Do not begin by asking whether a report is biased. That question usually leads to one of two unhelpful responses: accept the document because it looks serious, or reject it because the author has something to sell.

Begin with a more practical question:

Where does the evidence stop and the sales argument begin?

That question matters because much of the useful information reaching decision-makers about artificial intelligence comes from organisations with commercial interests. Vendors publish surveys. Consultancies publish case studies. Investors publish market maps. Platforms publish benchmarks. Analysts define categories that help buyers make sense of a difficult market, while also influencing which suppliers make the shortlist.

Some of this material is excellent. Some is thin. Most sits somewhere between the two.

The mature response is neither belief nor dismissal. It is disciplined extraction.

The five-minute commercial source check

Before a report shapes a budget, shortlist, strategy or board discussion, ask five questions.

1. Who produced or paid for it, and what outcome would suit them?

Identify the author, funder and route to market. Then ask what they would like the reader to believe, buy, commission or prioritise. Commercial interest is not a disqualification. It is metadata.

2. What was actually measured?

Look for the sample, recruitment method, dates, geography, question wording, denominators and definitions. A large number can still answer a narrow or poorly defined question. If the method is not visible, reduce the authority you give the finding.

3. Which statements are findings, interpretation and category construction?

A survey result is one layer. The author's explanation of that result is another. A newly drawn market distinction, such as safe versus unsafe, public versus applied, or experimental versus enterprise-grade, is another again. Do not let the layers merge simply because they appear on consecutive pages.

4. What happens when you leave the report?

Look sideways. Find the original source, search for the organisation and its commercial position, and check what independent or differently interested sources say. Repetition is not independent confirmation if every article traces back to the same survey or press release.

5. What would still need to be true before you acted?

Turn the report's conclusion into an evidence request. Ask what operational results, customer conditions, failure data, comparison group, product validation or contractual protection would be needed before the claim should influence a purchase or policy.

You do not need a complete investigation every time. One unanswered question may be enough to move the report from "decision evidence" to "useful signal", or from "act" to "verify".

A useful report with a commercial destination

Dext's 2026 report on artificial intelligence in accounting and bookkeeping is a good worked example because it is neither worthless nor neutral.

The report states that it combines Dext's internal research among 292 participants in North America, the United Kingdom and Australia with research conducted by Censuswide in December 2025 among 1,500 accountants and bookkeepers in the United Kingdom, Canada and France. It reports substantial concern about time spent correcting artificial intelligence errors, increased error frequency and financial loss. It also says that nearly a third of respondents encounter client mistakes caused by misleading artificial intelligence financial or tax advice each week. 1

Those findings are worth noticing. They suggest that unsupervised use of general-purpose systems may create a downstream verification and remediation burden for professional firms. They also reveal a buyer concern that suppliers of accounting technology would be foolish to ignore.

But the report itself does not publish the questionnaire, recruitment details, regional breakdowns, weighting, margins of error or full definitions behind its headline percentages. That does not make the findings false. It limits how precisely they can be interpreted and how confidently they should be generalised.

The document then moves through a well-constructed commercial sequence.

First, it establishes danger: errors, wasted time, financial loss, liability and weakened trust.

Second, it defines the market through a binary distinction between "public AI" and "applied AI". Public artificial intelligence is presented as broad, uncontrolled and prone to confident error. Applied artificial intelligence is presented as purpose-built, embedded in workflows and better able to deliver accuracy, efficiency and control. 1

Third, it offers sensible trust criteria. Its scorecard asks about data control, domain accuracy, auditability, workflow control and ownership when something goes wrong. These are useful procurement questions, regardless of whether a reader buys Dext. 1

Fourth, the argument narrows towards specialist tools and Dext's own product. The final page cites the volume of documents processed, the size of the development team and the number of accounting documents used for training, then claims "unparalleled accuracy" and "completely trustworthy" workflows before inviting the reader to book a discovery call. 1

This is not evidence that Dext is behaving improperly. It is evidence that the report has a commercial destination.

Competent marketing often works by turning uncertainty into a category and then occupying the reassuring side of that category. The reader's job is to notice each transition and ask whether the evidence supporting one layer also supports the next.

In the Dext report, evidence of client mistakes does not by itself establish that all general-purpose artificial intelligence is unsuitable for governed financial workflows. A useful trust scorecard does not prove that Dext scores highly against it. Large training and processing numbers do not, without performance evidence and scope conditions, establish complete trustworthiness.

