AI for RFP questions that need more than one contributor

Two requests under one question number

A client asks in an RFP: „Where are backups stored, and who approves their deletion?”. The infrastructure team supplies the location, and the process owner runs the approval of deletions. Assigning the whole item to the topic „backups” hides the need for a second person, so the team describes the location and the coordinator finds the gap only when the bid is being put together.

A good proposal shows the passage about the location and the passage about approval separately, each with a topic. Both parts stay linked to the full question and its number, and the original wording is visible. We keep recognising the topic apart from staffing: when the person responsible for a process changes, the company updates the assignment without retraining the model.

What a simpler change gives, and what training adds

In this count, a topic list with a description of responsibilities gives most of the effect for a fraction of the cost. Training makes sense at a larger scale or when the remaining errors are expensive. The most expensive error is a second request discovered while the bid is being put together, which is why we count missed parts and wrongly assigned topics separately.

The set of training examples has to be maintained: when the process, the naming or the base model changes, the evaluation is repeated on the same test set. We put that cost in the quote together with the training itself.

What to check before you adapt a model

First check the bid tool: whether it lets you assign several co-authors to one item or link sub-tasks to the client's question. For recurring questions, a shared topic list with a description of responsibilities helps. When splitting long sentences still needs a second read, compare a ready model with the current method. AWS describes custom document classification as defining the classes, training a classifier and applying it to documents. Gathering input from known authors is covered in the post about technical clarifications in a tender.

Organise questions in the bid team's tool

The comparison uses different requests from those used for adaptation: items that need several authors, and simple questions from one area. The coordinator counts how many proposals they correct before assigning authors. We adapt models as part of our private LLMs, and we build the app where an employee sees the result next to its source as part of our AI apps. For the free analysis, a few examples of an error that keeps coming back and a note of who corrects it today are enough.