Preparing transfer pricing documentation that survives a second read

Most transfer pricing documentation failures are not literary. They are structural. The industry chapter was copied from last year. The transaction schedule does not match the Form 3CEB. The FAR still describes a full-fledged manufacturer after the principal moved IP offshore. The annexures were "to follow." None of that is fixed by a better adjective in the executive summary.

This note is how we staff a Local File inside TP DOC GEN AI's documentation workflow, and what we refuse to let a model invent.

Start with the transactions, not the prose

OECD Local File guidance is organised around controlled transactions. India's rules are even more explicit: the international transaction is the statutory unit. If step one is "write the company overview," you will write a brochure. If step one is "list related parties, connected persons, amounts, and contracts," you will write a study.

In the product, that is a dedicated wizard step with Excel import. Each line should be specific enough to characterise: sale of finished goods is not the same as a limited-risk distribution arrangement with a marketing intangible debate attached. A "management fee" that bundles secondments, IT, and a guarantee will produce one method that fits none of the underlying deals. Split it.

FAR before method, method before search

Functional analysis is where files are won or lost. Who develops the intangibles. Who bears inventory risk. Who sets prices. Who funds losses. The tested party follows from that picture, not from which entity has the cleaner P&L. We score potential comparables 0–100 across industry, functions, risk, and operations, with a written rationale per score, precisely so a reviewer can disagree with a weight instead of a vibe.

Only then does method selection make sense. CUP where a CUP exists. TNMM where the tested party is the less complex entity and a reliable PLI can be computed. Profit split where both sides contribute unique intangibles. Working backwards from a desired operating margin is not advisory. It is a future assessment order.

What the AI is allowed to draft

TP DOC GEN AI drafts executive summary, entity and industry overviews, transaction commentary, FAR narratives, economic conclusions, and comparable-company descriptions grounded in live web search. It does not write tables, numbers, statutory citations, cover pages, the table of contents, or the glossary. Quartile math is deterministic code. Digit sequences in translated files are validated or the fragment stays in English.

That split is the documentation analogue of dual control. The model is a fast first-year. The worksheet is the senior. If generation fails, the report returns to draft. We would rather resubmit a job than publish a half-written chapter as if it were signed.

Annexures are the file

Reviewers skip to the annexures. Accept/reject matrix, comparables description, comparables margins, search strategy, entity margin, search keywords: these are the pages that get photocopied into an assessment folder. The platform auto-builds more than twenty tables and six annexures from uploaded data so those pages exist when the narrative does.

Master File work is a different artefact. It should be consistent with every Local File the group will file, assembled from the group's own agreements rather than a generic multinational template. See the methodology page for how OECD toggles (intangibles, services, financing, restructurings) are encoded.

Finish in the editor, then export

A generated Local File is a draft. The in-app editor, section navigator, and formatting panel exist because house style and legal caution still belong to the signing partner. Export Word and PDF only after that pass. Translations (English, Arabic, German, French, Italian, Spanish) are copies. The English original is never replaced. If you edit English later, translations badge as outdated until you re-run them.

If you want this run on an entity you actually file, book a walkthrough. There is no public self-serve trial. Documentation that will be read in audit should be demonstrated on your fact pattern, not on a vendor's sample company.

House style is a control, not a font

Sixteen admin-editable prompt templates exist so a firm can change how the executive summary sounds without a developer. That is also a risk surface: limit who can edit prompts. A Local File that suddenly reads like a different firm is a review failure even if the range is correct.

Further reading: using the Local File as audit defense and documenting international transactions.