Key takeaways
- No standard or regulator page we read says auditors 'accept' AI-prepared reconciliations as such. The sources say that the preparer stays responsible, review must be real and the information used must be reliable.
- The PCAOB staff spotlight says 'an engagement team member who uses a GenAI-enabled tool is still responsible for the results and documentation of the work' and calls itself 'not a rule, policy, or statement of the Board'.
- BDO separates two cases: AI that assists a preparer who then performs 'a meaningful review', and AI that functions as 'part of the control itself', which 'may require additional audit attention'.
- ISA 500 ties the reliability of company-produced information to its source, its nature and 'the controls over its preparation and maintenance'; ISA 520 asks the auditor to investigate fluctuations, and says nothing about who writes the explanation.
- Keep evidence an auditor can follow: source data, a reviewer who can see what the AI did, a log of rule changes and overrides, and a rule for what is never automated.
Auditors: AI-prepared reconciliations are neither accepted nor rejected as such, and no standard or regulator page we read says otherwise. The sources say something narrower: the preparer stays responsible, review has to be real, and company-produced information has to be reliable. This 2027 guide gives five mistakes to avoid, from sources read in November 2026. It is not audit advice.
For the wider finance picture, read our finance IT solutions guide. For the matching side, see our bank reconciliation automation post.
What does "accept" mean when an auditor looks at an AI-prepared reconciliation?
It means the auditor can rely on the evidence, and that depends on controls, not on the tool. CPA Practice Advisor (April 2026) puts three questions: "Can the output be traced back to its source? Was it reviewed and approved by a human? And is there enough documentation to withstand regulatory scrutiny?" It adds that "The answers to those questions depend less on the AI itself and more on how it was deployed and governed."
BDO (August 2026) says the audit approach stays the same in shape: "what is the risk, and what is the control that addresses it?" So the honest answer to "will they accept it?" is "they will test your control over it". Our account reconciliation software vs AI agents post covers the tooling side.
| What they say | Where | |
|---|---|---|
| Who is responsible | The person using the tool 'is still responsible for the results and documentation' | PCAOB staff spotlight |
| Is the review real | 'make sure ... review is more than a formality' | BDO, August 2026 |
| Is the information reliable | Reliability is influenced by source, nature and 'the controls over its preparation and maintenance' | ISA 500 |
| Does anyone say 'accepted' | No standard or regulator page we read does | All sources, November 2026 |
Source: PCAOB, BDO and ISA 500 as read in November 2026. Quotations shortened.
BufferSol Technologies · buffersoltech.com
What do the audit standards say about information a company produces?
ISA 500 covers it, and it does not mention who prepares the information. The 2012 handbook text says "When using information produced by the entity, the auditor shall evaluate whether the information is sufficiently reliable for the auditor's purposes", and that the information "needs to be sufficiently complete and accurate". A reconciliation or a flux schedule is information produced by the entity.
ISA 500 also says reliability "is influenced by its source and its nature, and the circumstances under which it is obtained, including the controls over its preparation and maintenance where relevant", and that "The reliability of audit evidence that is generated internally is increased when the related controls, including those over its preparation and maintenance, imposed by the entity are effective." That is the standard-level reason your controls matter more than the label "AI", and it is why BDO frames the question around the control. We read the extant 2012 text; we did not read any later revised version (see the closing line). Our AI bookkeeping post covers how the ledger numbers get built.
What did the PCAOB staff say about generative AI?
The PCAOB staff spotlight (July 2024) reports outreach, not a rule. It describes itself as "not a rule, policy, or statement of the Board". It says some preparers use GenAI "to assist in the performance of less complex and repetitive processes, such as preparing account reconciliations or to assist with identifying reconciling items", and that preparers emphasized that "reviewing GenAI output continues to be important".
On the audit side, the PCAOB spotlight says firms indicated "an engagement team member who uses a GenAI-enabled tool is still responsible for the results and documentation of the work", and that supervisors "are expected to apply the same level of diligence as when reviewing work where GenAI was not involved." It also says "some firms are designing their GenAI tools to document the relevant underlying source data used". Going Concern covered the same release. Read it as a picture of practice, not a requirement: the PCAOB staff describe what firms and preparers said, and the Going Concern article repeats the reconciliation example.
What changes in 2027 for finance teams?
