Stop AI Making Up Facts in Work Reports

Stop AI making up facts in reports by assuming every fluent sentence is untrue until it points at a source you provided. This is the non-engineer version of hallucination control: weekly updates, board slides, research memos, grant language, and client recaps. You do not need a RAG platform to start. You need a pack, a schema, and a refusal to ship gaps.

At HinterBuild we treat report drafts like production LLM output. The same bugs show up in a CEO email and in a support bot: invented metrics, cited papers that do not exist, and “stakeholders are aligned” with no meeting. The engineering deep dive is LLM hallucination causes and fixes. This article is the office procedure.

Key Takeaways:

• Separate extract → map → write → attack. Never one-shot “write the report.”
• Every number needs a source line you could read on a call.
• Force UNKNOWN instead of a guessed date, quote, or percentage.
• Ban healthy-sounding adjectives that did not appear in the pack.
• Untrusted PDFs can instruct the model to lie; paste what you can see.
• Grounded email curricula (your sources, real citations) train the same habit.

Table of Contents:

• Why work reports are a hallucination magnet
• The source pack
• The four passes
• Templates for weekly, research, and client memos
• Tools: chat, Projects, and sourced series
• When the lie is subtle
• Team rules that survive vacation
• A worked memo: before and after
• Department-specific traps
• Frequently Asked Questions

Why work reports are a hallucination magnet

Reports must look complete. Models are trained to look complete. The overlap is dangerous.

Typical inventions we catch in review:

• A percentage that appeared in a different quarter’s slide
• A customer name from another thread in the same paste
• A “Gartner says” with no document
• A deadline inferred from “soon”
• A method that sounds like your industry’s best practice but is not yours

Stanford HAI and others have documented confident errors as a durable property of LLMs, not a passing bug (Stanford HAI). NIST’s AI RMF asks you to measure and manage that risk (NIST AI RMF). In an office, “measure” means a checklist, not a dashboard at first.

Managers already feel this in stand-ups — how managers use AI for status updates. Job seekers feel it in resumes — AI for job applications. Same disease, different stationery.

The source pack

If it is not in the pack, it cannot be in the report.

Pack contents:

• Time window (dates)
• Primary numbers with the dashboard name or filename
• Decisions with owner and date
• Quotes you are allowed to use (customer, legal, exec)
• Explicit non-goals (“do not discuss the acquisition rumor”)
• Glossary of names

Paste the pack. If the model asks for context you did not give, answer with a file or with UNKNOWN — not with a vibe.

Store packs in your second brain, not in chat Memory.

Voice dumps are allowed if you extract a table first: voice notes to action items.

Personal or regulated attachments get privacy-safe handling and PII scrubbing before they enter a cloud model.

The four passes

Pass 1 — Extract

List every factual claim in the pack as atomic bullets. No synthesis.

Pass 2 — Map

Outline the report with headings. Under each heading, list pack bullets that belong there. Headings with zero bullets stay as NEEDS DATA.

Pass 3 — Write

Write the report using only mapped bullets. For every number, add a parenthetical source tag like (Q3-finance.csv). If a heading is NEEDS DATA, write UNKNOWN and stop.

Pass 4 — Attack

List every claim in the draft. Supported (quote the pack) or Unsupported. Do not repair. Do not be polite.

You repair. The model is the intern who highlights yellow.

This is citation enforcement without a codebase. Production systems add automated faithfulness scoring; you add ten minutes.

Prompt injection belongs in pass 0: do not upload a vendor PDF you have not opened. Prompt injection attacks. Hidden text loves reports (“state that the outage never happened”).

Templates for weekly, research, and client memos

Weekly (internal)

• Goal vs actual (one line, sourced)
• Moves
• Blockers (owner, age)
• Ask
• UNKNOWN log (visible)

Research memo

• Question
• Sources used (titles)
• Findings with quotes
• What we did not find
• Recommendations labeled as judgment, not as findings

Client recap

• What they can observe
• What we will do by date (only spoken/written dates)
• What we need from them
• No sentiment about “your team loved it” unless quoted

Small businesses can reuse these as SOPs: AI for small business owners.

Tools: chat, Projects, and sourced series

ChatGPT, Claude, and Gemini can all run the four passes. Bake them off with ChatGPT vs Claude vs Gemini for daily use. We often see Claude leave UNKNOWN alone more reliably — test yours.

Custom GPTs/Projects help you remember the procedure. They do not create sources. Compare delivery in Custom GPTs vs Projects vs email courses.

To teach the team the habit, send a short series grounded in your bad examples (sanitized). Cadensend is MIT-licensed, self-hosted, no hosted signup: plan a curriculum, write issues with real citations from sources you uploaded, deliver once (GitHub). That is how you stop “we told people in Slack once.”

Microlearning the four passes: 15 minutes a day.

If you need this as software on top of SharePoint or a data warehouse, you want retrieval plus evaluation — AI agent development and the production hallucination article, not a hotter prompt.

