What a month of studying proposal automation taught me about research, practitioners and staying connected to reality
Over the past month, I have spent a disproportionate amount of time trying to understand proposal automation.
I did not begin there completely by accident. I have spent years writing research proposals, coordinating technical work and trying to obtain funding. When I discovered the commercial RFP software ecosystem—Loopio, Responsive, AutogenAI, GovDash and others—it initially looked remarkably close to work I already understood.
That assumption turned out to be only partly correct.
Commercial RFPs and research grants share several operational similarities, but they sit inside very different economic systems. More importantly, the deeper I investigated commercial proposal work, the more I realised that my biggest research problem was becoming something else entirely.
It was not simply:
What should proposal software automate?
It was:
How do I prevent myself from constructing plausible theories about other people's problems without remaining sufficiently connected to the people actually experiencing them?
That question has changed the direction of my research considerably.
1. I Started From A Domain I Thought I Understood
Research proposals are familiar territory for me.
A Principal Investigator may need to identify a funding opportunity, understand the call, develop the scientific direction, find collaborators, determine the evidence required, assemble resources, write the proposal and survive multiple rounds of review. In my own work, proposal authoring is only one component of a much larger process involving decisions about what research is worth pursuing, which technologies deserve investment and what results need to be demonstrated next.
This initially made commercial RFP automation seem like an adjacent problem.
There are obvious parallels. Both environments involve requirements, deadlines, multiple contributors, previous material, technical judgment and substantial amounts of highly qualified human attention.
But the economics differ in an important way.
A company competing for a commercial contract may be able to connect proposal activity relatively directly to revenue. A $1 million contract remains a commercial opportunity for the bidder. A $1 million research grant is not equivalent to $1 million of economic value flowing to the PI who prepared the proposal; the funding exists to pay salaries, equipment, experiments and the research itself.
That distinction weakened my original assumption that the same proposal-software logic should naturally transfer between the two environments.
My earlier investigation also found that specialist proposal-authoring AI remains much less institutionally visible in research than response-management software has become in commercial bidding. There are genuine examples—such as the Colorado School of Mines testing Intellectible and Duke developing its internal Devil's Advocate proposal-review tool—but public evidence still looks more fragmented than the mature commercial RFP software category. (Colorado School of Mines, Duke University)
That earlier work, especially Why Is Proposal AI So Invisible In Research, made commercial RFPs worth investigating in their own right.
2. The Obvious AI Opportunity Became Less Obvious
At first, proposal automation looks technologically attractive.
A modern AI system can already read long solicitations, extract requirements, summarize documents, search previous material, draft answers, identify inconsistencies and perform various forms of compliance review. Specialist products go further by combining organisational knowledge, collaboration, governance and workflow.
Loopio says its platform supports reusable content and response workflows across large numbers of organisations. Responsive similarly combines organisational knowledge, response generation, project management and integrations with enterprise systems. (Loopio, Responsive)
The commercial market is clearly real. GovDash, for example, announced a $30 million Series B in January 2026 after reporting substantial growth in both revenue and customer count. (GovDash)
But the existence of a market did not answer the question that mattered to me:
Where is the remaining problem that is both important and poorly solved?
The more practitioner discussions I read, the less convinced I became that generic proposal generation was that problem.
In one detailed government-contracting discussion, practitioners described using ChatGPT, Claude and other general-purpose AI for requirements work, drafting and content reuse. Several comments kept returning to the same underlying condition: AI becomes much more useful when the organisation already has good information and a disciplined process around it. (Reddit: AI for proposals)
That shifted my attention upstream.
3. Practitioners Kept Pulling Me Toward The System Around The AI
The strongest correction came from proposal professionals themselves.
One particularly critical practitioner response to my research argued that proposal teams did not need another outsider starting with an AI solution thesis. The more valuable problem, in their view, was better and more intentional knowledge management. (Reddit discussion)
That criticism was uncomfortable, but useful.
I had been thinking of knowledge largely as a retrieval problem:
Can the system find the right previous answer?
Practitioner reality is harder.
Suppose an organisation has four different answers to the same technical question:
- an old proposal says one thing;
- the current architecture says another;
- marketing contains a more aggressive claim;
- a customer-specific implementation contains an exception.
Perfect semantic search can retrieve all four.
It still cannot automatically establish which statement the company is currently authorised to make.
The real questions become:
- Which source is authoritative?
- Is the information still current?
- Does it apply to this customer and product?
- Who owns the underlying fact?
- Does an SME need to approve a new commitment?
- What happens when two apparently authoritative sources disagree?
This is closer to governance than search.
APMP's own material treats proposal-content libraries as an ongoing management problem involving ownership, review, security and continuous maintenance rather than a folder that can simply be indexed once. Its 2026 AI conference even included a session specifically about building an AI-ready content library. (APMP)
The hierarchy I began to see looked more like:
AI capability → trusted organisational knowledge → proposal workflow and review discipline → competitive strategy
The model matters.
But the surrounding system may matter more.
4. Then I Ran Into A Different Problem: Access
At this point, the obvious next step was to speak directly with proposal practitioners.
