Project Overview
A US marketing agency was spending far more time writing new business proposals than the work warranted. Every proposal was built from scratch despite most of them sharing 60 percent of their content: agency credentials, process descriptions, case study summaries, and team bios. Account managers who should have been focused on client strategy were instead spending four to six hours per proposal on repetitive writing work. We fine-tuned a lightweight LLM on their best-performing historical proposals and built a simple drafting interface that cut that time to under 30 minutes without changing the voice or win rate.
The Challenge
The agency had no proposal template or content library to speak of. Each account manager wrote proposals in their own style, pulling from memory or searching through old files for content they had written before. The result was inconsistent quality across proposals, slower turnaround than the agency wanted, and creative energy spent on formatting and boilerplate rather than client-specific strategy.
- Proposal drafting took four to six hours per document, with most of that time spent on content that was largely repeated across submissions
- No central library of approved content: account managers recreated the same sections from memory each time
- Quality and tone varied significantly between account managers, creating inconsistency in how the agency presented itself to prospective clients
- The team was skeptical of AI tools and would revert to writing from scratch if outputs did not match their voice closely enough to be trusted
Our Solution
We fine-tuned a lightweight LLM on the agency’s top-performing historical proposals and built a simple interface where account managers entered the client brief and received a structured first draft within minutes. The system was designed to produce drafts that sounded like the agency, not like generic AI output, which was the condition for adoption.
- Proposal Data Preparation: We curated and structured 80 of the agency’s highest-performing historical proposals, annotating them by client type, service offering, and win outcome to give the fine-tuning process quality signal rather than just volume.
- Fine-Tuning for Voice: We fine-tuned a compact instruction-following model on the curated proposal dataset using supervised fine-tuning, with particular attention to replicating the agency’s sentence structure, tone, and section ordering rather than just their factual content.
- Brief-to-Draft Interface: Account managers entered a short client brief including industry, objective, budget range, and key differentiators, and received a complete structured first draft covering credentials, approach, timeline, and case study selection within minutes.
- Edit-First Workflow: The interface was positioned as a starting point rather than a finished product. Account managers were expected to edit and customize the draft, reducing resistance from those who wanted to maintain creative control.
Technical Approach
The fine-tuning used a compact base model selected for its instruction-following quality and low inference cost, allowing the agency to run drafts without significant per-query cost. The model was hosted privately so that client brief information did not pass through shared public endpoints. We ran a series of A/B comparisons during development where account managers evaluated AI drafts alongside their own without knowing which was which, and used those ratings to guide further fine-tuning iterations until output quality met the adoption threshold the team required.
Results & Impact
- Proposal drafting time reduced from approximately five hours to under 30 minutes, freeing account managers for client strategy and relationship work
- Win rate maintained in the quarter following deployment, confirming that the quality of proposals was not degraded by the shift to AI-assisted drafting
- Proposal voice and structure became consistent across the team for the first time, as all drafts started from a model trained on the agency’s best historical work
- Account managers reported spending recovered time on client discovery calls and strategy development rather than writing
- The fine-tuning process surfaced the agency’s best-performing content in a curated, organized form, giving the team a reference library they had not previously had
Lessons Learned
The skepticism from the team was the most important constraint to design around. Account managers with strong voices and good win rates had no reason to use a tool they did not trust. The A/B evaluation process that shaped the fine-tuning also served as a trust-building exercise: when account managers saw outputs they themselves could not distinguish from their own writing, resistance dropped significantly. Getting the team involved in evaluating outputs early was more valuable than any technical optimization we could have run in isolation.
Positioning the output as a first draft that required editing also mattered. Framing the tool as replacing their writing would have met resistance. Framing it as eliminating the blank page problem, specifically the most time-consuming and least valuable part of the process, was accurate and well received. The account managers who adopted the tool most enthusiastically were those who had the most proposals to write and the least time for the work they actually wanted to be doing.