How to Use AI to Draft and Review Term Sheets

The workflow has four stages, and each one has a clear gate. Do not skip a gate because the output looks clean.

Stage 1 is the one people skip, and it is the one that determines whether the rest works. A baseline is a short document you write once, listing your standard positions: 1x non-participating liquidation preference, broad-based weighted average anti-dilution, a board of two founders plus one investor plus one independent, an option pool of a stated size. When you later ask the model to flag deviations, it flags them against your written positions rather than against its own guess about what is normal.

The baseline also makes the comparison consistent across deals. Two analysts reviewing two term sheets by hand will produce two different memos. The same baseline plus the same prompt produces comparable memos.

The extraction prompt should ask for a table with fixed columns and demand a verbatim quote and a location for every value. Fixed columns matter because you will compare across documents later, and free-form summaries do not compare.

You are extracting terms from a venture term sheet. Use only the
attached document. Do not infer, complete or guess.

Return one Markdown table with these columns:
Clause | Value | Exact quote | Location

Clauses to extract:
- Instrument type (priced round, SAFE, convertible note)
- Pre-money valuation and investment amount
- Option pool size and whether pre- or post-money
- Liquidation preference (multiple, participating or not)
- Anti-dilution (type)
- Board composition
- Protective provisions (list each veto right)
- Pro-rata rights
- Drag-along threshold
- Vesting terms, including any credit for time served
- Exclusivity or no-shop period
- Expense reimbursement cap

Rules:
- Quote the clause verbatim in the Exact quote column, trimmed to
  the sentence that carries the term.
- Put the section number or page in Location.
- If a clause is absent, write ABSENT in Value and leave the quote
  empty.
- Do not add commentary.

Two details in that prompt do most of the work. The "ABSENT" instruction stops the model from inventing a plausible value when a clause is missing, which is a common failure on early-stage term sheets where vesting or drag-along are often omitted. The "do not add commentary" instruction keeps the output usable as data.

Google's technical writing guidance on using LLMs describes the same technique in general terms: give the model a role, define the goal of the document, add constraints, and iterate on the prompt rather than accepting the first result see the Google Tech Writing guide. The guide also notes that role and context setting improve output quality, which is why the prompt above names the document type and restricts the model to the attachment.

Run extraction on one term sheet first, then diff the table against the document line by line. If the model missed a clause or misquoted one, adjust the prompt before running the rest. A five-minute check on document one saves an hour on documents five through ten.

Once you have one table per term sheet, comparison is a text operation the model does well. Feed it two tables and your baseline and ask for a difference table with three columns: your baseline position, the investor's position, and a flag.

Below are: (A) our standard positions, (B) the extracted terms from
Investor 1, and (C) the extracted terms from Investor 2.

Produce a table: Clause | Baseline | Investor 1 | Investor 2 |
Deviation from baseline (MORE FAVORABLE / LESS FAVORABLE / SAME /
NOT ADDRESSED) | Exact quote supporting the flag.

Rules:
- Flag only deviations from the baseline, not differences between
  the two investors.
- Every flag needs a quote. If there is no quote, write NO QUOTE and
  do not flag it.
- Do not rank the investors or recommend one.

The "do not rank" instruction is deliberate. Ranking is judgment, and a model ranking two investors invites the founder to skip the analysis. A table of deviations invites the founder to read.

For a pricing comparison, keep the arithmetic out of the model and do it yourself or in a spreadsheet, because the inputs are what matter. The valuation cap is the maximum effective valuation at which a SAFE investor's shares convert, and for a post-money SAFE you divide the investment amount by the post-money cap to get the ownership percentage per CRV's guide. That guide gives the worked case: $200,000 on a $4 million post-money cap is exactly five percent at signing.

Post-money structure changes how a stack of instruments behaves, and the difference is large enough to model rather than estimate. Under a post-money SAFE, each investor's percentage is fixed at signing, so every additional SAFE dilutes the founders rather than the earlier SAFE holders as Avisen Legal explains. Their worked example is worth copying into your own model: $500,000 at a $5 million cap (10 percent), $500,000 at a $6 million cap (about 8.33 percent) and $1 million at an $8 million cap (12.5 percent) combine to roughly 30.83 percent for the SAFE holders, leaving founders about 69.17 percent before any priced-round dilution or option pool expansion.

