company, acquired 6 May 2026In-house machine learning plus Azure OpenAI; no single foundation model namedOpt-in; no customer data used for training; human validation required; activity logged
TeamMateTeamMate+ AI Editor (launched 2 Jun 2025); Expert AINot namedData isolation, no data retention, end-to-end encryption, a PII blocker; Expert AI content is not used to train models
Diligent One PlatformACL AI Studio (general availability 25 Sep 2025; agentic mode 2 Feb 2026); AuditAI (10 Mar 2026); orchestration, risk and internal-audit agents (15 Sep 2026)ACL AI Studio: running the Claude Sonnet 5 model in most regions as of 5 Nov 2026 (Claude 4.6 from 17 Sep 2026), per Diligent’s own release notes. The wider trust page names only Amazon Bedrock generallyNo training on customer data by default; output labeled as AI-generated
WorkivaFlowchart visualizer, evidence analyzer and control creator (Winter 2026); audit report and remediation agents; Tie-Out, Benchmarking and Sustainability agents plus an MCP gateway (29 Jul 2026, advanced tiers)Providers named on the Workiva AI page (Google, AWS/Anthropic and Microsoft/OpenAI); no specific model namedCustomer data is not used to train large language models; providers do not retain prompt data
OnspringOnspring AI (14 Oct 2025); Agentic GRC (27 Jul 2026), gated to the Silver service tier or aboveAnthropic’s Claude models, named on both releasesNot stated in any page reviewed for this guide
LogicGate Risk CloudAI Skills (included by default, opt-in); Spark AI Reporting Insights and Automated Evidence Testing; Config Newton; GRC Agents (general availability 1 Sep 2026)Named OpenAI models as of November 2026: gpt-4.1-mini, gpt-5-mini, o4-mini, whisper-1, text-embedding-3-small; a bring-your-own-key optionOpt-in for the OpenAI-processing step; not used to train models; 30-day retention under LogicGate’s own key
IBM OpenPagesOpenPages MCP Server and watsonx Orchestrate chat (9.2, 23 Mar 2026); eight AI model configurations including audit-plan drafting (9.2.1, 23 Jul 2026)No single model named: the AI configurations run on a supported provider the customer supplies credentials for, and the 9.2.1 announcement says multi-file AI analysis is available with Gemini and OpenAI modelsNo statement found on training, retention or opt-out for any AI feature
MetricStreamModel Gateway and an AI Governance and Trust Framework (7 Apr 2026); MetricStream Assistant and Policy Assistant (May 2026)Bring-your-own-model: the Model Gateway connects approved providers, including OpenAI, Anthropic, AWS Bedrock and Azure OpenAI, under the customer’s own agreements and API keys; no default model namedNo statement found
ArcherEvolv AI Compliance (15 Sep 2026); Evolv Foundation and Evolv Workplace (14 Sep 2026); earlier Archer Assurance AI and AI Governance (18 Sep 2024) and Evolv for Compliance (4 Feb 2025)Evolv AI Compliance: Amazon Bedrock Guardrails, run inside the customer’s own AWS account. Earlier features: unnamed, and Archer’s own pages disagree on the model countEvolv AI Compliance: prompts and model weights isolated from Archer in the customer’s account; no statement found for the earlier features
ServiceNow IRMNow Assist for IRM, being renamed Otto for Integrated Risk Management (both names live as of Sep 2026); On-Demand Rationalization (general availability); Proactive Clustering (early access)Not namedDocumentation says inputs, outputs and edits are collected and used to develop and improve ServiceNow models and AI products; customers can opt out of future collection
SAI360AI embedded across modules in GRC Elevate 6.0 (21 May 2026); the Essentials edition excludes the AI Chat AssistantNot namedNo statement found
ProtechtCognita AI assistant (announced 8 Sep 2025; full release Nov 2025)Not namedSingle-tenant AWS; no cross-customer sharing; never used to train third-party models
DataSnipperFour prebuilt SOX and controls-testing agents; Alwin, an agentic automation platform (22 Sep 2026), gated to the Elevate tierExcel and Disclosure Agents run on Azure OpenAI, via a Microsoft partnership dated 1 Oct 2025States it does not train on customer data
MindBridgeAgentic Risk Assessment and an MCP server (announced 22 Sep 2026); an Analysis Designer Agent in previewNot namedNo statement found

