Loopcard

All documents · Application jobs

Agentic application jobs — demand study

Date: 18 September 2026
Author: product researcher (application jobs, not infrastructure)
Scope: jobs where the product does the work a business or freelancer already pays a human (or a VA, or a $200+/mo tool) to do this week. One Ubuntu VPS. User’s LLM keys. Value in 60 seconds.
Out of scope (per brief + operator): MCP gateways, traces, permits, scanners, quantum hardware, social CryptoBook, token wallets.

Headline for PM: the honest list of jobs that pass every filter is empty. Five specific jobs are ranked below as the closest misses. Each has real demand. Each dies on clone, ChatGPT-substitution, legal, or “not 60 seconds on one VPS.” Do not stitch them into another Slip.

1. Filters (instant fail)

A candidate is KILL if any row fails. FIX means a narrower object on the same job might pass; it is not a pass.

# Filter Fail if
1 Pays this week A “should” (governance, GEO insurance, literacy). No card, no VA invoice, no existing $ tool they already resent.
2 Product does the work Copilot, dashboard, permit, receipt, gateway, scanner. The output must be the deliverable (email, PDF, listing, booked slot, resolved ticket).
3 One VPS + BYOK Needs a browser farm, GPU, quantum, or we hold wallets / resell tokens.
4 60-second start Sales call, SDK, IdP, crawl-the-whole-site, train-on-your-product for a week.
5 Not EU AI Act high-risk Hiring, creditworthiness, health/triage, education admissions, biometrics, law enforcement. Transparency-only OK. Do not claim “AI Act compliant.”
6 Not US-illegal Money transmission, broker/trading advice, CFAA (we log into third-party sites), fake SOC2/auditor.
7 Not a named clone Intercom / Sierra / Glean / Perplexity / Cursor. Also Fin, Decagon, Ada, Zendesk AI as the same support job.
8 User profit they already feel Saves time or money on a wound that exists on their machine this week.

Ignore TAM slides. Bain’s ~$100B “glue work” (May 2026) is real as an opportunity size; it is not a SKU. MarketsandMarkets $19.33B (Aug 2026) includes Microsoft/AWS/Salesforce.

2. September 2026 Product Hunt — clone vs open

Opened 18 Sep 2026: monthly leaderboard, week of 14 Sep, plus first-party product pages named below.

PH is not a census of who pays. It is a census of what builders launch. This month’s board is ~70% infra, coding agents, and payments. The application-shaped launches that match the operator’s list:

2.1 Naoma AI Demo Agent V2 — CLONE (do not ship)

Open? Only the price band below $249 for a scripted, non-video, non-live-product tour. That is Walnut/Supademo, also occupied, and it is not “the agent does the demo.”

2.2 Support agents — CLONE (explicitly forbidden)

Intercom rebuilt around Fin (~$0.99/resolution, 14-day trial, Salesforce deal ~$3.6B Jun 2026). Sierra $950M Series E May 2026, ~$15.8B, ~$150–200M ARR. Decagon $4.5B Jan 2026, Vendr median contract ~$386k–$433k/yr. Zendesk AI ~$1.50–$2.00/resolution. PH this week: ReplyPilot (“24/7 AI agent that resolves customer support tickets”), Widgo as sales-rep-on-site. Gorgias owns e-commerce. There is no indie VPS slice of “resolve the ticket.” Brief forbids Intercom/Sierra clones by name.

2.3 AGENT-NATIVE (and llms.txt) — OPEN as a utility, closed as a company

Open? The WordPress/one-URL transformer is unoccupied at quality + log of which models hit you — and still not a CryptoBook replacement. Kill as the SKU. Steal nothing.