That is the kind of number a founder should see next to the term sheet, and it is exactly the number a language model should not be trusted to compute silently. Ask the model to list the inputs (amount, cap, structure for each instrument), then run the arithmetic in your own model and paste the results back.

A flagged clause is a prompt for a question, not an answer. Four flags come up often enough to deserve a standing list of follow-up questions.

Liquidation preference. A 1x non-participating preference is the common seed and Series A position, while a 2x or participating preference is investor-friendly and materially shifts downside economics per Mercury's comparison table. The question to ask is not "is this standard" but "what does the founder receive at a range of exit values." Run three exits in a spreadsheet and the preference speaks for itself.

Option pool timing. A pool expressed as a pre-money percentage dilutes the founders before the investment lands, which is why the same headline pool size can mean very different outcomes depending on the label. Mercury's table notes that investors often push for larger pools at Series A. The follow-up question is whether the pool is stated pre- or post-money, and the model should be asked to quote the exact sentence rather than label it.

Pro-rata rights. Pro-rata rights give a holder the option, not the obligation, to invest in the next round to hold their percentage, and granting them to many SAFE holders can complicate a Series A because the allocation the lead wants may be consumed per CRV. The practical check is who holds the right, and whether it is limited to the next round only.

Exclusivity and expenses. These are the provisions most likely to be binding even when the financing terms are not. A term sheet can designate confidentiality, exclusivity, governing law and expense treatment as binding while the investment terms remain subject to definitive agreements per focal. The model should extract the exclusivity period and the expense cap with quotes, and a lawyer should read both.

A second comparison is worth running on the instruments rather than the clauses. A SAFE has no interest, no maturity date and no repayment obligation, while a convertible note is debt with interest and a maturity date per CRV. If the documents in your comparison set include both, the model should say which is which at the top of each table, because the review questions differ.

Drafting is the easier direction. Give the model your baseline positions plus a one-page deal summary (amount, valuation, investor name, board ask) and ask for a first draft using a standard template you already have. Review the output for internal consistency, then hand it to counsel. Do not let the model invent a clause it has not been given language for; if your template lacks a drag-along provision, the draft should say so rather than improvise one.

Guardrails come in four kinds, and all four are cheap to implement.

Redaction. Strip what the analysis does not need. Investor names, personal email addresses, bank details and any confidential commercial terms not relevant to clause comparison can be replaced with placeholders. Keep the clause language, which is the part the model needs.

Exact quotes. Every finding carries a quote and a location. A finding without a quote is discarded, not escalated. This single rule catches most hallucinations, because a fabricated clause cannot be quoted from the document.

Version records. Save the prompt, the model name and the output for each review. If a number turns up in a board memo three weeks later, you can trace where it came from.

Mandatory lawyer review. The model's output is a draft and should be labeled as such at the top of every file. Even standardized documents raise questions a model cannot settle: a practitioner guide to SAFEs notes that standardization does not remove the need for legal advice when the terms are negotiated as this SAFE guide explains. The same logic applies to term sheets.

One more guardrail is worth writing down: never send a model-generated summary of a term sheet to an investor without a lawyer's eyes on it. Summaries are where interpretation creeps in, because summarizing requires deciding what matters.

For teams that want to run this workflow across several models, or keep prompts and outputs in one place, a single API for text and structured output removes the plumbing. Quanta Inference puts text, image and embedding models behind one OpenAI-compatible key, prepaid and priced in US dollars: https://quantaintelligence.ai/inference.

Most failures in AI-assisted term sheet review fall into a short list, and each has a specific check.

The silent omission row deserves emphasis because it is the hardest to spot. If your clause list is complete and the prompt requires an explicit ABSENT for each missing term, you can audit by counting rows: eleven clauses in, eleven rows out. If a document returns ten rows, the model dropped one, and the comparison is incomplete.

The arithmetic row is next. The ownership math on a post-money SAFE is a division, but the inputs are easy to misread, and a model that quietly substitutes a pre-money cap for a post-money cap will produce a plausible and wrong number. Recompute. The worked examples in published guides exist precisely because founders get this wrong by hand.

A worked check: an investor offers $1 million at a $10 million post-money cap. The model should return ten percent and quote the cap sentence. If it returns nine percent, it has probably divided by a figure that includes the investment amount twice, and the quote check should surface the error.