Source health note: Dext commissioned and supplied research while selling products in the category it advocates. Treat the reported findings as useful signals where the method supports them, and treat the wider market distinction and product conclusions as commercial claims requiring separate evidence.

Why professional expertise is not enough

There is no basis here for claiming that accountants and bookkeepers are unusually poor judges of information. Their work depends on evidence, controls, traceability and professional scepticism.

The difficulty is that expertise is often local. Knowing how to evaluate financial records does not automatically provide a method for evaluating a polished technology report in which survey findings, category language, product positioning and a call to action are fused together.

Research into online source evaluation illustrates the problem. Sam Wineburg and Sarah McGrew compared professional fact-checkers with doctoral historians and Stanford undergraduates. The historians and students often remained inside a website, inspecting its design, domain name, internal links and apparent authority. Fact-checkers left the site quickly, opened other sources and reached better-supported conclusions in less time. The researchers called this "lateral reading". 2

The lesson is not that historians lack intelligence. It is that subject expertise and source-evaluation technique are different capabilities.

Finance leaders are also being asked to absorb artificial intelligence information during a period of acknowledged skills pressure. A December 2025 survey by the Chartered Institute of Management Accountants and the American Institute of Certified Public Accountants found that 56 per cent of 1,446 senior finance and accounting respondents identified generative artificial intelligence as the most prominent skills gap. Only 29 per cent described their organisations as well or very well prepared for artificial intelligence, while 33 per cent also identified gaps in communication, influencing and critical thinking. 6

That survey did not test whether finance professionals can distinguish interested from independent sources. It does show why practical, just-in-time help may be more useful than assuming professional scepticism automatically transfers into a fast-moving technology market.

Disclosure alone does not solve the problem

It is tempting to think that a visible declaration of interest settles everything. "Sponsored by" appears on the cover, the reader mentally applies a discount, and the problem is solved.

Behavioural research gives reason to be less comfortable. Daylian Cain, George Loewenstein and Don Moore found that conflict disclosure can fail to protect recipients when they do not discount the advice sufficiently. Their experiments also identified circumstances in which disclosure allowed advisers to give more biased advice through strategic exaggeration or moral licensing. 4

The practical conclusion is not that disclosure is harmful or should be hidden. It is that disclosure is the beginning of evaluation, not the end.

Evidence from medicine provides a stronger, though deliberately cross-domain, warning. A Cochrane review of 75 papers found that industry-sponsored drug and device studies were more likely than non-industry-sponsored studies to report favourable efficacy results and favourable conclusions. It also found less agreement between results and conclusions in industry-sponsored studies, while standard risk-of-bias measures did not fully explain the pattern. 5

That evidence concerns clinical research, not artificial intelligence white papers. It should not be used to accuse Dext or any other technology supplier of manipulating results. It demonstrates a broader point: commercial influence may enter through topic choice, comparison, framing and conclusion, even when a document appears methodologically respectable.

Separate the five layers

A useful way to read a research-shaped sales argument is to separate five layers that the document may present as one continuous story.

Evidence

What was observed or measured? What data, interviews, experiments, customer records or case material actually exist?

Interpretation

What do the authors say the evidence means? Could another reasonable interpretation fit the same findings?

Category

What market distinction has been created? Does it clarify a real difference, or simplify a spectrum into a binary that favours the author's position?

Product fit

What additional evidence shows that the author's service or product occupies the desirable side of the category?

Action

What does the document want the reader to do next: change policy, join a programme, buy software, commission advice, book a call or repeat the category language internally?

Each layer may be legitimate. Each requires its own support.

A report becomes most persuasive when the movement between layers feels effortless. A discerning reader slows it down.

Read sideways before you read deeply

The most efficient improvement is often not to interrogate every sentence in the report. It is to leave the report for a few minutes.

Search the organisation's product, pricing and investor or partner position. Find the original study or technical source. Look for a full methodology, questionnaire or appendix. Trace a striking statistic to its first appearance. Find one source that would have an incentive to disagree. Check whether the strongest claim is supported by operational evidence or merely repeated in promotional material.

This is lateral reading adapted for commercial decision-making.

It can be taught. In one study, 87 undergraduates completed four short online modules on source evaluation embedded in a nutrition course. Only 3 students used lateral reading at pre-test; 67 did so at post-test. Overall source-evaluation scores also improved significantly. The researchers were careful to note that the durability and everyday transfer of the improvement remained unknown. 3

That limitation matters, but so does the result. Readers do not necessarily need a grand media-literacy course. Repeated exposure to a compact method, applied to material they already need to assess, may improve the habit.

