The chatbot we talked them out of building
Losing deals to response time, with every proposal assembled by hand from prior engagements.
- Turnaround
- 5.5 days5 hours-96%
- Hours returned
- 058+58
- Win rate
- 1.0x1.11x+11%
Six months post launch against the prior six. Source: CRM proposal log, May 2025
Context
A consulting firm losing competitive deals on response time rather than on the quality of the work. Every proposal was assembled by hand from 400 plus past engagements stored in SharePoint.
The constraint
The thing that ruled out the obvious answer.
All prior client work product was under confidentiality. Nothing could leave the firm's tenant, which eliminated every hosted retrieval product they had already trialled.
Diagnosis
This was not a search problem. Analysts knew exactly which past engagement to draw from, usually within minutes. The days were going into reformatting, re-scoping and rewriting boilerplate to match a new client's context.
What we rejected, and what we chose instead
Any firm can describe what it built. The useful record is the option that was on the table and did not survive.
Rejected
A chat interface over the document library
It would have made analysts do the work of prompting to get back what they already knew. Chat solves discovery, and discovery was not the bottleneck.
Chosen
A structured generator with a partner review gate
Pick engagement type, client size and scope. It assembles a draft from approved prior sections. The analyst edits rather than composes, and a partner approves before anything leaves.
What was actually implemented
- 01Chroma running inside the firm's own VPC, embeddings generated locally
- 02Section library curated from prior work, each section explicitly approved for reuse
- 03Drafting through a zero retention endpoint, no client text used for training
- 04Partner review queue as a hard gate before any document can be exported
- 05Every generated document records which source sections it drew from
Measured, not estimated
Turnaround fell from 5.5 days to five hours. The firm measured 58 analyst hours returned per month and an eleven percent rise in win rate on competitive bids.
Six months post launch against the prior six. Source: CRM proposal log, May 2025
The most valuable recommendation was what not to automate.
More from ai and automation
AI and automation
Most AI projects fail because they add a chat box next to the problem instead of removing the step that costs the hours. We start from the workflow.