In theory, that sounds easy.
Identify companies that repeatedly bid for meaningful contracts. Find proposal managers, business-development leaders and technical SMEs. Ask them to walk through a recent tender and investigate where the process consumed time.
In practice, getting people to talk was extraordinarily difficult.
I tried cold outreach. I offered paid interviews. I researched companies and individuals. I experimented with different ways of explaining why I was doing the research.
The response was close to silence.
This does not prove that the RFP market is unattractive. It does not prove that proposal professionals are unusually inaccessible. And it certainly does not invalidate the businesses already succeeding in this space.
But it changed something important for my research process.
Market access is not an abstract variable.
If understanding a domain requires repeatedly pushing through a wall just to obtain basic practitioner contact, that matters when deciding whether it is the best place for me to concentrate my next months or years.
5. I Realised I Had Been Optimising The Wrong Research Loop
This is where the proposal investigation became useful beyond proposals.
I am naturally comfortable spending long periods researching a problem, decomposing it and constructing a system around it. That is useful in engineering.
It also creates a failure mode.
A possible business can progress through something like:
-
Observation
- I see something inefficient.
-
Research
- I find evidence that the market exists.
-
Technical reasoning
- I determine that the problem appears solvable.
-
System design
- I construct a plausible workflow.
-
Commercial hypothesis
- I infer why somebody should pay for it.
Every individual step can be logically sound.
The entire conclusion can still be wrong.
The missing variable is human behaviour.
People may already have a crude solution they consider good enough. They may hate the problem but have no intention of changing their process. The person suffering may not control the budget. Implementation effort may outweigh the apparent benefit. Or the problem that looks important from the outside may barely register among practitioners.
That cannot always be resolved by doing more desk research.
6. Reddit Unexpectedly Gave Me Something I Was Missing
During the same period, I started spending more time reading and responding to people in online communities.
The experience felt completely different.
Instead of beginning with a market thesis, I encountered problems that people had volunteered themselves. Someone had lost money because an employee left and subscriptions had not been properly cancelled. Someone else was struggling with a setback in their early twenties. Small-business owners discussed missed follow-ups, manual operations and systems they wished worked better.
These problems were often tiny compared with the sophisticated AI systems I had been thinking about.
But they were real.
More importantly, I could respond.
Sometimes people challenged my assumptions. Sometimes they appreciated the answer. Occasionally a question forced me to reconsider something I thought I already understood.
That interaction created a feedback mechanism I had been missing.
Real person → real problem → my interpretation → response → correction
The loop is much shorter than:
Market research → hypothesis → product → outreach → discover whether anyone cared
I am increasingly interested in the former.
7. This Is Not A Pivot Into “Being A Reddit Creator”
I want to be careful about what I am concluding.
I am not arguing that Reddit is a business model.
I am also not concluding that RFP automation is a bad market. Existing companies and practitioner demand clearly contradict such a simplistic conclusion.
What I am changing is the method through which I decide what deserves deeper investment.
For at least the next month, I intend to spend much more time participating in communities around areas reasonably close to my existing strengths: AI, automation, systemisation, delegation, technical work and small-business operations.
The objective is not to ask everyone:
“What product should I build?”
It is almost the opposite.
I want to observe problems people already care enough about to describe without being prompted. Where I have something genuinely useful to contribute, I will try to help. Where I do not understand the problem, I will listen.
Over time, I want to see which problems repeatedly pull me deeper.
8. The Signal I Am Now Looking For Is Human Pull
This changes what counts as interesting evidence.
A technically elegant problem is no longer enough.
A large theoretical market is not enough.
Even someone saying an idea sounds useful is weak evidence.
What interests me more is behavioural progression.
Does someone ask a follow-up question?
Do they actually try the suggestion?
Do they return with another related problem?
Do unrelated people describe essentially the same frustration?
Does somebody ask for help implementing the solution rather than merely discussing it?
Does a community repeatedly expose me to problems where my particular way of thinking appears unusually useful?
Those signals do not automatically establish a commercial opportunity.
But they tell me where deeper investigation may be deserved.
9. I Am Therefore Deprioritising RFP Automation Quite Heavily
This is more than a temporary pause.
Unless I encounter strong new evidence that pulls me back toward proposal automation, I do not intend to keep forcing the investigation forward simply because I have already spent a month on it.
That month was not wasted.
It taught me that commercial RFPs differ materially from the grant-proposal environment I originally knew. It showed me how much modern AI can already accomplish. It pushed me toward knowledge management, workflow discipline and organisational truth rather than proposal generation alone.
Most importantly, proposal practitioners repeatedly corrected ideas that looked quite reasonable from the outside.
That is exactly the mechanism I now want more of.
My next research question is therefore much broader than RFP software:
Can sustained participation in real communities help me understand where my capabilities genuinely matter better than repeatedly selecting markets from the outside and trying to force myself into them?
I do not know the answer yet.
But after more than three years of repeatedly trying different directions, I think changing the search process may be more important than immediately choosing another destination.
For now, that means less time trying to invent the next business inside my own head.
And considerably more time listening to people.