Deep dives
Five ideas where a bot can own most of a watch → enrich → decide → act → escalate loop. Scores are CRO bot-leverage ratings. No market sizes or pricing here.
Bid Scout / RFP scout is killed (cold trust + BidNet-class firehoses). Do not pair a new signal test this month. The live ops candidate is doc-to-data.
RFP / tender scout
Bot leverage: 5. Pre-award signal.
Public procurement is already an event stream: a new RFP or tender appears, a supplier either fits or does not, and the useful work is filtering before anyone writes a proposal.
| Step | Bot | Human |
|---|---|---|
| Watch | New notices on public bid / tender portals for one supplier niche | Name the niche and the watch list |
| Enrich | Pull scope, due date, set-asides, location, buyer, attachments worth reading | — |
| Decide | Fit vs the supplier’s capabilities; surface or drop | Tune the fit rules after the first misses |
| Act | Daily brief: opportunity | why fit | disqualifier | Read the brief; choose whether to bid |
| Escalate | Flag odd requirements, missing docs, or a close call | Bid / no-bid; proposal writing stays human |
The product is not “AI writes your RFP response.” The product is a short list they trust enough to stop scanning portals themselves. That filtering product still lost to BidNet-class alerts on cold trust — killed. See archived experiment and doc-to-data.
Public-data signal feed
Bot leverage: 5. Pre-award signal.
Same loop as the RFP scout, on a different public watch surface: permits, licenses, inspections, court or corporate filings, and other events that imply a buying window before a formal RFP.
| Step | Bot | Human |
|---|---|---|
| Watch | One public-data class in one geography or vertical | Choose the event type and the buyer |
| Enrich | Normalize the record; attach who might care and why now | — |
| Decide | Signal vs noise against an explicit trigger list | Correct false positives |
| Act | Digest: event, why it matters, who to contact, what would invalidate it | Outreach, if they want it |
| Escalate | Ambiguous records, sealed or paywalled docs, legal-sensitive items | Judgment calls |
Do not boil the ocean of “all public data.” One event type, one buyer, one cadence — same discipline as the RFP test. If the buyer will not pay for filtered timing, it is a newsletter, not a business.
Google review ops
Bot leverage: 5. Ops.
New Google reviews are a bounded stream. The loop is operational, not creative: see the review, classify it, draft or route a response, escalate the ugly ones.
| Step | Bot | Human |
|---|---|---|
| Watch | New or updated reviews on the locations you are hired for | Connect the locations; set tone rules |
| Enrich | Sentiment, topic, whether it needs a public reply vs a private follow-up | — |
| Decide | Auto-draft, hold for approval, or escalate (legal, safety, chargeback) | Approval policy |
| Act | Draft reply in the owner’s voice; notify on SLA | Edit and publish (or let high-confidence drafts post if they opt in) |
| Escalate | Threats, alleged injury, discrimination, fake-review fights | Owner / counsel |
This ranks with the signal businesses because the event is already structured and the action is bounded. It fails if you sell “we will make you famous on Google” instead of review intake → response → exception handling.
Doc-to-data
Bot leverage: 4.5. Ops.
Unstructured documents (PDFs, email attachments, scans) into a schema the buyer already uses — then write-back to their system of record.
| Step | Bot | Human |
|---|---|---|
| Watch | Inbox, shared folder, or upload hook | Name the document types and the target schema |
| Enrich | Extract fields; keep a confidence per field | — |
| Decide | Commit high-confidence rows; queue the rest | Exception queue |
| Act | Write to CRM / ERP / spreadsheet the buyer already lives in | Spot-check |
| Escalate | Unreadable scans, conflicting fields, unknown document types | Correction; schema tweaks |
4.5 not 5: extraction quality and write-back integration are the whole job. A demo on clean PDFs is not the loop. The loop is their messy docs, their fields, their exception pile. Charge for throughput plus exceptions handled — not for “an AI that reads documents” in the abstract.
Skeptical viability (Experiment 2): document-to-data deep-dive — CAREFUL. Same cold-trust trap as Bid Scout if you sell generic invoice OCR.
Inbound receptionist
Bot leverage: 4.5. Ops.
Missed calls, web forms, and chat: qualify, answer the FAQ, book, or take a message. Same event loop; the watch surface is inbound demand.
| Step | Bot | Human |
|---|---|---|
| Watch | Call, form, SMS, or chat | Publish the number / form; set hours and routing |
| Enrich | Intent, urgency, existing customer vs new, location / service requested | — |
| Decide | FAQ, book, capture a lead, or transfer | What must never be automated |
| Act | Reply, slot on the calendar, log the CRM row | Show up for the appointment |
| Escalate | Angry callers, medical / legal / emergency, anything off-script | Live takeover |
4.5 not 5: voice and trust still sit at the edge of the loop (same family of constraint as wholesale real estate, but here the bot can own intake). The product is coverage and clean handoff, not a fake human. If the business will not define escalate rules, do not take the account.
What these five share
- A watch surface that already emits events (portals, public records, review APIs, document inboxes, inbound channels).
- A decision that can be written down (fit / noise / confidence / intent).
- An action short enough to ship daily (brief, digest, draft, extract, book).
- A human escalate path so the bot does not need to close the hard case.
That is the filter. Ideas that fail it are on the kill list.
Sources
- ChatGPT Deep Research export — TBD from Google Drive
Research/Inbox. - CRO-approved structure (scores and idea list). Do not invent TAM, win rates, or price points on top of this.
- Document-to-data — Experiment 2 (CAREFUL).