5 Techniques to Prevent AI Hallucination in Office Work with WorkBuddy & Qwen Office

Why AI Hallucinates in Office Work

During AI office efficiency training for universities and state-owned enterprises, two recurring concerns emerged: token limits and AI fabricating facts. This article addresses the second issue. The academic term for "fabrication" is hallucination — AI generates plausible but untrue content. OpenAI's "Why Language Models Hallucinate" identifies a root cause: common training and evaluation methods reward guessing. Like an exam where wrong answers aren't penalized, models guess to maximize scores.

Adding "do not fabricate" to prompts helps but doesn't guarantee reliability. The key is establishing a verifiable workflow. Below are five techniques tested with WorkBuddy and Qwen Office.

Technique 1: Clarify Task Boundaries

Typing constraints make it hard to specify tasks fully. Use AI voice input (built-in or tools like Doubao Input, WeChat Input, Shandian Shuo, Typeless) for richer context. Also use the prompt enhancement tool in AI office software to expand vague prompts into detailed, professional ones.

Errors often arise when AI auto-fills ambiguous conditions. Example: "Summarize this year's revenue" requires explicit cutoff date, tax inclusion, refund handling .

PROMPT: Before starting, confirm key definitions that affect the conclusion; if ambiguous, ask me — do not assume.

In WorkBuddy, enable Plan Mode : the agent pauses at ambiguities, asks for missing info, and presents a plan for review before execution. Check the plan's statistical definitions and sources.

Technique 2: Provide Accurate, Complete Materials

When AI lacks necessary information (e.g., internal business data), it may invent plausible answers. Supply all required reference materials. For external info, require the agent to consult official sources and verify publication date and applicability.

PROMPT: First confirm you can read the required materials; if key information is missing, list the gaps. Do not fill in data not present in the materials.

Technique 3: Require Sources for Key Conclusions

Even with clear tasks and materials, AI may misinterpret or add unsupported conclusions. Require source citations for key claims to enable traceback.

PROMPT: At the end, append a verification table with columns: "Key Conclusion, Source File or URL, Original Text & Location". Mark conclusions without evidence as "To Be Verified".

Warning: Citations must be checked — open the source, confirm it exists, and that the original text actually supports the conclusion. Keep a draft with the verification table; remove it for the final version if needed.

Technique 4: Cross-Check with Another Model

Different models excel at different tasks. For long texts, complex requirements, or data analysis, after the first draft, switch to another model or agent for review. Example: the author used Hunyuan 4 for a visualization report, then asked a different model to check facts, data, and citations against the original task materials and draft.

If only the previous answer is given to the reviewer, it may simply agree. Even if two models concur, key conclusions must still be traced to original sources.

Select models by testing them on your typical tasks — observe which hallucinates less and which admits ignorance. Release date alone doesn't predict reliability for your specific work.

Technique 5: Delegate Calculations to Tools, Keep the Audit Trail

AI efficiency doesn't mean letting the model compute everything. For Excel aggregations, ratios, filtering, require the agent to use formulas or code and preserve input data range, filter conditions, formulas or code for re-calculation.

Still verify the model didn't pick wrong data or misinterpret definitions. Formulas aid verification but presuppose correct setup. For stable, well-defined operations, use dedicated tools directly — they're faster and less error-prone.

For deliverables involving names, amounts, dates, and policy bases , manually cross-check against original sources. Ultimately, humans bear accountability for errors.