2.5 Other Sep PH “agents that do work” — all clones or illegal here

Launch Job Verdict
Kilo Code (JetBrains + iOS/Android) Coding agent Cursor clone. Forbidden.
Mastra Factory Issue → production by agents Framework / Cursor-adjacent. Infra.
Nimble Web Search Agents Domain web research + retrieval Perplexity/Exa clone.
Anysite.io B2B lists by chatting to your agent Clay / Apollo / ZoomInfo. Enrichment often means scraping; CFAA residual.
Browzer (2 Sep, ~441) Repo → docs, blogs, changelogs, videos Mintlify / GitBook / Fern + a writer. GitHub connect is not 60s.
Slashy Assistant “Does email for you” Superhuman / Gemini-in-Gmail.
Replay Autonomous app testing QA Wolf / Rainforest. Browser infra.
Aside Browser that signs into your accounts Ninth Circuit Amazon v. Perplexity (4 Aug 2026): user-directed, user-machine is safer; vendor browser that logs in is the CFAA path the brief kills.
Viktor.com “AI coworker that actually does the work” Horizontal Glean/Sierra shape.
Text Agent Store SMS agent marketplace Two-sided; verticals include health coaches / recruiters / betting (Annex III + US).
Figo (17 Sep) Weekly competitor briefing See Job 1 below. Category is launching this week.
tiun. / CREEM 2.0 Auth+billing / sell AI products Infra + payments.
Voiskey Voice typing Not an agent that completes a business job.
Switch Agents in Slack/Teams Infra.
Elva / Monid / Weave APIs for agents, tool router, coding router Infra. Ignore.

Read-through: PH is celebrating agents that replace a workflow step. The steps that win votes (demo, support, code, search, sell-to-agents) are exactly the named clones or the legal kills. The leftover “open” launches are utilities (AGENT-NATIVE) and me-too competitor briefs (Figo).

3. Five specific jobs, ranked

Rank = residual chance of being a CryptoBook SKU, not TAM. Every row includes the kill. None survive all eight filters as specified.

Scoring: Demand this week × 60s × not-a-clone, 1–5. Legal/VPS are binary.

Rank 1 — Monday competitor brief from named public URLs

Verdict: KILL (clone of a SKU shipping this month)

One sentence. Paste 3 public URLs (pricing, changelog, homepage). Every Monday the product fetches them, diffs against last week’s snapshot, and emails a one-page brief: what changed, why it matters, one action. User never logs into the competitor.

Filter Result
Pays this week Yes. Product marketers and founders already have “check competitor.com” on Monday. They pay Crayon / Klue ~$10k–$16k+/yr (sales) or waste 45 minutes and miss a price change.
Does the work Yes — the brief is the work, not a ping.
One VPS + BYOK Yes, if we only fetch public HTML and summarize with the user’s key. No login.
60s Yes: paste URLs, see a canned “this week vs last week” on a demo competitor (e.g. a public pricing page we snapshot).
EU / US legal OK if public pages only, no credit/hiring, no “we make you compliant.” Disclose AI. CFAA fail if we fetch login-gated or ToS-blocked properties.
Named clone Not Intercom/Sierra/Glean/Perplexity/Cursor if we refuse to become a search engine or a knowledge base.
User profit One missed pricing change or a 45-minute Monday.

Evidence (opened / dated):

Kill reason. The interpreted Monday brief was the last unoccupied layer above Visualping’s ping. It is occupied at $19–$49 (Competiflow, Outmano) and was relaunched on Product Hunt yesterday (Figo). A fourth brief is a me-too. Widening to ads/social/AI-citations is Semrush/Ahrefs/Otterly. Logging into competitor apps is CFAA.

What would have been the wedge (do not build it anyway): only pricing + changelog + homepage, named URLs, public fetch, one action line, $19. That is Figo/Outmano’s landing page.

Rank 2 — Agency Friday pack (screenshots → client PDF + covering email)

Verdict: KILL (ChatGPT already is the product; no dedicated-SKU demand proof)

One sentence. Account manager drops 3 dashboard screenshots they already have open (Ads, GA, Search Console, Shopify). Product returns a branded 1-pager plus the email they send the client before Monday’s call. Optional: we keep last Friday’s pack so this week vs last week is automatic.