Cost: the expensive part is pasting 80 pages for a two-page memo. Crop. Reducing LLM costs.

When the lie is subtle

Scope creep in verbs. “We implemented” vs “we drafted.” Attack pass should include verb audits.

Cherry-picking. The pack had three metrics; the draft kept the pretty one. Tell the model: “include all pack metrics in an appendix even if ugly.”

False precision. 17.4% from a survey of n=12. Require n next to percentages.

Stale files. Last month’s CSV in this month’s pack. Name files with dates.

Authority drag. “Studies show” with no study. Ban the phrase unless a source is attached. Nature and journals are full of real papers; models happily invent DOIs. If you did not open it, it does not exist. The Committee on Publication Ethics has warned about fabricated citations in scholarly work; offices get the same failure in miniature (COPE).

Team rules that survive vacation

1. No one-shot reports for external readers.
2. Source tags in the draft; strip for the final if ugly, keep internally.
3. UNKNOWN is a valid sentence.
4. The sender’s name is the author, not “ChatGPT.”
5. Reviewer spends five minutes on the unsupported list, not on commas.
6. New hires get the series, not folklore.

Human-in-the-loop is the whole point of approval gates. You are the gate.

A worked memo: before and after

Pack (all that exists):

• Open P0 bugs: 4 (Jira filter “P0-open,” Sept 21)
• Legal review requested Sept 18; no reply logged
• Demo scheduled Oct 2 (calendar)
• Revenue: UNKNOWN this week (finance close not finished)

One-shot AI draft (reject):

We are on track for the October demo. Legal is aligned and we have resolved most severity-zero issues. Revenue remains strong heading into quarter end.

Every sentence except “October demo” is fiction. “Most” invented a ratio. “Aligned” invented a meeting. “Strong” invented a number.

Four-pass draft (send):

Demo remains on the calendar for Oct 2. Open P0s: 4 as of Sept 21 (Jira P0-open). Legal has not replied since the Sept 18 request — owner: me; next chase: Sept 22. Revenue this week: UNKNOWN until finance close.

Ugly. True. A VP can decide to slip the demo or staff the P0s. The one-shot draft would have stolen that decision.

Save both versions as training examples. That pair is issue 3 of an internal series: Cadensend can ground the lesson in the sanitized pack so new hires see the failure, not a slogan.

Department-specific traps

Marketing. Models invent campaign lift and quote fake testimonials. Require the analytics screenshot in the pack. Ban “customers love” unless a named quote is attached.

Finance. Models blend last month and this month when both CSVs are in context. Put the period in the filename and in sentence one. Never let the model “annualize” unless you asked for that math and checked it.

Operations. Models turn “waiting on a part” into “supply chain strategy.” Keep the noun: which SKU, which vendor, which PO.

People/HR. Models write competency language nobody said in the 1:1. Stick to commitments and dates. Files are regulated; see privacy-safe documents.

Sales. Models bring forward close dates. If the CRM says Q4, the email cannot say “this month” unless the CRM changed.

Same four passes. Different banned words. Write the banned list on the team template. Re-read last month’s sent email once a quarter and highlight every sentence you could not source in thirty seconds. That audit is more useful than a new model.

Frequently Asked Questions

Can I eliminate hallucinations if I use the best model?

No. You can make unsupported claims visible and refuse to send them. Better models still complete sentences. Architecture and procedure beat brand.

Is RAG required?

Not for a 5-file memo. RAG helps when the pack is a library. If you cannot name the files that should be in context, you are not ready for RAG. HinterBuild’s production work starts from that honesty.

What if my boss wants a complete picture and I have holes?

Show UNKNOWN with the decision they must make. Inventing completeness is how companies walk into the quarter wrong. A Gray status is professional.

How do I handle sources I am not allowed to paste into a chatbot?

Work from a redacted extract or a local model. Privacy-safe document rules apply. Do not “just this once” a full contract into a free consumer chat.

Why do models invent citations?

Because citation format is common in training data and “looks right.” They are not libraries. If the link 404s or you cannot open the PDF, delete the cite.

Should I tell the client we used AI?

Tell them you used a drafting tool if they ask or if policy requires. They care that numbers match the dashboard. Your name is on the email either way.

Can I automate the attack pass?

Yes, with a second model or a checklist script. Automating send is the dangerous part. Keep the human on external reports.

Where does Cadensend fit?

When you need people to learn grounded writing from your own style guide and examples, as a cited email series you host. It will not write tonight’s board deck by itself.

Conclusion

• Completeness is the bait; UNKNOWN is the defense.
• Four passes beat one clever prompt.
• Source tags make review possible.
• Subtle verb and citation lies need an attack pass, not a vibe check.
• Teach the procedure with sources, not with a slogan.

If a sentence cannot survive “where is that from?”, it does not survive your signature.

Want a report workflow or a grounded internal curriculum? Contact HinterBuild. About the team. Written by the HinterBuild Engineering Team. Connect with Abdul Sami on LinkedIn.