Send client decks you can stand behind

It's the night before the steering committee. The deck runs to 40 slides, three people have edited it, and the analyst who pulled the competitor figures on slide 12 rolled onto another engagement last week. The partner will read it once, on a phone, and ask one question: where is this from? This page shows how to check the claims in a client deck before it goes out, paste in suggested fixes for the ones that fail, and keep an Excel record of what you checked.

Clients are reading your footnotes

Two cases in 2025 made the risk concrete. Deloitte Australia delivered a report to the Department of Employment and Workplace Relations under a contract worth about A$440,000. It turned out to contain a fabricated quote from a federal court judgment and references to research papers that don't exist, and Deloitte refunded more than A$97,000. Weeks later, a 526-page health workforce plan that Deloitte prepared for Newfoundland and Labrador, which cost the province just under $1.6 million, was found to cite academic papers that were never written. The firm said AI had been "selectively used to support a small number of research citations."

Neither was a slide deck, and nobody needs AI to get a figure wrong. But the pattern will be familiar to anyone who has run QC on a big deliverable: a plausible, specific, unchecked detail, found by the client's side rather than yours. The deliverable carries your firm's name, and the fix costs more once it's in their hands.

What a document check gives you

Upload the deck as a PowerPoint file or PDF, or a report as Word or PDF, up to 20 MB. Reading it costs about 1 credit. CiteJury pulls out every checkable claim, and you choose which to check, up to 25. You can reword any claim first, which matters when a slide title squeezes a fact into four words.

Each claim then runs as its own check. CiteJury searches several independent search indexes, favoring primary sources such as filings, regulators and official pages, and looks for evidence against the claim as well as for it. The best passages are sealed into one evidence pack. Claude, ChatGPT and Grok answer separately from that same pack, every quote they cite is checked word for word, and a final review settles the verdict. Each claim comes back with a verdict, a one-line headline and a suggested fix you can copy.

The report puts the whole document in one place. The Claims tab lists every verdict. The Document tab shows your original text with each claim highlighted in its verdict color, so you can see at a glance which slides need work. You can download it as Excel for the QC file. Each claim also links to its own full check. Here's what one looks like, from a claim about sell-side research that came back Correct and still changed the footnote.

“CY27 EPS estimates for Goldman's 42-company basket of Indian “AI Enablers” are up 22% in 2026, while Nifty 500 estimates were cut 2%.”

Correct

Claude, ChatGPT and Grok agree12 sources usedTook 93 seconds

Goldman's own figures, as reported by Financial Express, match both numbers.

• Goldman has a 42-stock AI Enablers basket

  Confirmed by several outlets.

• CY27 EPS estimates are up 22% this year

  Stated in source 1, quoting Goldman.

• Nifty 500 estimates were cut 2%

  Stated in source 1. Nothing contradicts it.

Safe way to say it

Attribute the 22% upgrade and the 2% cut to Goldman Sachs research, as reported by Financial Express, since only that source carries them.

01financialexpress.comBacks it up

02curiouscats.aiBackground

08m.economictimes.comBackground

Both figures matched, and all three AIs agreed. But only one source, Financial Express, carried them, so the safe wording says to attribute them to Goldman Sachs research as reported by Financial Express. On a client slide, that footnote tells the reader exactly how far the figure has traveled.

A supplier-risk deck, checked

The file was a risk-assessment deck on seafood suppliers. Twelve claims were checked: five came back Correct, four Incorrect and three Sources disagree. Seven of the twelve needed attention before the deck could go anywhere. Here's the part of the report covering slides 5 and 6.

USDA Risk Assessment Sprint.pptx7 of 12 claims need attention

Slide 5

Trident Seafood: risk analysis

Trident recalled frozen fish due to potential plastic contamination. (Correct)

Trident settled with the EPA for around $3M under the Clean Air Act. (Incorrect)

Washington fined Trident $25K for oil spill negligence that led to a fire. (Incorrect)

Slide 6

Blue Harvest Fisheries: risk analysis

USDA awarded nearly $29M to Blue Harvest and two other firms. (Sources disagree)

Blue Harvest acquired High Liner Foods. (Incorrect)

Blue Harvest acquired Hygrade Ocean Products. (Correct)

Incorrect

Wrong amount: a $900,000 penalty plus up to $23M in fixes, not about $3M.