Reading the pattern

Four products in the table name an actual foundation model in writing: Diligent One’s ACL AI Studio on Claude Sonnet 5 as of November 2026, Onspring’s Onspring AI and Agentic GRC on Claude, DataSnipper’s Excel and Disclosure Agents on Azure OpenAI, and LogicGate, which goes further and names five specific OpenAI models by their model identifiers as of November 2026. Three of those four, Diligent, DataSnipper and LogicGate, also state, in writing, that customer data is not used to train the underlying model; the Onspring pages this site read name the model but state no training policy. That is not a coincidence: a vendor confident enough to name its model provider is usually also confident enough to state its data-use terms, because both statements can be checked against the model provider’s own published terms. The reverse holds too. TeamMate, Workiva, IBM OpenPages, MetricStream, Archer’s earlier Evolv Foundation features, ServiceNow’s Now Assist and Otto, SAI360, Protecht and MindBridge all describe an AI feature without naming a specific model, and four of those nine, IBM, MetricStream, SAI360 and MindBridge, have published no data-use statement for any AI feature this site could find. A buyer cannot verify what is not written down.

Watch for a second, quieter pattern: numbers that do not agree with each other on the vendor’s own site. Archer’s Evolv Foundation pages state both “492 purpose-built models” and “526 purpose-built GRC models” without reconciling the two. MindBridge’s Ensemble AI documentation states its analysis covers “more than 40” control points on one page and 28 on another. Optro’s Midship acquisition press release claims the AI can automate “up to 87 percent of SOX program management,” with no methodology published for how that figure was measured. Archer’s own 30 June 2026 release headlines “95% Verified Accuracy, 80x Faster, 92% Lower Cost” from Archer’s own test of 55 regulatory documents against an unnamed general-purpose model; the case study it links describes the test design and names no outside validator. None of these numbers is necessarily false. All of them are unverifiable as published, and a vendor’s own inability to keep its marketing pages consistent with each other is itself useful information about how carefully the claims underneath were checked before they went live.

The twelve questions to put in writing

Ask these before the first demo, not after. A vendor that answers in writing and specifically is telling you something different from one that answers “let me follow up,” and the difference is usually visible within a day.

The AI evaluation checklist

Data use and opt-in. Is the feature opt-in or on by default, and who in our organization can turn it off? Is our data used to train the model, for us or for any other customer, and where is that written down?

Model provider. Which named model and provider actually power this feature, not “our proprietary AI”? If the provider or the model changes, will we be told before it happens, or only after?

Retention. How long are our prompts and any uploaded documents retained, by the vendor and by its model provider? What happens to that data on termination, and how long do we have to request deletion?

Output checkability. Can every AI output be traced back to its source document inside the product itself, not just described in a marketing page? What is the vendor’s own published accuracy or error-rate figure, and what methodology produced it?

Audit trail of AI edits. Does the system log every AI-suggested edit separately from a human edit, showing who accepted, rejected or changed it and when? Can that log be produced for an external auditor or examiner in a format they can actually use?

Pricing. Is the feature included in the base subscription, a paid add-on, or gated to a specific tier, and what is the exact incremental cost in writing, not a verbal estimate?

Evaluation evidence. What independent evidence exists beyond the vendor’s own press release: a named analyst report, a customer reference we can call ourselves, or a live demo we control rather than the vendor’s rehearsed sample?

The five-task demo script

A vendor-run demo is built to show its best case, using data the vendor chose. These five tasks are built to show the product’s actual case: ask for each one live, in the vendor’s own environment, using a document or dataset your own team brings rather than the vendor’s rehearsed sample.

TaskWhat to ask for, liveWhat a genuine feature showsRed flag
Draft a risk statement from a real controlPaste in one of your own control descriptions and ask for the linked risk statementA draft naming the specific risk in the control’s own language, with a visible link back to the source textA generic risk statement that would fit almost any control
Summarize a real workpaper or evidence fileUpload a redacted workpaper and ask for a summary with a completeness callA summary that flags what is missing or unclear, not only what is presentA confident summary that never flags a gap
Flag anomalies in a sample of journal entriesBring an anonymized extract of your own ledger data and ask the tool to flag unusual entriesSpecific entries flagged with a stated reason, and a way to see why each one was scoredA vague risk score with no visible reasoning behind it
Trace an AI-suggested edit to its author and timestampAccept, then reject, an AI suggestion inside a workpaper and ask to see the change logA log entry showing the suggestion, your action and the timestamp, separate from human editsNo visible distinction between an AI edit and a human one
Reproduce the same output twiceRun the same task twice, a day apart, with the same inputThe same or a materially consistent output both times, or a stated reason for any differenceMeaningfully different answers with no explanation offered

What AI actually costs, and why so few vendors will say

Forrester’s confusing-pricing finding above is not abstract. Across the fourteen products in this guide’s table, only a handful state in writing whether the AI feature is included, an add-on, or gated to a tier, and fewer still name a price for it.

