Five questions every AI-generated file should answer before you send it

By Dr Carlos Velez, expert trainer of The AI for Pharma Business Development & Licensing Course and Business Development & Licensing Course For Pharma & Biotech.
 
 

A working check for business development & licensing professionals in life sciences. 

AI can produce a partner brief, an rNPV model or a due-diligence memo in minutes. Whether that file is safe to send to a licensor, licensee or your management team is a separate question.

It comes down to five checks: source traceability, labelled assumptions, live calculations, verified numbers and independence from the person who built it. If a file fails any one of these, it is not finished yet, no matter how polished it looks.

Business development & licensing teams in pharma and biotech use large language models to draft partner briefs, build rNPV and revenue models, and summarise due-diligence findings, whether they're scouting an in-licensing candidate or preparing an out-licensing data room. The output looks professional: clean formatting, confident language, numbers that line up neatly across tabs and slides.

But looking finished and being correct are different things, and the gap between them is where AI-generated errors travel quietly into deal decisions. A hardcoded cell, an unsourced regulatory claim or a slightly wrong prevalence figure reads exactly the same as a verified one, right up until a licensor, a licensee or your own management team relies on it.

The check itself is not complicated. Before an AI-generated brief, model or memo leaves your outbox, it should be able to answer five questions.
 

The checklist
 

1. Can you trace every claim back to a source?

   Every factual statement in the file should carry a visible source: a document name, a date, a reference. If you cannot say where a claim came from, treat it as unverified, not true.

   Language models are fluent, not evidential. They will state a trial result or a patent expiry date with exactly the same confidence whether it came from an approved source or from a plausible pattern in their training data.
    

2. Are assumptions clearly separated from facts?

   Assumptions, estimates and placeholders should be visibly labelled as such, not folded into the same sentence as sourced data.

   "Peak sales of $800 million" reads identically whether it is a sourced forecast or an AI-generated placeholder, unless the file tells you which.
    

3. Do the calculations update automatically?

   In any AI-built financial model, check whether changing an input actually changes the output. Hardcoded values dressed up as formulas are one of the most common, and hardest to spot, AI errors in spreadsheets.

   A discount rate cell can look live but was pasted in as a static value. Change the assumption and the valuation does not move, and nothing in the formatting tells you so.
    

4. Has every number been checked against its original source?

   Cross-check figures against the underlying document, not the AI's summary of it. Small transcription errors are easy for a model to introduce and easy for a reader to miss.

   A prevalence estimate of 250,000 instead of 200,000 will not look wrong on the page. It will only look wrong next to the source.
    

5. Would the file hold up without you in the room to explain it?

   If a colleague forwarded the file with no commentary from you, would every number and claim still make sense on its own? If the answer is no, the file is not ready to leave your hands.
    

The discipline behind the questions

None of this is about distrusting AI tools. It is about applying the same verification discipline you would apply to a junior analyst's first draft: check the sources, separate fact from assumption, and confirm the numbers actually work before your name goes on the file. The tools are new. The standard is not.
 

FAQ