IMD product opportunity research

Snapshot: 14 October 2026. Decision: do not build a product yet; run one paid demand test for the closest hypothesis below. No funds or tokens were spent, and no contracts or tokens were deployed.

Executive decision

The strongest near-term hypothesis is a low-cost, file-only security-questionnaire response service for small B2B software vendors that answer recurring customer questionnaires but do not use a GRC/questionnaire suite. It would return the customer's original spreadsheet/document with draft answers tied to exact source evidence and mark unsupported answers for a human. Questionnaire work has evidence of recurring time cost and commercial products already sell automation. The unproven part is the opportunity: Vanta, Drata, and Conveyor already offer AI-backed response, evidence libraries, review and workflow features. A cheaper, narrower product might serve teams priced out of those tools, but neither the size of that segment nor its willingness to pay is established here.

Recommendation: validate by selling the service manually to qualified prospects; do not implement the service or issue a token unless the threshold in §6 is met. The cost of building a basic uploader is low, but getting sensitive evidence safely, fitting arbitrary spreadsheet/portal formats, and earning trust are the actual barriers. IMD's public job flow charges 0.5 IMD per action; that is not an evidenced USD SaaS price or proof that IMD can bill questionnaire customers conveniently.

The evidence against the recommendation is substantial: incumbents already cover much of the promised workflow. Vanta claims 80%+ automatic answer coverage and 95% acceptance for its generated answers. Conveyor's live pricing page offers a free trust-center tier but excludes questionnaire automation; its full platform is Enterprise/custom-priced. These are vendor claims and pricing, not independent performance studies, but they make “AI answers questionnaires” an uninteresting differentiator and show that lower-cost access is an open question rather than a proven gap. If buyers confirm that current tools meet their needs, abandon the idea.

1. Existing IMD capabilities and evidence limits

The live IMD Explorer displayed 4,325 jobs, 1,092 agents, 75 jobs finished that day and 8 failed jobs at the time inspected. Recent examples include bespoke contract launches and audits, token artwork, websites, research and project continuations. These are activity indicators, not evidence of unique customers, paid repeat usage, net revenue, or customer satisfaction. The visible sample includes many one-off launches and iterations. Explorer

IMD's docs describe a 0.5 IMD charge for job.open, job.continue, launch.open, workflow.open, oracle.request, and each scheduled run; recurring jobs/oracle questions can run on a cadence. A schedule has pre-paid runs, does not expire, and remains active until runs are spent. The documented worker runs the operator's own Claude Code or Codex login and quota. It requires a paired IMD NFT and a paid model plan; the docs explicitly say that operating a worker does not guarantee earnings. Research jobs can specify citation requirements and publish Markdown to GitHub. These are real useful primitives, but they are an execution network and pay-per-action intake, not a customer-facing recurring SaaS with CRM, spreadsheet, identity, billing, privacy or service-level commitments. IMD payment and worker docs

Relevant public IMD work confirms that scheduled jobs, oracle questions, contract builds/reviews, and public repository/site delivery are feasible. For example, the schedule-pack job records checks of recurring research jobs and oracle-question drafts; the Marginalia contract job shows published contract work and audit findings. These are examples, not proof of external demand. The public worker distribution is inspectable and documents local runtime use, worker-side tools, and public GitHub/IPFS/site delivery. I found no public evidence in the inspected Explorer/docs of customer billing in USD, paid questionnaire connectors, a support SLA, or a public customer-retention/revenue series. The exact reusable skill/tool inventory is discoverable from the documented public GET /skills endpoint; this assignment did not exercise paid workflows or verify every skill at runtime.

2. Adversarial review of the existing ideas

Swarms, Recall, Arena42 and OpenArena verdict: all four are live alternatives or substitutes for generic agent commercialization, scoring or competition. They establish that these themes are not novel. Their published counts and features prove a product exists; they do not establish durable paid use of agents for business outcomes. Recall distinguishes its simulator/paper API from live competitions that use self-funded wallets and real on-chain spot swaps; the latter is stronger execution evidence, but a leaderboard/trade record still does not establish persistent, risk-adjusted alpha or buyer retention. Recall competition API, live spot competition rules

3. Strongest opportunities considered

Scores are qualitative (1 low–5 high), based on available evidence, not market sizing. “Demand evidence” scores problem/category evidence; it does not establish demand for IMD's solution.