Three dated points frame 2027. First, the guidance is still thin and still moving. NIST says the AI RMF 1.0 "is being revised as part of the White House AI Action Plan", and its AI 100-1 document says the framework is "intended to be voluntary". Second, audit firms are now writing practical notes for controllers, such as BDO's August 2026 piece.
Third, the standard on audit evidence has been under review for years, and ISA 500 as we read it dates from the 2012 handbook: the IAASB's audit evidence project page records the Board considering "revisions to ISA 500 or other actions", though the page we read describes 2019 meetings and we did not open a final revised standard. BufferSol view: plan the 2027 close for the controls the auditor will test (review, evidence, change control) rather than for any one tool. Our financial close automation post shows where those controls sit, and our AI workflow automation guide covers the wider design.
What is a confabulation, and why does it matter for flux commentary?
It is confident text that is wrong, and a flux paragraph is exactly where it can appear. NIST's generative AI profile (July 2024) defines confabulation as "The production of confidently stated but erroneous or false content", and says "Risks from confabulations may arise when users believe false content – often due to the confident nature of the response – leading users to act upon or promote the false information."
For a controller, the risk is a fluent explanation that names a driver the ledger does not support. ISA 520 asks the auditor to inquire of management and obtain "appropriate audit evidence relevant to management's responses", so an unsupported explanation gets tested. CPA Practice Advisor says AI "can't determine whether management's assumptions are reasonable, conservative, or biased". Auditors also bring professional skepticism, which ISA 200 defines as "An attitude that includes a questioning mind, being alert to conditions which may indicate possible misstatement due to error or fraud, and a critical assessment of audit evidence." It lists being alert to "Information that brings into question the reliability of documents and responses to inquiries", and a fluent explanation with no ledger lines behind it fits that description.
BufferSol view: make the agent cite the ledger lines behind each driver, and refuse to name a driver it cannot cite. Our AI agent evaluation post shows how to test that.
1. Are you asking auditors to "accept" AI instead of showing them the control?
This is the first mistake, and it starts the wrong conversation. No source we read gives auditors a rule that says "AI-prepared is acceptable" or "not acceptable". BDO says "The core principle is similar; the important question is whether the tool is assisting the preparer or functioning as part of the control itself." The conversation to have is about the control.
ISA 500 points the same way: reliability follows controls "over its preparation and maintenance". And the PCAOB spotlight reports that supervisors review GenAI-assisted work with "the same level of diligence". So bring the control, not the label, and keep the CPA Practice Advisor questions in front of you.
Fix it: before year-end, write a one-page description of the reconciliation process: what the AI does, what the person does, what evidence each leaves. Use the BDO scenarios as the vocabulary, and the PCAOB spotlight as a reminder that the person using the tool owns the result. Our AI agent architecture post shows where each step sits in the design. Our AI accounting software post shows the same split for bookkeeping, and our AI for accounting firms post covers the firm-side view.
2. Has the AI become both the preparer and the control?
This is the mistake that changes the audit most. BDO says: "if the organization is relying on the tool to produce a complete and accurate result without that first layer of review, then the tool may become part of the key control environment, requiring more attention to access, change control, and governance over the underlying logic."
BDO's three invoice-to-journal scenarios show where the line falls. In scenario 2, "AI generates a journal entry based on the invoice, the preparer reviews and submits the AI-generated entry, and a reviewer posts it", and "there are still effectively four eyes", so "a traditional audit approach may still be appropriate". In scenario 3, "AI generates and submits a journal entry" and then a reviewer posts it, so "the technology may become more critical to the process and may require additional audit attention, including understanding third-party assurance reporting".
We apply the same test to a reconciliation below, as BufferSol view. Our AP automation guide covers the AP version of the same split.
| Scenario (BDO's wording, adapted by BufferSol) | Who prepares | Who reviews | Audit attention, per BDO |
|---|---|---|---|
| 1. Manual reconciliation | Preparer | Reviewer | "Traditional manual process that auditors have evaluated for years" |
| 2. AI drafts, preparer reviews and submits, reviewer approves | AI drafts, a person owns it | Two people | "A traditional audit approach may still be appropriate" |
| 3. AI prepares and submits, reviewer approves | AI | One person | "May require additional audit attention" |
3. Is your reviewer's sign-off a formality?
A signature without a view of the work is the common failure. BDO says "If AI is used in areas such as journal entries, transaction processing, estimates, or reconciliations, make sure preparer and reviewer responsibilities are clear and that review is more than a formality", and that "An understanding of the full workflow is critical: what the AI is doing, what the reviewer can see, whether data can be changed, and whether the review is substantive or just a sign-off."