Suggested fix: Trident settled Clean Air Act claims with the EPA, paying a $900,000 penalty and agreeing to spend up to $23 million on refrigerant fixes.

Incorrect

Cause and effect reversed: the fire caused the oil spill.

Suggested fix: Washington's Department of Ecology fined Trident $25,000 after a fire ruptured hoses and spilled oil.

Incorrect

It bought one business unit, not the company.

Suggested fix: Blue Harvest acquired High Liner Foods' scallop business and its New Bedford processing plant.

The errors on slide 5 are worth reading closely, because neither would jump out on a partner's read:

• A wrong amount. The slide said Trident settled with the EPA for around $3M under the Clean Air Act. The sources showed a $900,000 penalty plus an agreement to spend up to $23 million on refrigerant fixes. The suggested fix states both.
• Cause and effect reversed. The slide said Washington fined Trident $25K for oil spill negligence that led to a fire. It was the other way round: the state's Department of Ecology fined Trident $25,000 after a fire ruptured hoses and spilled oil.

On slide 6, the deck said Blue Harvest acquired High Liner Foods. It bought one business unit, High Liner's scallop business and its New Bedford processing plant, not the company. A USDA award of nearly $29 million to Blue Harvest and two other firms came back Sources disagree, which is a prompt to read both sides and word the slide carefully, not a reason to cut it.

A wrong amount, a reversed cause and a part mistaken for the whole are classic errors in a deck built fast from secondary coverage. Each fix was one sentence you could paste straight onto the slide.

Before the deck goes out

1. Freeze a version. Check the file you're about to send, not one that's still changing.
2. Upload it as a .pptx, or as a PDF with a text layer.
3. Pick the claims the recommendation rests on. Anything with a figure, a date, a named company or a regulatory fact, plus every claim on the executive summary. Up to 25 per document.
4. Tighten the wording. Add the year, currency and geography the slide assumes but doesn't say.
5. Choose a tier. Standard, with three AIs and a final review, runs about 8 credits per claim and usually takes one to three minutes. Quick, with one AI, suits a first pass. Deep suits claims where sources are thin or contested.
6. Work the Claims tab from the worst verdicts up. Fix Incorrect first, then decide how to present Sources disagree, then chase Can't confirm.
7. Download the Excel report and file it with the deliverable, or share a link to any claim's check with your engagement manager.

QC habits worth keeping

• Separate client data from public facts. Figures from the client's own systems can't be checked against public sources. Label them as client data and don't spend checks on them.
• Cite the primary, not the aggregator. If a number reached you through a news story about a report, footnote the report, or write "as reported by".
• Distrust recycled slides. A market figure lifted from last year's engagement may now be the wrong year. Check it again.
• Present disagreement honestly. When sources disagree, show the range or the stronger source and say why. A caveat on the slide reads better than a correction after the meeting.
• Check the units. Millions against billions, fiscal against calendar year, nominal against real.
• Confirm every reference exists. CiteJury checks what your sentences claim. It doesn't validate a bibliography, so an AI-drafted appendix still needs someone to confirm each cited paper is real.

Earlier in the project, a sourced research brief shows what's established and what's contested before you build the storyline, and a sourced comparison gives you a vendor table where every cell links to its source. The document flow itself is covered in fact-checking a document.

What it won't do

• It reads text, not charts. Numbers inside chart images or pictures aren't checked. Scanned PDFs without a text layer can't be read, because there's no OCR.
• It can't read paywalled research. Subscription databases and reports behind a login are out of reach. The result lists the pages it couldn't open, so you know where your firm's own subscriptions come in.
• It checks facts, not logic. Whether your recommendation follows from the facts is still your team's call.
• It isn't legal or financial advice, and a verdict is only as good as the sources found. Every sentence links to its source so you can check before you rely on it.

On confidentiality, only upload what your engagement terms allow. To run a check, the text goes to the AIs you've turned on in Settings, and search queries based on it go to search services. CiteJury keeps the document's text, not the file, and doesn't train AI models on your content.

A check won't make a weak deck strong. It helps the facts in a good one survive the client's first read, and it leaves you a record of what you checked.

Questions people ask