Automating business documents without losing your voice

Automating business documents without losing your voice means standardizing what repeats from one mandate to the next — structure, clauses, rates, legal blocks — while keeping bespoke what never repeats: the client's specific context and the recommendation addressed to them. The dividing line isn't a matter of taste; it separates what a firm decides once, outside of any mandate, from what gets decided inside a particular one. Standardizing both at once produces documents that are fast to output but interchangeable, which weakens the argument instead of strengthening it. Where that line falls also depends on the tool chosen, a trade-off set out in the article on custom software versus off-the-shelf SaaS.

An element can be standardized when its wording doesn't change from one client to the next, and stays bespoke the moment it does. Applied to contractual terms, that test yields a short, stable list: offer validity, payment terms, exclusions, confidentiality, intellectual property, limitation of liability, termination, governing law. The set OffrePro settles on has eight, with defaults that are a company decision rather than a negotiation: offer valid thirty days, fees payable net thirty days, confidentiality maintained twenty-four months after the engagement ends. Those values are edited per proposal and are not legal advice. The rest of the document plays out elsewhere, in the sections a winning business proposal has to contain.

A shared library alone won't hold that line, because it creates the opposite risk: fixing a clause at the company level would retroactively rewrite documents already negotiated. The pattern that solves it is the snapshot. When a proposal is created, the account's active clauses are copied into rows owned by that document — key, title, body, position — and it is the copy the responsible person reviews. Editing the liability clause of one proposal leaves the library untouched, and the copy runs only once: if the document already carries its clauses, the operation is skipped.

The same pattern applies to mandate phases, with a consequence expressed in numbers. Catalogue phases are copied per service line, along with their objective, steps, actions and minimum and maximum hour bounds; the proposal's phases then become independent of the catalogue. On an hourly line, the hour envelope drives the quantity: the summed bounds are reduced to their rounded midpoint, with a floor of one hour, and that quantity resynchronizes after every phase edit. The approach becomes the source of truth for hours, rather than a number typed in beside the text.

The reasoning specific to a mandate can't be standardized, and some sections gain nothing from being handed to a language model either. The timeline is the clearest case: its figures are quasi-contractual, so it is rendered deterministically from computed durations, because the model was observed in production dropping them. That boundary is worth carrying beyond any particular tooling: anything that commits the firm to a date, an amount or a scope should be computed, while anything that explains a recommendation should be written.

A document assembled automatically can come out complete and generic without a single visible error, and that is the limit worth facing squarely. When the model fails or omits a section, that section is filled with pre-written template content: the document stays valid and quietly loses its specificity. The counterweight is an equally mechanical rule in the other direction: a section edited by hand is never rewritten by a later regeneration. Document automation follows the same principle as the rest of generative AI adoption in a small business, where the rule is put this way: 'the system proposes, a person validates, then the system executes'.

An automated document has to stop being editable the moment it leaves the firm, otherwise the version discussed with the client and the version kept on file can drift apart untracked. The rule can be coded: priced line items and content stay editable only at the working stages — draft, generated, ready to send, change requested — and freeze as soon as the document is sent, accepted or declined. That requirement lines up with the definition of integrity set out in section 6 of Québec's Act to establish a legal framework for information technology: integrity is ensured when it can be verified that the information has not been altered, that it is kept in its entirety, and that the medium provides the required stability and permanence. Section 5 of the same Act adds that a document's legal value is neither increased nor diminished merely because a particular medium or technology was chosen. What that integrity requires of a signature affixed to the document is taken up in what makes an electronic signature valid in Québec.

Freezing the content isn't enough if the file that went out isn't the one everyone assumes. A PDF is a dated rendering of content that keeps moving while the document is being prepared: a corrected clause, an added service line, a changed lead, a rewritten introduction. Detection runs on timestamps — the PDF counts as stale as soon as the most recent change to a section, line item, responsible person, clause or introduction postdates the file itself — and that comparison beats human discipline, because it doesn't depend on what the sender happened to remember.

The last constraint is time. Revenu Québec requires records and supporting documents to be kept for the six years following the end of the last year to which they relate, on paper, electronic or microfilm media, which means that documents serving as supporting records have to outlive the template that produced them. The real gain from automation therefore isn't that 'the document writes itself': it's cycle time, consistency from one document to the next, and traceability. OffrePro, for instance, standardizes the structure of proposals, letters and invoices, attaches the legal blocks on the proposal side, and logs proposal sends in a timestamped event record — without generating the recommendation in place of the person responsible for the mandate, and without an equivalent record on the invoice side, where a failed send shows up only in the failed-jobs list.