"The more automation is involved, the more important it is that human review is real and well understood," BDO adds. A practitioner page on SAP's close agents asks the named reviewer to "inspect support, reperform material calculations, challenge assumptions, resolve contradictions, evaluate unusual entries, and record accept/modify/reject/escalate". That page is a practitioner view, not a standard, but it shows what a real review can look like. Our workflow exception handling post covers the queue design, and our single-agent vs multi-agent post explains why one agent should not both prepare and approve.
Fix it: give the reviewer a screen with the source lines, the AI's matches, the unmatched items and the rule that produced each match. Record who reviewed, when and what they changed, since BDO separates "substantive" review from "just a sign-off". ISA 230 is about the auditor's own file, but it models the habit: it asks for documentation "sufficient to enable an experienced auditor, having no previous connection with the audit, to understand" the work.
4. Can you show the evidence trail behind each reconciliation?
If you cannot trace a number to its source, the audit slows down. CPA Practice Advisor says "A properly configured AI system can log data sources, rule changes, and user overrides automatically", and that this makes it easier for auditors "to trace how a balance was constructed without reconstructing weeks of manual work."
The PCAOB spotlight says some firms design their tools "to document the relevant underlying source data used", so auditors themselves ask for this. And the population matters: ISA 500 says information "needs to be sufficiently complete and accurate". The SAP practitioner page lists the same worry as a failure mode: an incomplete source population that looks fine because every received item was processed, with the control "independent source counts, totals, statuses, cutoff and exclusions".
Fix it: make the agent write, for every reconciliation, the source files, the date range, the line counts, the control totals and the rule version. Our three-way match automation post shows the same audit-log idea for AP.
5. Is AI-written flux commentary investigated or just published?
Flux commentary is where a fluent paragraph can hide a missing check. ISA 520 defines analytical procedures as "evaluations of financial information through analysis of plausible relationships among both financial and non-financial data", which "also encompass such investigation as is necessary of identified fluctuations or relationships that are inconsistent with other relevant information or that differ from expected values by a significant amount."
It then says that when such differences appear, the auditor "shall investigate such differences by: (a) Inquiring of management and obtaining appropriate audit evidence relevant to management's responses". The standard does not say who writes the explanation. CPA Practice Advisor draws the limit: AI "can apply a rule consistently across 100,000 rows, but it can't determine whether management's assumptions are reasonable, conservative, or biased."
A vendor such as Nominal says "Flux analysis is one of the most time-consuming parts of close preparation", which is a vendor's view of the workload, not an audit position. Fix it: let the agent draft the movement and its candidate drivers from the ledger, and require a named owner to confirm the business reason with evidence. Our workflow exception handling post shows the flag-versus-verdict split, and our AI agent vs chatbot post explains why a drafting chatbot is not an agent that acts.
How would an ISA 520 test apply to AI-written flux commentary?
Treat each flux line as a small analytical procedure and test the explanation, not the prose. ISA 520 frames the work as "analysis of plausible relationships among both financial and non-financial data" and investigation of "fluctuations or relationships that are inconsistent with other relevant information". The test below is BufferSol view, built on that wording.
| Step | What to check | Who does it | Source of the idea |
|---|---|---|---|
| 1. Set the expectation | Was the expected value set before the AI wrote anything? | Controller | ISA 520: "expected values" |
| 2. Pick the lines | Which differences are significant, by a written threshold? | Controller | ISA 520: "by a significant amount" |
| 3. Draft the driver | Does each driver cite ledger lines or non-financial data? | AI agent drafts | NIST AI 600-1: confabulation risk |
| 4. Confirm the reason | Does a named owner confirm it with evidence? | Budget owner | ISA 520: "Inquiring of management" |
| 5. Keep the record | Draft, edits, owner and date are stored | Agent logs | ISA 230: who reviewed and when |
| AI agent | Named person | |
|---|---|---|
| Movement | Computes the change and ranks the biggest lines | Confirms the threshold for significance |
| Driver | Proposes candidate drivers and cites ledger lines | Confirms the business reason with evidence |
| Wording | Drafts the explanation | Edits it and owns it |
| Record | Logs sources, draft and rule version | Signs with name and date |
Source: BufferSol view, built on ISA 520 and NIST AI 600-1 wording as read in November 2026.