What a finance leader should do with the Dext report

Imagine a finance director receives the report before a meeting on artificial intelligence tools for the accounting team.

A weak response is to forward the headline statistics as proof that general-purpose artificial intelligence is dangerous. An equally weak response is to dismiss the report as sales collateral.

A stronger response is to separate what can be used now from what still needs evidence.

The director can retain four useful signals: clients may be creating avoidable clean-up work; verification burden deserves to be measured; data control, auditability and workflow ownership belong in supplier discussions; and specialist accounting tools are competing partly on reassurance and control.

Before those signals become a buying conclusion, the director can ask for the survey questions and denominators, independent evidence on error rates, product validation under comparable workflows, documented failure modes, audit and escalation arrangements, and evidence that the proposed system performs better than a governed general-purpose alternative for the task in question.

The report has then done something useful without being allowed to decide the meeting.

The real skill is disciplined extraction

Truly uninterested sources are scarce. Governments have policy aims. Universities seek funding and reputation. Consultancies want engagements. Vendors want demand. Investors want markets to become legible. Publishers want attention.

The answer is not to pretend that all interests are equivalent. Nor is it to wait for a perfectly neutral source before acting.

The answer is to classify the source, trace the claim, compare outside the document and stop the conclusion at the boundary of the evidence.

Commercial interest does not tell you that a report is wrong. It tells you which questions become compulsory.

What did they measure? What does it support? What interpretation was added? What are they selling? What still needs to be verified before the conclusion should shape a real decision?

The objective is not to find a source with no interests. It is to keep the source's interests from doing your thinking for you.

Executive view

  • A disclosed commercial interest changes the questions a decision-maker must ask; it does not invalidate every finding.
  • Keep measured findings separate from interpretation, market categories, claims of product fit and calls to action.
  • A useful buyer response is to retain the signal while requesting methods, independent comparisons and task-specific product validation.
  • Lateral reading is a compact method for checking source context before a report shapes a decision.

Sources and evidence

Six primary sources support the article. The five-layer reading method and finance-director example are Andrew’s Intel editorial synthesis, not a tested procurement intervention.

Source posture: Dext is a commercially interested vendor source. Its public landing page supports the headline risk figures, category language and sales destination; the approved source document cites further full-report details that are not independently reproduced on that landing page. Treat those details as Dext-reported, not independently established. The disclosure experiments and medical sponsorship review are cross-domain analogies, not findings about AI vendors. The lateral-reading studies concern online source evaluation, not finance procurement. The finance skills survey reports perceived gaps, not source-discrimination ability.

Evidence checked 30 October 2026.

  1. Dext, AI in 2026: Research on AI in the Accounting and Bookkeeping Profession / The Power of Applied AI Report 2026, especially pp. 3, 5-8 and 11-15; public report page published 15 April 2026. Source posture: commercially interested material combining internal and commissioned research. Used as the worked example and as market signal, not as neutral proof of Dext product performance or of a universal public-versus-applied artificial intelligence distinction. Primary source.
  2. Sam Wineburg and Sarah McGrew, "Lateral Reading and the Nature of Expertise: Reading Less and Learning More When Evaluating Digital Information", Teachers College Record, 2019. Context: online credibility evaluation on social and political topics; applied here as a transferable source-evaluation method, not as direct evidence about accountants or white papers. Primary source.
  3. Joel Breakstone et al., "Lateral reading: College students learn to critically evaluate internet sources in an online course", HKS Misinformation Review, 23 February 2021. Context: an 87-student course intervention. The study reports substantial short-term improvement but does not establish durability or routine workplace transfer. Primary source.
  4. Daylian M. Cain, George Loewenstein and Don A. Moore, "When Sunlight Fails to Disinfect: Understanding the Perverse Effects of Disclosing Conflicts of Interest", Journal of Consumer Research, 2011. Context: experimental advisory settings; used to show why disclosure alone may be insufficient, not to estimate behaviour in artificial intelligence purchasing. Primary source.
  5. Andreas Lundh et al., "Industry sponsorship and research outcome", Cochrane Database of Systematic Reviews, 2017. Context: drug and medical-device studies. Included as cross-domain evidence that sponsorship can affect outcomes and conclusions; not direct evidence about the Dext report or technology research generally. Primary source.
  6. Chartered Institute of Management Accountants and American Institute of Certified Public Accountants, Future-Ready Finance: Technology, Productivity, and Skills Survey, reported 17 December 2025. Sample: 1,446 senior finance and accounting leaders and managers. Used to establish reported preparedness and skills pressure, not source-discrimination ability. Primary source.