Two days of proposal work, done before the coffee
A governed skill library turned a two-day proposal draft into a fifteen-minute one, demonstrated on the firm's own portfolio.
- Services
- AI Foundations · Advisory, platform and build
- Period
- Jul – Sep 2026, a fixed forty-hour block
- Firm
- About twenty-five staff across Melbourne and Sydney. Microsoft throughout, no internal technology team.
Everyone was using AI, on their own terms
The firm runs on proposals. Partners put proposal writing at roughly a fifth of the firm's capacity: two to three consultant-days each, over about a week and a half, with up to seven working days of setup before anyone wrote a word.
A baseline survey found every member of staff using at least one AI tool, eighty-three percent daily or more, almost all self-taught. Trust in the output averaged 2.4 out of 5. Microsoft Copilot was the only sanctioned tool, but people had opened personal accounts on whatever else suited them.
Copilot could not handle the firm's own material. It could not tell a proposal from a delivered project, so it offered case studies for work the firm never won. Finding past proposals was the top complaint of ninety-three percent of the people who write them: there is no single source of truth
.
Partners were also finding AI slop in client work: machine prose they could spot, and juniors submitting output they could not explain. The Managing Partner set the limit. You do not have permission to let the tool own the narrative.
Better proposals, not faster ones
A senior engagement manager framed the test at kickoff. Not faster proposals or less-effort proposals — how do we do better proposals?
That ruled out a push-button proposal generator.
So we measured first: a firm-wide survey, a workshop mapping the proposal process end to end, and a systems session across SharePoint, Salesforce and the slide library. Then one recommendation report, and partner approval before anything was built.
The platform decision was to standardise on one sanctioned tool and keep Copilot as the opt-out tier, because some government clients prohibit generative AI outright. The build then followed the firm's existing proposal process, stage by stage.
A skill library the firm owns
The work sits in the firm's own Claude workspace, behind Microsoft single sign-on, distributed as a plugin from a GitHub organisation the firm controls. Fifteen curated proposals are the baseline the skills are tested against.
- Briefing
- Reads the RFP, the emails and the transcripts, creates the SharePoint folder, interviews the team, and returns a briefing document with win themes and a slide storyboard.
- Assembly
- Word and PowerPoint builders that edit the firm's real slides rather than generating new ones.
- Quality check
- An agent that reads a draft for the tells of machine writing before it reaches a client.
- Context
- A won and lost layer drawn from Salesforce, so the system can tell a proposal from delivered work.
- Distribution
- A plugin marketplace in the firm's own GitHub organisation, synced to its workspace, with a custodian named inside the firm.
Approved in eighteen days, in use before handover
From the kickoff workshop to partner approval of the target architecture and the move to a single core AI platform.
A content-complete, on-brand PowerPoint draft, against a baseline of two to three consultant-days. The working group judged it well past half of client-ready.
From kickoff to the plugin going live in the firm's own marketplace, with an editable handover artefact.
The whole firm was provisioned and trained in one all-hands session, and five backlog items were delivered at no extra cost. The Managing Partner had adopted the workflow for his own work before handover.
The Partners met yesterday and agreed to press go on the target architecture, including moving to Claude as our preferred, core in-house AI toolkit.
Have a problem shaped like this one?
Fixed-scope discovery to start.
Strata: H&CO — Unsplash