BufferSol Technologies · buffersoltech.com
Which of BDO's three scenarios is your workflow?
Map each reconciliation and each flux schedule to one of the three scenarios, and write the answer down. The scenarios come from BDO, and the mapping to reconciliations is BufferSol view. If you cannot say who prepared and who reviewed, you are probably in scenario 3 without having chosen it.
| Task | Typical set-up | Scenario | What to evidence |
|---|---|---|---|
| Bank reconciliation | AI matches lines, preparer reviews exceptions and signs | 2 | Source files, match rules, preparer review, reviewer approval |
| Intercompany reconciliation | AI matches and posts differences without a person | 3 | Access, change control, logic, third-party assurance if any |
| Flux commentary | AI drafts, controller edits and owns the explanation | 2 | Draft, edits, owner, supporting evidence |
| Flux commentary | AI drafts and it is published unchanged | 3 | The reviewer's real check, or do not publish |
What evidence do auditors need, and where does each item come from?
Here is the evidence list with the source line behind each item. Where a source is silent, the table says so, rather than stretching a quote. Treat it as a starting checklist, not an audit programme.
| Evidence | Source line | Note |
|---|---|---|
| Output traced to source data | "Can the output be traced back to its source?" (CPA Practice Advisor) | Also "document the relevant underlying source data used" (PCAOB) |
| Named human reviewer | "Was it reviewed and approved by a human?" (CPA Practice Advisor) | Review must be "substantive or just a sign-off" (BDO) |
| Preparer responsibility | "still responsible for the results and documentation" (PCAOB) | Management keeps responsibility (BDO) |
| Accurate and complete inputs | "sufficiently complete and accurate" (ISA 500) | Control totals and counts are BufferSol view |
| Logs of rule changes and overrides | "log data sources, rule changes, and user overrides" (CPA Practice Advisor) | No source we read uses the word "versioned" |
| Change control over the logic | "access, change control, and governance over the underlying logic" (BDO) | A practitioner page asks for a new version and new approval on any change (profession.it.com) |
How do you version the rules, and why does it matter?
Versioned rules are an inference, not a quoted requirement, and we say so. No source we read says "versioned". BDO asks for "change control, and governance over the underlying logic", and CPA Practice Advisor says systems can log "rule changes". Versioning is the practical way to meet both, and it fits the "Govern" and "Manage" functions in NIST AI 100-1.
The SAP practitioner page goes further: "If any field or support changes, preserve the old packet, create a new hash/version, rerun affected deterministic checks, show the difference, and require new approval." As BufferSol view, treat matching rules, tolerances, prompts and model versions the same way: each change gets a version, a date, an owner and a test on last month's data. The NIST AI RMF organizes its core as Govern, Map, Measure and Manage, which is a useful outline for who owns each.
What did the sources not say?
Knowing the gaps keeps you from overclaiming to your auditor or your board. Here is what we looked for and did not find, as of November 2026, on the pages we read.
- No "accepted" rule. The PCAOB spotlight, ISA 500 and BDO give no test that says an AI-prepared reconciliation is accepted or rejected as such.
- No flux guidance. BDO's page does not mention flux analysis, and ISA 520 does not say who writes the explanation.
- No "versioned rules" requirement. The nearest lines are BDO's "change control" and CPA Practice Advisor's logging of "rule changes".
- No final revised ISA 500. We read the 2012 handbook text and the IAASB project page only.
- No rule in the spotlight. The PCAOB staff call it "not a rule, policy, or statement of the Board", and Going Concern reported it as outreach.
Will auditors test the adjusting entries an AI agent proposes?
Yes, in the way they test any adjusting entry. ISA 240 says the auditor "shall design and perform audit procedures to" "Test the appropriateness of journal entries recorded in the general ledger and other adjustments made in the preparation of the financial statements", and to "Select journal entries and other adjustments made at the end of a reporting period". A reconciling adjustment the agent proposes becomes one of those entries once it is posted.