Filter Result
Pays this week Partial. Agency AMs spend Friday on reporting. They pay AgencyAnalytics / DashThis / Whatagraph ($49–$249) or a VA. The screenshot path exists because those tools need two-week API integrations.
Does the work Yes — the PDF and the email.
VPS + BYOK Yes. Vision model on user key. PDF render on this box.
60s Yes: three example screenshots preloaded.
Legal OK if we don’t store client PII longer than needed; this is not Annex III.
Named clone Not the five names.
User profit 1–3 hours Friday → 10 minutes, if they trust the numbers we OCR’d.

Kill reason. (1) No primary evidence of a card-swipe SaaS whose only job is screenshot→client PDF that is converting this week. AgencyAnalytics demand is for connected data, not screenshots. (2) Claude/ChatGPT with vision does the 60-second aha for $20. A prettier PDF is not a felt $29/mo wound. (3) OCR of revenue numbers will be wrong once; that is how you lose the agency. (4) Retention is weekly, which is good, but activation is “I already have Claude.”

FIX that still dies: “connect GA4” becomes a two-week integration and a Glean-shaped dashboard.

Rank 3 — Invoice PDF / aging CSV → send-ready chase pack (we never send, never collect)

Verdict: KILL (clone of ChaseAI; sending/collecting would be legal KILL)

One sentence. Drop a PDF invoice or an aging CSV. Get the 7 / 14 / 30-day email sequence, in the user’s voice, with amount, due date, and a human-sounding “reply to this thread.” User copies into Gmail. We do not send mail, take cards, score credit, or threaten.

Filter Result
Pays this week Yes. Late invoices are cash-flow. Freelancers and 2–20 person agencies chase weekly. Upwork/agency Slack is full of “still waiting on invoice.”
Does the work Drafts are the work only if they send them. If they still paste into Gmail, ChatGPT also drafts.
VPS + BYOK Yes.
60s Yes: sample overdue PDF.
Legal Draft-only + no credit score = not Annex III, not FCRA, not money transmission. The moment we send mail, run collections, or score “will pay,” we die. Chaser’s collections take a % of recovered invoices — that is a collections agency, not us.
Named clone Not the five names.
User profit One $2,400 invoice paid 10 days sooner pays a year of $29.

Evidence:

Kill reason. The self-serve, PDF-in, AI-copy, sub-$20 job exists (ChaseAI $9). Chaser vacated the floor by moving to turnover pricing; ChaseAI (and QBO reminders) caught it. If we send the email we become an ESP + collections-adjacent. If we don’t send, we are a GPT wrapper they already have.

Rank 4 — Google Business Profile review → public reply + (if 1–2★) internal make-good pack

Verdict: KILL (self-serve $9–$29 layer is full)

One sentence. Paste a Google review (or connect GBP). Get three reply drafts in brand voice. If the star rating is 1–2, also get an internal checklist: refund / remake / call, and a one-line SMS to the owner. User posts the reply. We do not claim we “make you rank.”

Filter Result
Pays this week Yes. Local operators get reviews this week. Google emails them. Unanswered 1-stars are a felt wound. 98% of consumers read reviews (vendor claim; treat as directional).
Does the work The reply is the work. Auto-post via GBP API is the agentic version.
VPS + BYOK Yes.
60s Yes: paste a 1-star sample.
Legal OK as a draft. Auto-post is still not Annex III. Do not do medical/legal advice in replies (dentists, lawyers — stay generic or refuse those verticals).
Named clone Not Intercom/Sierra (those are inbound chat). Podium is the expensive cousin.
User profit 10–20 minutes/day of owner time; one recovered Yelp/Google customer.

