Can AI Accelerate Research?
Desk research is mostly finding, reading and pulling out the facts that matter, and AI tools can now do much of the finding. The best measured gains come from extraction against a fixed list. The weakest point is the citation itself, which means the time saved on retrieval is partly spent again on checking it, and nobody we found has measured the net.
On this page
Better: Note and confidence
Low to moderate confidenceAccurate extraction; unreliable citationsIn a small peer-reviewed study, a model extracted trial data with 96.3% accuracy. In other independent tests, chatbots invented or misattributed a large share of the references and sources they were asked to supply.Faster: Note and confidence
Low confidenceRetrieval is faster; net time saved is unmeasuredThe only time-saving figure we found for consulting research is McKinsey's own, self-reported claim of up to 30% of search and synthesis time, as the press reported it. We found no independent study timing desk research with the checking counted.Cheaper: Note and confidence
Low confidencePlausible, but nobody we found has measured itNone of the sources we found measured the cost of a research task. Whether faster retrieval lowers cost depends on how much verification the output needs.Safer: Note and confidence
Governance-gatedSafe when a named person checks every sourceThe major evidence synthesis organisations say authors remain responsible and AI should be used with human oversight. An Australian parliamentary committee inquiry published false, AI-generated allegations against KPMG and Deloitte in an academic submission.The problem, in numbers
Much of professional services work starts with desk research: what has been written, what the data say, what the firm already knows. That work is slow because it is thorough. The most carefully measured version is the systematic review in medicine. Across 195 reviews registered on PROSPERO, the mean estimated time to complete and publish one was 67.3 weeks, with a mean of five authors, and searches returned anywhere from 27 to 92,020 studies to sift.Source 1 That is a stricter standard than a client briefing, but it shows where the hours go: finding, screening and extracting.
Firms have built tools aimed at exactly those hours. McKinsey's internal tool, Lilli, searches a store of more than 100,000 documents, case studies and playbooks, as Consultancy.com.au reported at its launch in 2023.Source 2 By 2026 McKinsey said the tool and related systems were saving up to 30% of the time consultants once spent searching and synthesising information, as Entrepreneur reported.Source 2 That is the firm's own, self-reported figure, and we found no independent measurement of it.
What improves, and why
In our reading, desk research has three parts with very different risk. Retrieval is finding candidate material. Extraction is pulling defined facts out of a document already in hand. Synthesis is deciding what it all means. AI helps most where the task can be checked against a fixed target.
Extraction is the clearest case. In a peer-reviewed proof-of-concept study, Claude 2 extracted 160 data elements from 10 open-access trial reports with 96.3% accuracy, and four of its six errors were missed data items.Source 3 The sample was small and chosen for convenience, so it shows what is possible on clean documents, not what a firm should expect on its own files.Source 3
Retrieval is faster but less trustworthy, because the model is now choosing what exists. That is where the time saved can be spent again. Every source a tool returns has to be opened and confirmed before it goes into client work.
Synthesis, deciding what a body of material means for this client, is where we found the least measured evidence of any kind. In our reading, it is also where the question is most specific to the engagement, so it is the part least likely to be done well by a tool trained on general material.
AI makes finding sources cheap. It does not make trusting them cheap, and the checking is the part a client is paying for.
What the evidence actually says
The research on AI in evidence synthesis is growing fast but is still early. A 2025 scoping review of 37 studies found large language models applied to 10 of 13 systematic review steps, most often literature search (41%), study selection (38%) and data extraction (30%).Source 4 Most were validation studies, the authors judged the use promising in about half, and they found few fully validated applications.Source 4 Its last search was in February 2024, so newer tools are not covered.Source 4
The citation is the weak point. In a peer-reviewed comparison across 11 systematic reviews, the share of references counted as hallucinated (two or more of title, first author and year wrong) was 39.6% for GPT-3.5, 28.6% for GPT-4 and 91.4% for Bard.Source 5 Those were mid-2023 model versions, and the paper does not say whether they could search the web.Source 5
Chelli et al., Journal of Medical Internet Research, 2024. References generated for 11 systematic reviews; hallucinated means two or more of title, first author and year were wrong. Model versions from July 2023.
Tools with live search did not avoid the problem. In 2025 the Tow Center at Columbia Journalism Review gave eight AI search tools excerpts from 200 news articles, 1,600 queries in all, and asked each to name the headline, publisher, date and URL. Collectively they answered more than 60% of queries incorrectly, from 37% for Perplexity to 94% for Grok 3, and premium tools answered more prompts correctly than free ones but also had higher error rates, mainly because they gave definitive wrong answers instead of declining.Source 6 In our reading, that is the core risk for desk research: a tool that cites fluently but wrongly costs more checking time than one that says it does not know.
Controlled experiments with consultants, covered in our review of delivery costs, found faster work inside the model's competence and worse work outside it. We found no study that times professional desk research with and without AI and counts the verification.
