The rollout that quietly became a shortcut

A twenty-person advisory firm buys seats on an AI tool in January. By March, three people use it daily, eleven use it occasionally, and the rest have gone back to copying last quarter's document. Nobody cancelled anything. There was no decision. The tool just drifted from “how we draft” to “a thing you can paste a paragraph into when you're stuck on a sentence.”

Ask the eleven occasional users what happened and you get a version of the same answer. Every document starts from nothing. You open the tool, and before you can get useful output you have to explain the document type, remind it how the firm writes, correct the terminology twice, paste the source material, and then read the result closely enough to catch what it made up. That is fifteen minutes of setup before the first useful sentence, repeated per document, per person, forever. It is cheaper to copy the old file.

The failure here is not the model. It is that the firm bought a generation engine and never built the thing the engine runs against. A working AI document workflow needs persistent state: assets that live above any single document, get set up once, and apply to everything drafted afterwards. Five of them, in practice. Here is what each one is, what breaks without it, and roughly what it costs to capture.

1. The template: the shape of the document

Not a prompt, and not a blank Word file with your logo on it. A template in this sense is a structured object: the section list in order, sub-sections, a guideline for each section describing what it has to accomplish, and the shape each section takes.

That last part matters more than people expect. A methodology section is prose. A pricing schedule is a table with fixed columns. A deliverables list is neither: it is one drafted entry per item, where the items themselves have to be discovered from the source material first. SkyDraft templates support all three (prose, table, repeating), and getting the shape right is what stops an AI from writing four paragraphs of narrative where the client asked for a compliance matrix.

Without it: every draft has a structure the AI proposed rather than the structure your client, your panel, or your regulator requires. You spend the first editing pass moving content between headings that should never have existed.

Cost to capture: upload one good existing document of that type. Structure extraction is a single pass. A partner then spends twenty minutes correcting the guidelines, because extraction gets the skeleton right and the intent approximately right. We wrote at length about why this artefact is worth the twenty minutes in templates as institutional memory.

2. Brand voice: how the firm writes

Sentence length. Formality. Whether you hedge or commit. Whether you open a recommendation with the recommendation or with the reasoning. Whether your headings are noun phrases or full sentences. Firms carry these decisions implicitly, and the consistency is a real part of what a client is paying a premium for.

Generic AI has a voice, and it is the average of the internet: balanced, sub-clause heavy, allergic to a flat declarative statement. It is not bad writing. It is nobody's writing, which for a firm that has spent fifteen years sounding like itself is worse.

Without it: the document reads fine and reads wrong. Reviewers cannot articulate the problem, so they rewrite sentences one at a time, which is the slowest possible way to fix a systemic issue. Two consultants using the same tool produce drafts that do not sound like the same firm.

Cost to capture: effectively free if you do it by extraction rather than by writing a style guide. Voice comes out of the same upload pass as the template. It lives at workspace level, so every document drafted afterwards inherits it without anyone pasting anything into a prompt.

3. The glossary: the words that are not negotiable

Every firm has thirty to fifty terms where the model's default is wrong. You run engagements, not projects. Your discovery phase is Phase 0. Your delivery framework has a name and it takes a capital letter. The client is the Department, not the customer. Your remediation levels are P1 to P4 and a P2 means something specific that no model can guess.

A glossary is the cheapest asset on this list and the one that produces the most visible improvement per minute invested, because terminology errors are the ones a client notices immediately. A proposal that calls the engagement a “project” four times tells the reader the document was assembled rather than written.

Without it: a find-and-replace pass on every document, done by hand, forgotten under deadline. The errors that survive are the ones in the sections nobody reread.

Cost to capture: thirty minutes for the first pass, then it grows by one or two entries a month as edge cases surface. Update it once and every subsequent draft picks up the change.

4. Company identity: who the document is written as

This is the asset most firms skip, because it looks like a duplicate of brand voice. It is not. Voice is how you write. Identity is who is writing: what your firm does, the sectors you work in, the size and shape of your team, how you position against the alternatives, and what you are entitled to claim.

The distinction shows up the moment a section has to make an assertion. “Our team has delivered comparable programmes across the sector” is either true, false, or unverifiable, and the model has no way to know which unless something tells it. Get the voice perfect and a draft with no identity still argues from nowhere: correctly toned sentences making claims the firm cannot stand behind, in a document that will be read by a procurement panel scoring exactly those claims. We pulled this apart properly in why your firm needs an AI-author identity.

Without it: generic capability statements, and a slow leak of unsupportable claims into documents that get signed.

Cost to capture: one focused conversation with whoever writes the firm's capability statements, captured as a workspace asset alongside voice and glossary.

5. The required-information checklist

The fifth asset is the one that is hardest to see, because it is not about output at all. It is the list of things you always need to know before drafting this document type can honestly begin. For a proposal: client name, decision maker, budget envelope, timeline, incumbent, evaluation criteria. For an engagement report: scope boundary, period covered, who was interviewed, which systems were in scope, what was explicitly excluded.

This lives on the template, but treat it as its own asset because it is maintained separately and it does separate work. It is the difference between an AI that invents a plausible timeline and an AI that stops and asks you for the real one. Missing inputs do not disappear when you skip the checklist. They get filled with confident guesses that are indistinguishable from facts by the time the document reaches review.

The checklist is also what makes the clarifications loop worth having. Answered required information plus per-section clarification questions means the gaps get surfaced at the beginning and at the point of use, rather than discovered by a reviewer on the last afternoon.

Without it: the fabrication problem, arriving as specifics. Dates, headcounts, durations, and prior-engagement references that read exactly like the ones you supplied.

Cost to capture: fifteen minutes, and it comes partly pre-filled from the same document upload as the template.

What the five assets buy you together

Individually each one removes a category of rework. Together they change what a document costs to start. The fifteen minutes of per-document, per-person setup goes to zero, because the setup already happened at workspace level and every draft inherits it.

That is the part that decides whether a rollout survives past March. A tool that requires setup per document gets used by the three people who enjoy tooling. A tool where the setup already exists gets used by the person with a tender due Thursday, because opening it is now faster than copying the old file. The adoption curve follows the friction, not the quality of the output.

The second effect is on your junior staff. When the firm standard lives in five assets rather than in three senior heads, a second-year consultant drafting their first engagement report produces something that passes review, not because they suddenly know what good looks like, but because the workflow does. That is a margin change, and it is the one that gets leadership attention.

Two failure modes worth naming

Building all five before drafting anything. The assets improve fastest against real documents. Capture rough versions in an afternoon, draft one live engagement, and fix what the draft exposes. A template refined against three real documents beats one workshopped for two weeks.

Letting them go stale. These are living assets. A glossary that has not changed in a year is not stable, it is abandoned, and the terminology drift is being corrected by hand in every document instead. The maintenance cost is small but it is not zero, and it needs an owner.

Where to start this week

Pick the document type you produce most often. Find the best recent example of it, the one a partner would point to. Upload it, let the structure, voice, glossary, and required-information checklist come out in one pass, then spend twenty minutes with someone senior correcting what the extraction got approximately right. Add the company identity by hand. That is your afternoon.

Then run the next real engagement through it. Not a test document, a real one with a deadline. The assets are only trustworthy once they have survived contact with a document somebody actually has to send. How it works covers the mechanics in detail, and use cases shows the document types this shape fits best.