So the risk sits at posting, not at matching. BDO names "automated journal entries" as a low-oversight use case that "deserve[s] caution". BufferSol view: the agent proposes, a named person approves, and the posting is done under that person's name. Our AI bookkeeping post shows the same rule for coding, and our AI invoice processing post shows it for AP postings.
| Output | The agent may | A person must | Why |
|---|---|---|---|
| A matched line | Match it and log the rule | Sample a share of matches | Evidence of review (BDO) |
| A reconciling adjustment | Draft the entry and its support | Approve and post it | End-of-period entries are tested (ISA 240) |
| A flux explanation | Draft with cited ledger lines | Confirm the business reason | Inquiry of management (ISA 520) |
| An estimate or accrual | Gather inputs and show the arithmetic | Own the assumptions | AI "can't determine whether management's assumptions are reasonable" (CPA Practice Advisor) |
What should you not automate?
The sources give cautions, not a closed list, so this is a short list with the line behind each. BDO says "low-oversight use cases tied to financial transactions or reporting deserve caution, particularly automated journal entries, invoice approvals, and accounting estimates." It also warns that "the biggest immediate AI risk is not that the model is wrong; it is that the surrounding control environment is not ready for the speed and reach of AI-enabled tools."
- Sign-off: the PCAOB spotlight reports firms expect GenAI "to augment, but not replace, humans".
- Estimates and judgments: CPA Practice Advisor says AI "can't determine whether management's assumptions are reasonable".
- Anything a reviewer cannot see into: BDO asks "what the reviewer can see, whether data can be changed".
The PCAOB spotlight also notes that "Some firms currently do not allow GenAI to be used when performing audit or attest procedures", so check what your own auditor allows. Our human-in-the-loop guide and touchless invoice processing post cover where approvals go, and our AI workflow automation vs RPA post shows which steps can stay as plain rules.
What does an AI reconciliation agent do, step by step?
Here is the pattern BufferSol builds, with a person at the points auditors care about. The steps are our design, labelled BufferSol view, and the evidence column maps back to the sources above.
- 011 CollectPulls the statement and ledger lines and records files, dates and line counts
- 022 MatchApplies versioned rules and logs which rule matched each line
- 033 ProposeDrafts adjustments and flux drivers; posts nothing
- 044 ReviewA named preparer sees source lines, matches and exceptions, and decides
- 055 ApproveA separate reviewer signs; overrides and changes are logged
- 066 KeepStores the packet: sources, control totals, rule version, reviewers
BufferSol Technologies · buffersoltech.com
Our autonomous workflows service describes how we deliver this, our intelligent invoice processing guide covers the AP side, and the intelligent invoice processing case study shows the same log-everything pattern in AP. For custom builds, see our AI agents service and our AI agent pricing table.
What does a completeness check look like in practice?
It is a handful of totals that anyone can recompute. Here is a worked example, a BufferSol estimate with made-up numbers to show the method, tied to ISA 500's "sufficiently complete and accurate" wording. A bank statement has 1,200 lines. The agent reads 1,200, matches 1,080 and flags 120. The control totals must tie: opening balance plus statement movements equals closing balance, and 1,080 plus 120 equals 1,200.
Then the review: the preparer clears all 120 flagged items and checks a sample of 108 matched lines, 10 percent, chosen by someone other than the agent. If the line count is 1,199, the check fails before anyone reads a match. The SAP practitioner page names the same control: "independent source counts, totals, statuses, cutoff and exclusions". Keep every number in the packet, as ISA 230 models for the auditor's own file. Our cash application automation post shows the same review-queue arithmetic for AR.
Source: BufferSol estimate with made-up numbers to show the method. Completeness wording from ISA 500, November 2026.
BufferSol Technologies · buffersoltech.com
Are vendor claims about audit readiness reliable?
Read them as marketing and check each against a standard. Vendor pages describe audit benefits in general terms. We saw pages that attribute wording to the PCAOB that is not in the PCAOB's spotlight, so we checked every PCAOB quote in this post against the PDF itself. The spotlight says what firms told the staff; it sets no duties.
The practitioner page on SAP's agents is useful because it names controls, but it is not a standard. Nominal sells flux and reconciliation agents, so its audit-readiness advice is a vendor view. A fair test, in the spirit of BDO: can the vendor show the reviewer's screen, the log of rule changes and the source population totals for one real reconciliation? Our AI vendor security questionnaire post and DPA checklist cover the data side, and our AI automation agency pricing post covers what a custom build costs to scope.
Which questions should you ask a reconciliation vendor in writing?
Ask for evidence you can hand to an auditor, not a brochure. These questions come from the lines in this post, and the framing is BufferSol view. Vendors such as Nominal describe flux and bank reconciliation agents, but a product page rarely answers them.
