Deep DiveNo. 01
RFPs: from 3 weeks to 8 minutes
How a software company is using two agents and a knowledge library to prepare more proposals
Gonzalo GilgesHead of Practical AI
- Sent
- Reading time
- 5 min
Preparing a proposal used to take between three and four weeks. The solution we developed produces a first draft in eight minutes, with preliminary answers for 81% of the questions during the second testing period.
That 81% does not represent the solution’s ceiling. As we incorporate validated corrections and define more precisely the knowledge required for each service, we estimate that coverage could reach 95–97%. For tightly defined services, the potential may be higher still. The team retains control of the review and the commercial decision.
The cost of an RFP is not limited to the hours required to answer it. It also includes the contracts a company cannot compete for while its people search for documents and coordinate approvals.
You may have the product and the experience to win, but not the operational capacity to submit a bid.
At Fluency Labs, we organised the solution around two agents: one to find and qualify opportunities, and another to prepare draft answers. The commercial opportunity is to pursue more relevant proposals with the same operating structure.
Fig. 1Two agents, with human decisions
- 01 Agent 1
Finds and qualifies opportunities. In development.
- 02 Team
Decides which RFPs to pursue.
- 03 Agent 2
Consults the library and proposes answers. In beta.
- 04 Team
Reviews, corrects and completes.
- 05 Draft
The team approves the offer and submits it.
01
First agent: find opportunities worth the effort
Tenders appear across different platforms. Finding them, reading the conditions and checking whether they fit requires daily work. Automating that search takes more than keywords: the system needs to understand what the company sells, where it operates and which requirements it can meet.
Using those criteria, the search agent is designed to review sources, remove duplicates and collect the information needed to assess each opportunity. If the description does not clarify a requirement, it may need to consult an annex before recommending how to proceed.
The intended output is a reasoned shortlist: why each contract may be a fit, when it closes and what still needs to be checked. The team retains the decision to bid.
02
Second agent: prepare answers using company knowledge
Once an opportunity has been selected, the agent identifies the questions and submission requirements: files, formats and sections. It then consults a library of the company’s previous answers.
The AI searches by meaning: two buyers may ask the same thing in different words. But finding a similar answer does not guarantee that it is suitable. A second AI review checks whether the precedent actually answers the question.
The system distinguishes three situations:
- There is a suitable precedent: it presents it for review.
- There is enough information to draft an answer: it prepares a proposal and labels it as such.
- There is not enough evidence: it identifies what the team needs to provide.
The reviewer can compare the question, the precedent and the proposed answer; then accept, correct, reject or complete it. Instead of rebuilding every answer by searching through files, the team starts with prepared content and clearly identified gaps.
That reviewed content is used to prepare the document draft. The initial focus is on answers to requirements. Pricing, commercial commitments and approval of the complete offer remain with the team.
03
What the first tests show
During the first testing period of the MVP, the AI answered or proposed answers for an average of 68% of the questions. During the second, it reached 81%: an improvement of 13 percentage points.
Each period included several tests. The figure measures coverage of preliminary answers, not approved answers or proposals ready to submit. Processing took an average of eight minutes, before final review.
The move from 68% to 81% shows where the system is heading. Each validated correction strengthens the library and closes gaps for the next proposal. The ceiling depends on scope: the more clearly defined the service, the greater the proportion of questions the solution can prepare.
In figures
- Average processing time for a first draft, before final review. Preparing a proposal used to take between three and four weeks.
- 8 minutes
- Questions the AI answered or proposed answers for, on average: 68% in the first testing period of the MVP, 81% in the second.
- 68% to 81% +13 points
The figure measures coverage of preliminary answers, not approved answers or proposals ready to submit.
04
The library is part of the solution
When an answer needs correcting, the team must understand what was missing: product information, a rule or an undocumented criterion. Correcting only the current proposal leaves the same problem unresolved for the next RFP.
The team adds validated answers and their sources, replaces outdated information and clarifies the rules: which product each answer applies to, under which conditions and when a person needs to be consulted.
We then test again. We measure how many answers are usable and how much correction they require. In this way, a specialist’s knowledge becomes available to other proposals, even when a buyer phrases the question differently.
In a second part, we will return with results on how the solution performs and how it affects commercial outcomes: opportunities pursued, proposals submitted, contracts won and revenue.
The objective is not simply to prepare answers faster. It is to use that efficiency to compete for more opportunities and convert them into new revenue.
See you in the next issue.