Where it falls short
The speed figure is self-reported. The 30% saving is McKinsey's own claim, as the press reported it.Source 2
Citations fail often. Independent tests found invented references and wrong attributions at high rates, even with live search.Source 5,Source 6
The best accuracy result is small. The 96.3% extraction figure comes from ten clean, open-access trial reports.Source 3
Few uses are validated. Most published work tests whether a step can be done, not whether a whole review holds up.Source 4
Where governance decides the outcome
Australia has already seen what unchecked AI research costs. In 2023 a group of accounting academics made submissions to a parliamentary joint committee inquiry into the consultancy industry that contained false case studies, which one of them later blamed on Google's Bard. As the press reported, they wrongly linked KPMG to the 7-Eleven wage theft scandal and accused Deloitte of fudging the accounts of Patisserie Valerie, a company Grant Thornton audited, and the material was published under parliamentary privilege. Emeritus Professor James Guthrie took responsibility and apologised to the firms and the committee, KPMG's chief executive said AI had been relied upon "without comprehensive fact-checking", and the committee said submitters using AI should "ensure the output is truthful".Source 8 The Deloitte report for a federal department, covered in our review of delivery costs, shows the same failure on the firm side. The four major evidence synthesis organisations, including Cochrane, say authors are ultimately responsible for their synthesis, that AI "should be used with human oversight", and that any AI use that makes or suggests judgements should be fully reported, with the tool's name, version and date.Source 7 In our reading, a firm should apply the same standard to client research, and keep its data handling in line with our review of client confidentiality.
How to think about adoption
Use AI first for extraction from documents you already hold and trust: tabulating figures from reports, pulling clauses or data points against a fixed list, summarising a known set of sources. Treat retrieval as a lead list, not a bibliography. Every source that reaches a client should be opened, confirmed and recorded by a named person, along with the tool and version used. Pilot on real engagements and time the whole job, checking included, before claiming a saving. In our reading, retrieval does compress desk research, but the skill a client pays for moves to verification, and firms that price and staff for it will keep the speed without the corrections.
BFCS.ai does not fabricate figures. The independent evidence on research speed in professional services is thin. The only consulting time-saving figure we found is a firm's own claim, reported by the press. The accuracy studies come from medical evidence synthesis and journalism, tested mostly on 2023 and 2024 models, and are used here as the nearest measured settings. We found no study, Australian or otherwise, that times professional desk research with and without AI once the checking is counted.
Sources
- 1
Borah, R., Brown, A. W., Capers, P. L. & Kaiser, K. A., "Analysis of the time and workers needed to conduct systematic reviews of medical interventions using data from the PROSPERO registry," BMJ Open 7(2), e012545, 27 February 2017, doi:10.1136/bmjopen-2016-012545 (peer-reviewed; abstract read via PubMed, the publisher page refused automated access). pubmed.ncbi.nlm.nih.gov
Back to textWhat we took from it
195 registered and completed reviews with associated publications, as of 1 July 2014; "The mean estimated time to complete the project and publish the review was 67.3 weeks" (IQR 42); mean of 5 authors per review (SD 3); searches returned between 27 and 92,020 studies, with a mean yield of included studies of 2.94%. Medical evidence synthesis, not professional services.
- 2
Two press reports of McKinsey's own statements about Lilli, its internal generative AI tool (self-reported, not independently measured). consultancy.com.au
Shibu, S., "McKinsey Consultants Are Letting New Technology Take Over an Essential Part of Their Work," Entrepreneur, 11 June 2026: entrepreneur.com
Back to textWhat we took from it
Consultancy.com.au, "McKinsey unleashes internal generative AI tool for employees," 23 August 2023: Lilli lets users scan the firm's intellectual property archive, with repositories of more than 100,000 documents, case studies and playbooks across 40 curated knowledge systems; it was in beta for about 7,000 employees; a McKinsey senior partner told VentureBeat it provides attribution and links to its sources.
"McKinsey has said that Lilli and its related tools are saving up to 30%" of the time consultants once spent searching and synthesising information; around three quarters of employees use Lilli; more than 500,000 prompts a month, as reported by McKinsey. We could not open McKinsey's own pages, which timed out, so these figures are as the press reported them.
- 3
Gartlehner, G., Kahwati, L., Hilscher, R., Thomas, I., Kugley, S., Crotty, K., Viswanathan, M., Nussbaumer-Streit, B., Booth, G., Erskine, N., Konet, A. & Chew, R., "Data extraction for evidence synthesis using a large language model: A proof-of-concept study," Research Synthesis Methods 15(4), 576 to 589, 2024, doi:10.1002/jrsm.1710 (peer-reviewed; abstract read via the RTI International publication page). rti.org
Back to textWhat we took from it
A convenience sample of 10 English-language, open-access randomised controlled trial publications, 160 data elements; Claude 2 "demonstrated an overall accuracy of 96.3%", with test-retest accuracy of 96.9% and 95.0%; 6 errors, of which the most common (4) were missed data items.
