Almost every article about AI and RFPs sells the same promise: answer them faster. The benchmark data does not support that as a goal, and the gap between the promise and the numbers is the most useful thing to understand before you automate any of it.
The adoption paradox
Loopio's seventh annual RFP Response Trends and Benchmarks Report, drawn from more than 1,500 teams worldwide, contains three numbers that only make sense together.
| Measure | Latest | Prior year | Direction |
|---|---|---|---|
| Teams that have used generative AI in the RFP process | 79% | 68% | Up sharply |
| Average hours spent per RFP response | 33 | 35 | Down slightly |
| Average advancement rate to shortlist | 46% | 54% | Down sharply |
| Average RFPs submitted per organisation, per year | 166 | 153 | Up |
| Share of received RFPs a team answers | 55% | 63% | Down |
Adoption jumped eleven points and the time saved was two hours per response. Meanwhile the rate at which those responses reached a shortlist fell eight points. We are not going to claim AI caused the decline, because the report does not show that and plenty else moved in the same window, including a rise in volume that would depress advancement on its own. The honest reading is narrower and more useful: whatever teams automated, it was not the part that decides outcomes.
One more figure sharpens it. Teams with higher win rates spent about 35 hours per submission against the 33-hour average. Longer, not shorter. If more time on a response correlates with winning it, then an agent whose only contribution is finishing sooner is optimising against the thing you want.
That reframes the goal. The agent is not there to reduce the 33 hours to 20. It is there to move hours out of retrieval and into argument, so the total can stay where it is and land somewhere better.
Split the questionnaire before you automate it
Every RFP is two documents wearing one cover. Separating them is the whole design decision, and it takes about twenty minutes on a past bid.
| The recall half | The argument half | |
|---|---|---|
| Typical content | Security questionnaires, compliance and certification questions, integration lists, SLA and uptime terms, company facts, standard capability descriptions | Executive summary, win themes, why-us positioning, pricing narrative, implementation plan tailored to their environment, references chosen for this buyer |
| Where the answer lives | Already written, somewhere: a past response, a policy document, a product page | Nowhere yet. It is made from what you know about this specific account |
| What makes it hard | Volume and retrieval. Finding the approved version and reformatting it to their template | Judgment. Deciding what to emphasise and what to concede |
| Cost of a wrong answer | Contractual. You committed to something untrue | Competitive. You sounded like everyone else |
| Agent role | Draft it, cite where each answer came from, flag anything with no approved source | Assemble the raw material, then stop |
The recall half is usually the larger by question count and the smaller by influence. That asymmetry is why automating it is worth doing and why doing only that will not move your win rate by itself. It buys back hours. What you do with the hours is the strategy.
What to hand the agent
The instruction that works is much narrower than "respond to this RFP". Specify the sources, the exclusions, and what the agent must do when it is unsure.
| Element | What to specify |
|---|---|
| Source set | A curated answer library, current product and security documentation, and past responses a human marked reusable. Nothing else, and specifically not a raw file share |
| Attribution | Every drafted answer carries the source it came from and the date that source was last approved |
| Exclusions | Pricing, legal terms, contractual commitments, security claims not backed by a current certificate, and anything about roadmap |
| Format | Match the buyer's template exactly, including their question numbering and any word limits |
| Uncertainty rule | If no approved source answers the question, say so and name the person who should. Never compose a plausible answer from adjacent material |
| Output | A draft plus two lists: questions answered with high confidence, and questions requiring a human |
The uncertainty rule is the one teams leave out, and it is the one that decides whether the whole thing is safe. An RFP question like "describe your data residency guarantees for the EU" has a real answer that either exists in a signed document or does not. An agent asked to answer it without an approved source will produce something reasonable-sounding and wrong, and unlike a bad blog paragraph, that answer may end up attached to a contract. Getting a blank with a name next to it is a better outcome than getting prose.
This is the same permissions-and-approval discipline that applies to any agent touching systems of record, and it is worth reading AI agent security best practices and how to add a human approval step to an agent before you wire anything to a live document store.
The higher-leverage job: qualifying
Here is the part the vendor pages skip, because there is no software licence in it. Organisations now answer 55% of the RFPs they receive, down from 63% the year before, while submitting more in absolute terms, 166 a year against 153. Teams are getting choosier, and the report ties the bandwidth squeeze directly to rising volume.
Which means the highest-return thing an agent can do on an RFP is often not drafting it. It is reading the request the day it arrives and preparing the bid-or-decline decision:
- Extract the hard requirements, deadlines and submission format, so nothing is discovered in the final week.
- Flag disqualifiers immediately: a certification you do not hold, a reference class you cannot supply, a jurisdiction you cannot serve.
- Summarise how close this request sits to deals you have actually won, using your own past outcomes rather than a similarity score on the text.
- Estimate the response effort by counting questions that have no approved answer in the library, which is the real predictor of how painful the bid will be.