- Show the reviewer's screen: source lines, matches, unmatched items and the rule behind each match, as BDO asks about "what the reviewer can see".
- Explain whether data in a reconciliation can be changed after the agent runs, and by whom.
- Provide the log of rule changes and user overrides, as CPA Practice Advisor describes.
- State how the product proves the source population is complete, with counts and totals, in the spirit of ISA 500.
- Say how a flux driver is sourced and what stops it from citing something the ledger does not hold, given the NIST confabulation definition.
- Offer a SOC 1 or equivalent report if the platform becomes part of the control, since BDO mentions "third-party assurance reporting".
- Describe where your data is stored and whether it trains models, as our AI vendor security questionnaire post lists.
When is a person-prepared reconciliation the better choice?
When the account is small, the judgment is high or the evidence is thin. A person who prepares a rare, one-off reconciliation does not need an agent, and BDO says "Approve use cases before deployment, define ownership, document workflows, restrict access to sensitive data, and keep a human at the helm". An agent earns its place on repetitive, high-volume reconciliations with clear rules.
What an agent does not do: it does not own the account, it does not judge management's assumptions, and it does not replace the reviewer. BDO says "AI can help improve efficiency, but it does not replace management's responsibility for the result." Our how long to build an AI agent post explains what a first slice should contain, and our AI agent development guide covers the build side.
How should you brief your auditor before year-end?
Early, in writing and with a demonstration. This is BufferSol view, not audit advice, and your auditor decides what they need. A verb-first plan:
- Describe the process in one page: what the AI does, what a person does, which scenario it matches.
- Show the reviewer's screen on a real reconciliation from last month.
- Hand over the evidence packet: sources, control totals, rule version, reviewers and dates.
- List every change to rules, prompts and models since the last audit, with owner and test.
- Say which tasks stay manual, such as estimates and sign-off.
- Ask what additional evidence they want, and agree it before fieldwork.
The PCAOB spotlight says some preparers run new technology "in a parallel environment prior to deployment", a habit worth copying: run the agent beside the manual process for one close first. Our AI readiness assessment post helps you pick the first process, our AI automation agency guide covers who can build it, and our free AP automation savings calculator shows how to size review time.
Which related guides should you read next?
Pick the post that matches your next decision. Each is part of the same finance cluster.
- Month-end and close: our financial close automation post and AI accounting software post.
- AR and cash: the accounts receivable automation software post and the AI invoice processing post.
- Bookkeeping and firms: the AI bookkeeping post and the AI for accounting firms post.
- Roles and oversight: the will AI replace accountants post and the human-in-the-loop guide.
- Build or buy: the best accounting automation software post and the AI automation agency guide.
What does a 30-day pilot deliver?
A working reconciliation agent on your real data inside 30 days, which matches the process on our autonomous workflows page. In about two weeks you see a prototype that reads your statements and ledger and drafts matches. Then a live pilot runs on one account or one entity, beside your manual process, with a person reviewing every item and the evidence packet produced each time.
What you provide: a sample of statements and ledger exports, your matching rules and tolerances, and a named preparer and reviewer. How we measure: match rate, exceptions per 100 lines, reviewer minutes, any control total that failed and whether the packet answers your auditor's questions. We do not promise full production for every account in 30 days. Our AI agents service page lists the delivery steps, and our free AI agent ROI calculator helps size the prize first.
Is this right for you?
An AI reconciliation pilot fits if most of these are true.
- You reconcile many accounts each month with repeatable rules.
- A named preparer and a separate reviewer exist for each account.
- You can export source files and ledger lines and keep them.
- Your auditor is open to a walkthrough before year-end.
- You want evidence for each reconciliation, not only a faster close.
If that sounds like you, the financial reporting and forecasting Done For You automation and the month-end reconciliation Done For You automation fits finance teams that want it live in 3 to 6 weeks. Our finance IT solutions guide goes deeper on each step.
Which sources did we read, and what did we not re-check?
Facts checked November 2026 on these pages: the PCAOB staff spotlight, ISA 500, ISA 520 and ISA 230 (2012 handbook text), BDO, CPA Practice Advisor, NIST, NIST AI 100-1, Going Concern, profession.it.com, Nominal and the IAASB project page. BDO, Nominal and the practitioner pages are not standards.