All three have a concrete problem and plausible cheap prototype. None has a verified, defensible gap plus IMD-specific payment evidence. The reported “demand” for candidates 2 and 3 is category demand only. Candidate 1 wins for the test because it is separable from on-chain speculation, has a repeated deliverable, admits a low-cost manual trial and has a simple paid signal. It does not yet win permission to build.

4. Economics for the closest hypothesis

These are test assumptions, not a market forecast or observed sales. Price in USD is proposed solely for the experiment; IMD's current public task intake price is 0.5 IMD and is not a comparable USD market price.

Customer/use case: a 5–50 person B2B software company; a founder, sales engineer or security owner receives a prospect's 25–100 row security Excel questionnaire and needs a grounded draft quickly. Exclude regulated/high-consequence submissions and forms requiring a portal/login during the first experiment.

Proposed transaction price: $49 for one questionnaire up to 50 questions, with a human-reviewed cited answer sheet. An eventual $99/month subscription for up to three submissions is only a price hypothesis. Do not make the subscription until repeat usage is observed.

Cost model (assumptions): model/vendor API or worker runtime $2–8 per questionnaire; document conversion/storage and service infrastructure $0.50–2; 15–25 minutes of human review/support at an assumed $40/hour = $10–17. Total direct cost is about $12.50–27 per $49 run, or 45–74% gross margin before payment fees, customer acquisition, security/compliance, refunds, and founder time outside review. At low usage (one run/month), one-run contribution is $22–36.50. At moderate usage (three runs/month at $99), assume $6–24 model cost, $2–6 infrastructure, and 45–75 minutes review/support ($30–50): direct cost $38–80, gross margin about 19–62%. The wide range is deliberate: no IMD run cost, output quality, inference price, handling time or questionnaire volume has been measured. If it needs more than 25 minutes of review/run or frequent bespoke portal work, the $49 price is not viable.

Acquisition to first 100 relevant users: no paid ads assumed. Start with five accelerator/fractional-security partners plus B2B founder/security communities; seek 20 qualified introductions from each partner over a quarter (100 total), then publish a one-page before/after sample with synthetic evidence and offer a limited paid concierge batch. The 20-intro yield is an explicit assumption, not observed conversion evidence. The first 15 qualified prospects form the experiment; the path to 100 trials remains unverified. Consultants may be better channel partners than direct cold acquisition, but that is a hypothesis.

Why repeat: customer questionnaires recur as new enterprise prospects arrive; a reusable approved answer/evidence library could reduce each subsequent response's work. However, the same knowledge library creates product lock-in for Vanta/Drata/Conveyor. Buyers may only receive these requests rarely; return frequency is unknown for target-sized firms.

Smallest willingness-to-pay test: manually process a redacted or synthetic sample end-to-end for 15 qualified small vendors, quote $49 for their next questionnaire and take payment before doing the work. Accept only questionnaires/evidence that the customer has redacted or explicitly approved for this pilot; do not collect credentials, portal access, source code, secrets or confidential customer data, and do not use a public IMD task. If buyers will only participate with confidential material, that itself is a blocker until private processing exists. The output is draft-only and the buyer remains approver. Measure paid conversion, minutes of review, number of material corrections, turnaround, and second paid order within 60 days.

Minimal first version and cost: the validation version is a manual concierge service, not software: take an approved redacted spreadsheet and evidence bundle, draft citations and answers with a human reviewer, and return a copy of the same file. It needs no deployed contract, token, paid IMD job, hosted app or customer credentials. Assume $0–100 of setup/tool cost for the sample workflow, plus $12.50–27 direct cost per paid run from the scenario above; founder time and data-protection overhead are not included. If it passes, a single-tenant private intake/export prototype is roughly 5–10 engineering days and an assumed $100–500 of hosted model/storage costs for 10 test runs; both are planning estimates only, excluding engineering labor, security/legal work and support. Public IMD job/result storage cannot handle confidential questionnaire data.

5. IMD execution fit, dependencies and failure modes

Reusable today: workers execute Claude Code or Codex using operator-paid model quota; scheduled paid actions can repeat; research jobs can require citations; public jobs can deliver a Markdown artifact or GitHub repository; Explorer activity and job status are public. This is enough to test a manual research/document workflow, not enough to claim a secure SaaS.

Dependencies before productization: safe upload/retention/deletion; Excel/DOCX/PDF parsing and writing without corrupting original formatting; immutable source/evidence references; customer workspace isolation; authentication and access control; a USD billing/subscription provider; reliable job callbacks/retries and support; privacy/security review; explicit human approval before export. None is confirmed as a ready-made IMD component in public docs. Optional worker tools and runtime keys are operator-local; there is no documented customer-safe mechanism for ingesting private evidence at scale.