ProductWhat this site found on AI pricing and gating
OnspringThe AI add-on requires the Silver service level or above; no separate published price
DataSnipperAlwin is gated to the Elevate tier, which is itself quote-only; no separate AI price found
SAI360The Essentials edition excludes the AI Chat Assistant, Advanced Analytics and SSO; higher editions are not priced publicly
LogicGateAI Skills are included by default with no extra line item stated; the OpenAI-processing step itself is opt-in
Optro, TeamMateNo separate AI fee found for either; AI appears bundled into each vendor’s existing pricing model
IBM OpenPagesThe pricing page lists watsonx.ai, watsonx Assistant and two Watson services as available at additional cost
Workiva, Diligent One, MetricStream, Archer, ServiceNow, Protecht, MindBridgeNo public statement found on whether any AI feature carries a separate charge

Treat every blank cell above as a question for the RFP, not an assumption that the feature is free. The site’s internal audit software pricing guide has the base-subscription evidence for each of these vendors; this table is only about the AI layer sitting on top of it, and that layer is exactly where Forrester found the least agreement on how a vendor should charge at all.

Write the usage policy before the first demo

A usage policy is not a compliance exercise to finish after the contract is signed. Written first, it turns the checklist above into terms with consequences: the vendor either meets the policy, or the feature does not get turned on, regardless of what the sales deck promised.

Internal audit AI usage policy: outline

Approved uses. Name the specific tasks AI may assist with today, such as drafting a first pass at a risk statement, summarizing a long workpaper, or flagging anomalies for a human to review, and state plainly that final judgment, risk rating and sign-off stay with a named auditor.

Data boundaries. State which data classes may and may not be pasted, uploaded or connected to an AI feature, tied to the answers gathered from the data-use and retention questions above, not to a general assumption that the vendor’s usual security terms already cover it.

Human review requirement. Require a named reviewer to accept, reject or edit every AI-generated output before it enters a workpaper, and require the system to record which of the three happened and by whom.

Disclosure in workpapers. Require that any AI-assisted step be marked as such in the workpaper itself, the same way reliance on another party’s work is marked, so a reviewer or external examiner can see it without asking.

Vendor accountability. Require the vendor’s written answers to the twelve questions above as part of the contract file, not a verbal assurance from a sales call, and require advance notice before the underlying model or provider changes.

Review cadence. Set a fixed date, tied to a QAIP cycle, to re-run the demo script above against the live product, since a feature that passed at purchase can change without a new contract ever being signed.

Worked example: MidState Beverage tests two AI claims

MidState Beverage, the fictional three-state drinks distributor with a six-person audit function used for worked examples elsewhere on this site, is down to two finalists on its shortlist and wants an AI feature that can draft first-pass risk statements from its own control descriptions. Both vendors’ sales teams describe their feature as “agentic.” MidState’s audit manager sends the twelve questions above to both, in writing, ahead of any demo.

The first finalist answers within two days. Its feature runs on Anthropic’s Claude, named in writing; it is opt-in per user; and MidState’s data is not used to train the model. What the written answer cannot confirm is retention: how long prompts and uploaded control descriptions are kept, and what happens to them if MidState leaves the platform. That gap, a named model sitting next to an open retention question, is Onspring’s own documented position as of November 2026, with one difference: the Onspring pages this site read state no training policy either. MidState’s manager treats the open question as a signing condition rather than a footnote, and puts it in the contract before anyone signs rather than after.

The second finalist’s sales team cannot name the underlying model at all, describes data handling only as “enterprise-grade security,” and cannot produce a published accuracy figure. Run against the demo script above, its feature reproduces a materially different risk statement from the same control description on two separate days, with no explanation offered for the difference. That combination, no named model, no data-use statement and an unreproducible output, is the agent-washing pattern Gartner and Forrester both warn about, whatever the sales deck calls the feature. MidState’s audit committee gets a one-page memo recommending the first product with the retention answer as a signing condition, and passing on the second until it can answer the same twelve questions in writing.