Facts we also read: ISA 200 for professional skepticism, ISA 240 for journal-entry testing and NIST AI 600-1 for confabulation.
Facts we did not re-check: PCAOB AS 1105 and AS 2201 (the PCAOB site returned an access error, so we do not quote them), the final revised ISA 500 (we read only the 2012 text and the project page), and SOX or COSO texts. This post is not audit advice; agree your approach with your own auditor and check current standards.
If you want AI-prepared reconciliations your auditor can follow, ask BufferSol for a scoped estimate for a reconciliation agent with an evidence packet now, so the 2027 budget covers it and the pilot can start in Q1 2027.
Frequently asked questions
Will external auditors accept AI-prepared reconciliations?
No standard or regulator page we read says auditors accept or reject AI-prepared reconciliations as such. BDO says that if a staff accountant uses AI to support a reconciliation and then 'performs a meaningful review, the focus remains on the individual's control activity'. If the tool replaces that review, it 'may become part of the key control environment' and get more audit attention.
Do auditors accept AI-written flux commentary?
The standards we read do not mention flux commentary or its author. ISA 520 defines analytical procedures to include 'investigation as is necessary of identified fluctuations', and requires the auditor to inquire of management about significant differences. So the test is whether management's explanation is supported and investigated, not whether software drafted it.
What evidence do auditors need for an AI-prepared reconciliation?
The sources point to traceability to source data, a reviewer who can see what the AI did, documentation of who reviewed and when, and logs of rule changes and overrides. CPA Practice Advisor asks whether output can be traced to its source and whether a human reviewed and approved it. ISA 500 adds accuracy and completeness of the information used.
Does the PCAOB allow auditors to use generative AI?
The PCAOB staff spotlight reports what firms said, and it is not a rule. It says some firms 'do not allow GenAI to be used when performing audit or attest procedures due to data privacy concerns', while others expect GenAI 'to augment, but not replace, humans'. It also says supervisors apply 'the same level of diligence' as for any other work.
What should a company never automate in the close?
The sources give cautions, not a closed list. BDO says low-oversight use cases tied to transactions or reporting 'deserve caution, particularly automated journal entries, invoice approvals, and accounting estimates'. Sign-off and management's responsibility for the result stay with people, and work a reviewer cannot see into should not be automated.
Should we tell our auditor before using AI in the close?
BufferSol's view is yes, early. Show the workflow, the reviewer's view, the evidence trail and the change log before year-end, not during fieldwork. BDO says 'an understanding of the full workflow is critical'. This is not audit advice; agree the approach with your own auditor and check current standards.
Sources
- Spotlight: Staff Update on Outreach Activities Related to the Integration of Generative Artificial Intelligence in Audits and Financial Reporting, PCAOB
- ISA 500, Audit Evidence (IAASB Handbook), IAASB / IFAC
- ISA 520, Analytical Procedures (IAASB Handbook), IAASB / IFAC
- ISA 200, Overall Objectives of the Independent Auditor (IAASB Handbook), IAASB / IFAC
- ISA 230, Audit Documentation (IAASB Handbook), IAASB / IFAC
- Reviewing AI in Finance Processes: Practical Audit Considerations for Controllers, BDO
- AI is Co-Writing End-of-Month Financial Reporting. Here's What that Means for Auditors, CPA Practice Advisor
- AI Risk Management Framework, NIST
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, NIST
- The PCAOB Asked Big Firms How They're Using Generative AI, Going Concern
- SAP Financial Close Agent Control Workflow, profession.it.com
- Financial Audit with AI: How Modern Accounting Teams Stay Compliant at Scale, Nominal
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, NIST
- ISA 240, The Auditor's Responsibilities Relating to Fraud in an Audit of Financial Statements (IAASB Handbook), IAASB / IFAC
- Audit Evidence project page, IAASB
Cite this article
BufferSol Editorial Team (2026). Auditors, AI-Prepared Reconciliations and Flux: 5 Mistakes to Avoid Before the 2027 Audit. BufferSol Technologies. https://buffersoltech.com/blog/auditors-ai-prepared-reconciliations-2027
BufferSol Editorial Team
AI & software engineering team
Senior engineers, AI specialists and product leads at BufferSol Technologies who design, build and run AI agents, autonomous workflows and custom software for companies in the US, UK and Europe.
Reviewed by Sufyan K., Senior Writer, BufferSol Technologies.