A human still decides. But deciding well on the first day, with the disqualifiers visible, is worth more than shaving four hours off a response you should not have written. Teams that already run structured evaluation on the buying side will recognise the shape of this; it is the mirror image of what we described in AI agents for procurement teams and the AI agent procurement checklist.
A working setup, end to end
The RFP lands. The agent parses it, produces the qualifying brief above, and posts it where the deal owner sees it that day. If the answer is bid, the agent moves to the recall half: it maps each question to the answer library, drafts what it can with attribution, formats to the buyer's template, and hands back the two lists.
A human then spends their time in the right place. The flagged questions get real answers from the people who own them. The argument half gets written by someone who has spoken to the buyer. The reclaimed hours go into the executive summary and the win themes rather than into submitting sooner, which is the whole point of the exercise.
The pattern generalises to the rest of the deal cycle, and the pieces connect: sales call follow-up keeps the account knowledge current enough for the argument half to be specific, proposal follow-up handles what happens after submission, and the wider set of sales agents covers where else this fits. For the retrieval side specifically, knowledge search across Slack deals with the common case where the approved answer exists only in a thread.
What goes wrong
The library is the product, and it rots. An agent pointed at an answer library is only as good as the library's last review. The failure is silent: a security answer approved eighteen months ago is still returned confidently today. Put a review date on every entry and have the agent surface answers whose source is older than your threshold instead of using them quietly.
Reformatting is underestimated. Buyers send questionnaires as spreadsheets, portals, and locked documents with word limits per field. A draft that ignores the template creates a second manual pass that can cost more than the drafting saved. Specify the output format as strictly as the content.
Volume becomes the metric. The easiest way to show the agent is working is to answer more RFPs. The benchmark data suggests that is precisely the wrong response, given advancement rates fell while volume rose. Measure the share of responses that reach a shortlist, not the count submitted.
The agent writes the differentiator. If the executive summary starts arriving pre-written, someone has quietly moved the boundary. That section is where the bid is won or lost, and a generated version reads like every other generated version in the buyer's pile.
None of these are reasons to skip automation. They are reasons to automate the recall half deliberately, keep the argument half human, and spend the saved hours where the data says the difference is made.
Frequently asked questions
Can an AI agent write a full RFP response?
It can produce a complete first draft, and for the repeated compliance and capability questions that draft is often close to final. It should not write the sections that decide the bid: the executive summary, the win themes, the pricing narrative, and anything that makes a claim about your company you would have to defend in a contract. Those need a human who knows the account.
How much time does an AI agent actually save on an RFP?
The realistic saving is on retrieval and first-draft assembly, which is the bulk of the repeated questions. Loopio's seventh annual benchmarks report puts the average response at 33 hours, down from 35 the previous year, with generative AI adoption at 79%. That pattern suggests AI is compressing hours rather than eliminating the response, so plan for a meaningful reduction in grind work, not for the RFP to answer itself.
Does using AI on RFPs improve win rates?
Not on its own. In Loopio's 2026 data, generative AI adoption rose from 68% to 79% while the average advancement rate fell from 54% to 46%, and teams that spent longer per response, 35 hours against the 33-hour average, won more. Speed alone does not win bids. The gain comes from spending reclaimed hours on the differentiating sections rather than submitting more responses faster.
What should the agent be allowed to access?
A curated answer library, approved product documentation, and past submitted responses that a human has marked as reusable. Do not point it at a raw file share. Most bad RFP answers come from an agent confidently reusing an outdated or never-approved claim, so the source set matters more than the model.
What does it cost to run an RFP agent?
On Gravity the free tier is $0 a month for one agent, which is enough to test the pattern on a single questionnaire. Paid plans start at $20 a month and include $20 of usage, with extra usage available beyond the plan. Dedicated RFP response platforms sit in a different bracket and are usually quoted per seat with an annual commitment.
Should the agent decide which RFPs to bid on?
It should prepare the decision, not make it. An agent can extract the requirements, flag disqualifiers such as certifications you do not hold, and summarise how close the request sits to deals you have won. A human still decides. Loopio reports organisations now answer 55% of the RFPs they receive, down from 63%, so the qualifying step is where teams are already finding leverage.
Sources
- Loopio, "Average RFP Win Rates & More: 38 Proposal Statistics to Know in 2026", drawing on the seventh annual RFP Response Trends & Benchmarks Report of 1,500+ teams, published 26 March 2026, checked 23 August 2026, loopio.com. Source for the 79% generative AI adoption figure (from 68%), 33 hours per response (from 35), 35 hours among higher-winning teams, 46% advancement rate (from 54%), 166 RFPs submitted annually (from 153), and the 55% response rate (from 63%).
- Loopio, "RFP Response Trends & Benchmarks Report", the underlying annual study, loopio.com/trends-report, checked 23 August 2026.
- Gravity, "How it works" and pricing, gravity.fast, checked 23 August 2026.
Want to test this on one questionnaire before committing to a process? Join the Gravity waitlist and start with the free tier.