Where this fits your QAIP and the standards

Evaluating a vendor’s AI feature is a different exercise from using AI in your own audit work, and this guide covers only the first. If your function is also using AI directly, drafting testing procedures, summarizing evidence, or querying a risk register with a prompt, the site’s AI audit framework and prompt library cover that separately, with the same emphasis on checkable output and a named human sign-off. The Global Internal Audit Standards do not name AI specifically, but Principle 12, Enhance Quality, is the natural home for a written AI evaluation and usage protocol: it is exactly the kind of leading practice a quality assurance and improvement program should test for. The site’s QAIP playbook and standards reference map cover where a policy like the one above belongs in that program. Keep the usage policy and each vendor’s written answers to the twelve questions in the same file a QAIP self-assessment would look for, not in a separate drawer only the person who ran the demo can find.

Questions about evaluating AI in audit software

Is any audit software AI actually “agentic” today?

Some of it, by the plain meaning of the word: software that can carry out a multi-step task toward a goal with limited step-by-step prompting. Onspring’s Agentic GRC and Diligent One’s agentic-mode ACL AI Studio both name a model provider and describe multi-step behavior in writing. Most of what is marketed as “agentic” in the table above is closer to a single-shot AI feature, such as a summary or a drafted first pass, wearing the word because it tests well. What launched, when, and with what disclosed, is the more useful question than whether the marketing word is technically defensible.

Which vendor in this guide publishes the most complete AI data-use statement?

On the evidence gathered for this guide, LogicGate’s is the most specific: it names five OpenAI models by identifier as of November 2026, states data is not used for training, and states a 30-day retention period under its own key. Onspring and DataSnipper both name their model provider. DataSnipper states a no-training position, and its security page gives a 24-hour limit for two of its AI features; Onspring publishes neither a training statement nor a retention period this site could find. Four products in the table, IBM OpenPages, MetricStream, SAI360 and MindBridge, had no data-use statement this site could locate for any AI feature at all; ServiceNow’s documentation has one, and it says inputs and outputs are used to improve ServiceNow’s models unless a customer opts out.

Should we let a vendor’s AI draft anything that goes into a workpaper?

Only with the human-review requirement in the usage policy outline above in place, and only where a vendor’s own answers to the retention and output-checkability questions are satisfactory in writing. A drafted risk statement or summary that a named reviewer checks, edits and signs is different in kind from an unreviewed AI output entering the file; due professional care applies to the first exactly as it would to a junior auditor’s first draft.

Does it matter if a vendor names its AI model when we never see the model ourselves?

Yes, because a named model is a claim you can check against a third party. When a vendor states it runs on Claude 4.6 or a specific OpenAI model, that model’s own published data-use and training terms become a second, independent source to compare the vendor’s claims against. When a vendor will not name the model, the vendor’s own statement is the only source there is, which is a weaker position for a buyer before anything else about the feature is even evaluated.

How often should we re-run the demo script against a product we already own?

At minimum once a year, tied to a QAIP cycle, and again whenever the vendor announces a new AI release or a model change. AI features across this program have shipped, changed and been renamed within a single year more than once, Now Assist becoming Otto for Integrated Risk Management being the clearest example, and a feature that passed the demo script at purchase is not guaranteed to still be the same feature eighteen months later.

Is agent-washing unique to audit software?

No. Forrester’s minimal-value and confusing-pricing findings above come from a GRC platforms study, not an audit-specific one, and Gartner’s agent-washing warning describes a pattern analysts are flagging across enterprise software generally in 2026. Audit software is not being singled out; it is simply one category where the buyer, an audit or risk professional, already has both the standing and the professional obligation to ask the harder questions before signing.

internalauditguide.com has no commercial relationship with any vendor named on this page. The site takes no vendor money, runs no affiliate links and accepts no sponsored placements, and no vendor saw this page before publication. Product and company names are the trademarks of their owners. Corrections: desk@internalauditguide.com.

Sources and verification

Reviewed in November 2026 against the sources below. How guides are sourced, checked and corrected is set out in the editorial standards. Corrections made in this review are recorded in the corrections log.

Free PDF · 15 one-page cheat sheets

The Internal Audit Desk Reference, free by email.

GIAS 2024 on one page, the sample-size tables, the risk and control matrix, the five-C finding, the SOX deficiency ladder, the CIA exam at a glance, and nine more. Printable. Written by working auditors.

Free. Confirm your address once; the PDF link follows. Unsubscribe anytime.

Cover of The Internal Audit Desk Reference: fifteen one-page cheat sheets and templates (PDF)
Cover of The Internal Audit Desk Reference: fifteen one-page cheat sheets and templates (PDF)

Free PDF

The desk reference: 15 cheat sheets, one page each.

GIAS map, sample sizes, the RCM, the five-C finding, SOX deficiency ladder, CIA at a glance. Free, printable.

Confirm once; the PDF link follows. Unsubscribe anytime.

Comments

Leave a Reply