The setup

The scenario is one every consultancy runs weekly. A senior consultant has spent 90 minutes with a client running a discovery workshop. There's a Teams recording, an auto-generated transcript, a page of typed notes, and a folder of supporting files (the RFP the client sent, a couple of internal policy documents, their org chart). The deliverable is a 20-page assessment report, due next Wednesday, in the firm's standard structure: executive summary, current-state assessment, findings, recommendations, roadmap, investment estimate.

Old way: four to eight hours of typing over two evenings. Every consultant reading this knows the shape of those evenings. It's not writing, it's transcoding, which is the boring bit that nobody trained for and nobody enjoys.

New way: twenty minutes from upload to a first draft you'd actually send to a partner for review. What follows is what happens in those twenty minutes, in a SkyDraft pilot workspace with the consultancy's template already set up.

Minute 0-2: upload and required-information check

Drag-and-drop the transcript, the notes, the RFP, and the policy documents into the sources panel. SkyDraft parses them (transcripts get speaker turns and rough timestamps preserved; PDFs and DOCX are extracted cleanly) and shows a required-information checklist tied to the assessment-report template.

The checklist has ten fields. The AI has pulled seven from the sources: client legal name, engagement scope, key stakeholders, meeting date, in-scope systems, current-state pain points, budget indication. Three remain unanswered: preferred delivery date for the final report, confidentiality classification, and whether the roadmap should include a "phase 3 optional" section. You type answers into all three in under a minute.

Minute 2-4: template scan + brand-voice check

Before generating anything, SkyDraft shows you the template outline for this document type, side-by-side with the workspace assets it'll apply: brand voice, glossary, company info. All three come from your workspace-level configuration set up months ago; you don't re-enter them per document. The side-panel confirms voice is "practical, plain-speaking, first-person plural" and glossary enforces the terms you care about ("engagement" not "project", "advisory" not "consulting", the specific framework names your firm uses).

This is the last human confirmation before drafting. Two seconds to scan, then hit Generate on section one.

Minute 4-6: executive summary

Section one, Executive Summary. Two paragraphs, drafted in about 40 seconds. The AI has picked up from the transcript that the client's primary concern is regulatory compliance in a specific system, that the CFO is the sponsor (not the CIO), and that the budget indication was "sub-$150k for a three-month engagement." All of that lands in the summary.

The summary ships with one clarification card: "The transcript mentions the client is 'considering' moving to a new platform in Q4. Should this Executive Summary flag the platform-migration angle, or treat it as out of scope for this assessment?" You answer: out of scope for now, flag as future work. Accept the answer; it becomes a pinned tile in the notes panel.

Minute 6-9: current-state assessment

Section two is the longest section in this document type (three to four pages typically). The AI drafts it from the transcript plus the internal policy documents, which describe the client's current architecture at the level the assessment needs. Draft comes in at 640 words in about 90 seconds, with three sub-sections: People, Process, Technology.

Two clarifications this time. First: "The transcript references 'the outage we had in April' but doesn't specify severity or which systems were affected. Include the reference or drop it?" You add context (major incident, main financial reporting system, 6 hours downtime). Second: "The policy document lists 12 in-scope systems; the transcript only discusses 4. Should the assessment cover all 12 or just the 4 discussed?" You answer: all 12, but detailed for the 4 discussed.

You accept both. Both answers become notes; both will propagate to Findings, Recommendations, and Roadmap.

Minute 9-13: findings

Section three. Findings is a repeating section in this template (one drafted entry per finding discovered from the sources), which means SkyDraft first proposes a list of findings extracted from the material, then drafts each one individually.

The proposed list has six findings. You look at it and spot that finding #5 ("inconsistent audit-log retention across the four systems") is really two findings glued together (audit-log gaps in system A, retention non-compliance in systems B/C). You split it, which is a one-click UI action. Now there are seven.

Then generate. Each finding drafts in 20-30 seconds. You read each one before advancing. Two of the drafts are clean. One (finding #3) makes a claim about the client's ISO 27001 status that isn't supported by the sources; the AI has flagged this exact concern as a clarification card, and offers to soften the claim or drop it. You drop it. Two more findings ship with clarifications you accept. The last two are accepted without changes.

Minute 13-16: recommendations

Section four is Recommendations. The AI sees all seven findings, all the accepted clarifications, the in-scope systems list, and the budget indication. Drafts recommendations by grouping (immediate, 90-day, 12-month), which is your firm's standard shape.