- 4
Lieberum, J. L., Toews, M., Metzendorf, M. I., Heilmeyer, F., Siemens, W., Haverkamp, C., Böhringer, D., Meerpohl, J. J. & Eisele-Metzger, A., "Large language models for conducting systematic reviews: on the rise, but not yet ready for use, a scoping review," Journal of Clinical Epidemiology 181, 111746, published online 26 February 2025, doi:10.1016/j.jclinepi.2025.111746 (peer-reviewed; abstract read via Europe PMC, the publisher page refused automated access). ebi.ac.uk
Back to textWhat we took from it
37 included articles, last search February 2024; LLMs supported 10 of 13 systematic review steps, most often literature search (41%), study selection (38%) and data extraction (30%); 57% were validation studies; authors judged LLM use promising in about half of studies; few fully validated applications exist.
- 5
Chelli, M., Descamps, J., Lavoué, V., Trojani, C., Azar, M., Deckert, M., Raynier, J. L., Clowez, G., Boileau, P. & Ruetsch-Chelli, C., "Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews: Comparative Analysis," Journal of Medical Internet Research 26, e53164, 22 May 2024, doi:10.2196/53164 (peer-reviewed; abstract read via Europe PMC). ebi.ac.uk
Back to textWhat we took from it
11 systematic reviews across 4 fields; 471 references analysed in total, of which 362 were returned by the models (139, 119 and 104); a reference counted as hallucinated if two or more of title, first author and year were wrong. Hallucination rates: GPT-3.5 39.6% (55 of 139), GPT-4 28.6% (34 of 119), Bard 91.4% (95 of 104). Precision: GPT-3.5 9.4%, GPT-4 13.4%, Bard 0%. Model versions from July 2023 (GPT-3.5 and GPT-4 "July 19 version", Bard released 13 July 2023); the full text (PMC11153973) does not say whether the models could search the web.
- 6
Jaźwińska, K. & Chandrasekar, A., "AI Search Has a Citation Problem," Columbia Journalism Review, Tow Center for Digital Journalism, 6 March 2025. cjr.org
Back to textWhat we took from it
Eight generative search tools with live search; ten articles from each of 20 news publishers; 1,600 queries asking each tool to identify an excerpt's headline, publisher, date and URL. "Collectively, they provided incorrect answers to more than 60 percent of queries"; Perplexity answered 37% incorrectly and Grok 3 94%; premium tools answered more prompts correctly than free ones but had higher error rates, mainly by giving definitive wrong answers; more than half of responses from Gemini and Grok 3 cited fabricated or broken URLs.
- 7
Flemyng, E., Noel-Storr, A. et al., "Position statement on artificial intelligence (AI) use in evidence synthesis across Cochrane, the Campbell Collaboration, JBI, and the Collaboration for Environmental Evidence 2025," JBI Evidence Synthesis 23(11), 2162 to 2166, 6 November 2025 (published simultaneously in several journals). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
"Evidence synthesists are ultimately responsible for their evidence synthesis"; "AI and automation in evidence synthesis should be used with human oversight"; AI may be used "as long as they can demonstrate that it will not compromise the methodological rigor or integrity of their synthesis"; "Any use of AI or automation that makes or suggests judgments should be fully and transparently reported", including the tool's name, version and date, purpose and limitations.
- 8
Three press reports. news.bloombergtax.com
Consultancy.com.au, "Senate publishes false KPMG and Deloitte claims invented by AI," 3 November 2023: consultancy.com.au
The Mandarin, "Parliamentary committee reminds parties to verify submissions generated by AI," 7 November 2023: themandarin.com.au
Back to textWhat we took from it
Nesbitt, D., "Big Four False Allegations Based on Google AI, Academic Claims," Bloomberg Tax, 3 November 2023: Emeritus Professor James Guthrie of Macquarie University's Department of Accounting and Corporate Governance claimed sole responsibility for false allegations in two submissions made jointly with other academics to the parliamentary joint committee inquiry into consultants' ethics, blamed Google's Bard, and apologised to the big four firms, other accounting firms and the committee.
the false case studies included KPMG being complicit in the 7-Eleven wage theft scandal and Deloitte fudging the accounts of Patisserie Valerie, which Grant Thornton audited; the material was "published under parliamentary privilege"; the academics offered "an unreserved apology"; KPMG's chief executive, Andrew Yates, said AI had been relied upon "without comprehensive fact-checking". This report calls the inquiry a Senate inquiry; Bloomberg Tax and The Mandarin name the Parliamentary Joint Committee on Corporations and Financial Services, which we follow.
after two major accounting firms wrote to correct errors in the academics' submission, the committee issued a statement that submitters using AI tools should "ensure the output is truthful". We could not open the Guardian's original report or the Parliament submissions page (access refused), so we do not state whether the submission was later withdrawn or replaced.