Security and quality risks: questionnaire packages may include SOC 2 reports, architectural diagrams, subprocessors, incident records or customer-specific confidential material. Publicly submitted job bodies/results are inappropriate for such inputs. A model can invent controls, use an expired document, or answer a nuanced question incorrectly; citations can point to a document without supporting the exact claim. Do not process real confidential evidence on public/shared jobs. A future version needs private execution/storage, explicit deletion, tenant isolation, audit trails, human approval, and no automatic portal submission. Agent review is not a compliance certification.

Likely recurring costs/failure modes: inference and conversion per file; human correction and buyer support; model/API quota or worker-seat cost; parser drift as templates change; occasional spreadsheet macros/portal formats; stale source documents; output formatting failures; customer acquisition; and sensitive-data incident exposure. At 0.5 IMD per IMD action, scheduled processing is metered per run, but the network fee alone says nothing about local operator inference cost or end-customer subscription margin.

6. Go/no-go test and decision

Run the concierge test only. Recruit 15 companies that (a) have 5–50 employees, (b) completed at least one 25+ question customer security questionnaire in the last 90 days, and (c) do not already use Vanta, Drata or Conveyor's questionnaire automation. Show a sanitized sample with citations and unsupported items marked. Offer $49 for the next redacted questionnaire/evidence bundle; do not accept unredacted confidential material or upload data to a public IMD task.

Go to a private, minimal product prototype only if all conditions hold: at least 5 of 15 qualified buyers pay $49 before processing; at least 3 place a second paid order within 60 days; median hands-on review is at most 25 minutes per ≤50-question response; and at least 90% of answered cells have direct supporting evidence with no critical false claim in the human-reviewed sample. This is a validation bar chosen for this experiment, not an industry benchmark.

No-go / abandon this product: fewer than 5/15 pay; fewer than 3 repeat by day 60; buyers say their current suite already solves it; redaction/privacy makes customers unwilling to trial; median review exceeds 25 minutes; or any material unsupported claim survives the review. A failed test means do not retreat to a generic agent marketplace, token or leaderboard.

7. Token decision and next task

Token: never for this product. Questionnaire customers buy a service outcome and need ordinary invoice/card purchasing, predictable USD pricing, confidentiality and recourse. An IMD token introduces volatility and purchase friction without an evidenced customer benefit. IMD may use its existing internal payment rail for agent work if appropriate, but that does not make an agent/product token necessary now or later.

Exact next implementation task: none yet. The evidence supports a paid concierge demand test, not software development. If the go criteria pass, the first implementation task should be a private, single-tenant questionnaire intake/export prototype that accepts one spreadsheet plus an explicitly selected evidence bundle, writes cited draft answers back into a copy of the same file, marks unsupported answers blank, requires human approval, and deletes inputs/outputs after a stated short retention window. Do not connect an agent to customer portals or enable auto-submission in v1.

8. Sources, confidence and unresolved questions

Facts above are linked to live product pages, primary documentation, repositories, or the live Explorer where possible. Product-feature and pricing pages are vendor-published. Vendor-reported adoption/performance claims (for example Vanta's completion/acceptance rates, Arena42 registrations, Virtuals counts, and Swarms listings) are explicitly treated as claims, not independently verified outcomes. Official docs identify Recall's Base token contract, but BaseScan returned HTTP 403 during inspection; no holder, transfer, stake or service-payment volume is used as adoption evidence here. Whistic's survey is vendor-produced and older; it demonstrates a reported category workload, not the target segment's demand. Market-category breadth is evidence that the customer problem exists, not evidence of an open niche.

Reasonable inferences: enterprise questionnaire platforms may be uneconomic for some very small, low-volume vendors; a private human-reviewed file workflow may be technically simple; current public IMD primitives reduce the cost of a manual pilot. These are hypotheses.

Unknown/unverified: number of small vendors with frequent questionnaires but no incumbents; their willingness to pay $49; repeat frequency; realistic inference and support cost; available private execution and retention controls; USD/token-equivalent costs and worker earnings; competitor cohort retention/GMV; and whether IMD can acquire customers through a channel partner. A limited web/document search cannot establish novelty across all markets. The explicit experiment, not this report, should resolve the decision-critical unknowns.