Two things happen here that only work because of the earlier sections. First: the "platform-migration" clarification you dropped in the Executive Summary propagates. Recommendations doesn't suggest anything about the future-state platform because you scoped it out. Second: the budget indication (sub $150k, three months) constrains what the AI proposes. You don't get an eight-workstream plan; you get three prioritised workstreams that fit the constraint, plus a "future work" callout for the rest.

This is the section where the value of a clarifications loop is most obvious. Without answers propagating forward, you would spend this section re-explaining the same constraints to the AI, or worse, missing that it had invented a different budget.

Minute 16-19: roadmap + investment estimate

Section five is Roadmap. Drafts as a table with three rows (immediate, 90-day, 12-month), each with dates, owners, dependencies. The dates come from the engagement start date you entered in the required-info checklist; the owners are placeholders like "Client lead: TBC, Assessor: Firm PM". Two clarifications, both small (Q3 or Q4 for the 12-month milestone; whether to show budget-per-phase or budget total). You answer both. Section drafts done.

Section six is Investment Estimate. The AI pulls the budget indication from the transcript, the three prioritised workstreams from the Recommendations, and the roadmap phasing, and drafts a summary table with phase-level totals and a bottom-line. It flags the usual caveats: "estimate subject to detailed scoping"; "T&M with not-to-exceed cap"; "excludes client-side effort". None of those numbers are specifically committed yet (that's what the SOW is for), but the estimate is coherent with the earlier sections. You'd be comfortable sending this to a partner for review.

Minute 19-20: review pass, DOCX export

Scroll top to bottom. Every section reviewed already in the act of drafting; the review pass is a once-over for tone consistency (the brand voice anchoring means it's mostly consistent, but you spot two paragraphs where the AI slipped into slightly formal register; two-second edit each). Export to DOCX. The template's cover page, header, footer, and heading numbering all applied. You send it as an email attachment to the senior partner: "First cut for review. Assumptions captured in the clarifications panel. Ready to walk through when you're free."

What the human owned

The AI wrote most of the words. The human still owned the parts that matter.

Every review gate. Nothing shipped without you seeing it. There was no "draft the whole report" button pressed in a moment of optimism.
Every clarification answer. Fifteen or so, across the report. Each one was a small judgment call about what to include, what to scope out, what to soften. That's the actual expertise being applied, and it's exactly the part that doesn't transfer to a chat window.
The one substantive correction (splitting the audit-log finding, dropping the unsupported ISO 27001 claim). The AI wasn't trying to hide either issue, but you were the person who caught them and made the call.
The framing decisions. Whether the platform migration was in-scope, whether the assessment covered all twelve systems or just the four discussed, whether recommendations should assume T&M or fixed-fee. The AI drafted around the answers; the answers came from you.

What went wrong (yes, some things did)

This isn't a fairytale. A few things needed cleanup:

The AI initially over-quantified the audit-log finding ("resulting in $2.4M annual compliance risk"), a number nowhere in the sources. The clarification caught it, you dropped the number.
One recommendation drifted into slightly generic consultant-speak ("establish a centre of excellence"). You rewrote a phrase; every subsequent section respected the edit because the workflow treats your edits as context.
The DOCX export needed a manual page break before the roadmap table (a template polish issue, not a content one).

Total human intervention time: about six minutes of the twenty. Not "AI wrote it and I clicked send". More like "AI transcoded the raw material, and I made the judgment calls where judgment was needed."

The pattern to remember

Twenty minutes is not the point. The point is which minutes disappeared. The four to eight hours of transcoding are gone. The judgment work is preserved. The pattern is the same across engagement reports, status updates, technical assessments, and technical briefs. The pattern is: AI drafts, clarifications surface, human answers, drafts pick up the answers, document ships.

For the full mechanical walkthrough of templates, workspace assets, and every moving piece of the generation flow, see how it works. For the specific use cases where this pattern lands hardest, see use cases.

Where to start

Bring us one of your recent engagement reports and one of the transcripts that produced it. We'll set up the template with you in a working session, then run the transcript through the workflow live. The first engagement report you produce this way will be the one you use as the benchmark for every one after.

SkyDraft pilot workspaces are open. Free during pilot, founder-led setup, no credit card. Pilot pricing locks in when standard pricing publishes.