Evidence (2026 pricing, multiple recaps through Sep):

Tool Public entry Motion
Podium Core $399/mo (Sep 2026 PulseSignal extract of podium.com/pricing); AI replies often a ~$99 add-on Sales
Birdeye $299–$449 per location Sales
NiceJob ~$75/mo Review collection more than AI replies
RepliFast $29/mo Self-serve Google replies, 5 min setup
MARA $28/mo Same
StellarReply ~$25 / £20 SMB + Airbnb
Reply Champion $10/mo (1–2 locations) Owner
ReviewMankey $8.99–$39.99; free 10 AI replies Self-serve, live 17 Sep 2026

Kill reason. The “Podium is $399 so there is a $29 hole” story was true in 2024. It is false in Sep 2026. RepliFast, MARA, ReviewMankey, Reply Champion, StellarReply already sell the 60-second reply. Adding a make-good checklist is a feature, not a company.

Rank 5 — Client email dump → scoped Statement of Work

Verdict: KILL (ChatGPT + $9 generators)

One sentence. Paste the messy client thread. Get a 10-section SOW: in-scope, out-of-scope, assumptions, revision cap, deemed-acceptance, change-order. Freelancer sends it before kickoff. We are not a lawyer; label it a draft.

Filter Result
Pays this week Yes, as a human task. Freelancers and tiny agencies write SOWs (or skip them and eat scope creep) every time a job starts. Upwork AI-mention jobs are 20% of posts (Aug 2026, UpHunt 3.5M-post index); proposal/SOW writing is still a weekly ritual.
Does the work The SOW file.
VPS + BYOK Yes.
60s Yes: sample thread.
Legal Draft + “not legal advice” is OK. Do not sell it as an enforceable contract. Unauthorized-practice risk if we claim lawyer-grade.
Named clone Not the five names.
User profit One change-order that sticks is $400–$6,000 (GigRadar 2026 anecdote: $6k withheld for “not premium enough”).

Evidence:

Kill reason. The 60-second aha is ChatGPT. Stecya already charges $9 for the download. A Virtus-pink SOW PDF does not beat the model the user already pays for.

4. Scoreboard

Rank Job Demand this week 60s VPS/BYOK Legal Named-clone Dedicated SKU still open? Verdict
1 Monday public-URL competitor brief 5 5 5 4 (public only) 4 No — Figo/Outmano/Competiflow/Visualping Reports KILL
2 Agency Friday screenshot pack 3 5 5 5 5 No evidence of a paying SKU vs Claude KILL
3 Invoice chase drafts (no send) 5 5 5 5 if draft-only 5 No — ChaseAI $9 KILL
4 GBP review reply + make-good 5 5 5 5 5 No — $9–$29 shelf full KILL
5 Email dump → SOW 4 5 5 4 (draft only) 5 No — ChatGPT + Stecya $9 KILL

Honest list that passes all eight filters: empty.

5. Also killed (so we do not “invent a sixth”)

Tempting job Kill in one line
Naoma-class live demo agent Clone + browser/video farm + $249 floor. §2.1
Support / website sales rep Intercom Fin / Sierra / Decagon / Widgo. Forbidden.
AGENT-NATIVE / llms.txt Free plugin; paid job is citation tracking, not doing work.
ZeroClick / x402 storefront $55M + money transmission.
B2B list building Clay/Apollo; scrape/CFAA. Anysite is on the Sep board.
Coding agent / repo→PR Cursor. Kilo/Mastra are the PH proof.
DevRel content from git Browzer + Mintlify.
Inbox agent Slashy / Gemini / Superhuman.
Meeting recap email Fathom / Fireflies / Otter already send it.
Transcript → SOP Scribe/Tango own clicks; Claude owns the transcript.
Changelog from git log git-cliff is free; Browzer is the hosted version.
Airbnb listing from photos Hostaway/Guesty AI; Fiverr $20–50; seasonal, not weekly for one host.
App Store review replies Same shape as Job 4, smaller market, AppFollow exists.
AP invoice-vs-PO exception email Feature of Bill.com / Vic.ai / Stampli; bookkeepers already have Hubdoc/Dext.
RFP → proposal for agencies Loopio / AutogenAI (sales); Claude for the 5-person shop.
Bank CSV → books Expensify / Ramp / QBO. Don’t become credit.
Aside-class browser agent CFAA. Brief kill.
Quantum anything Hardware we don’t have; prior research kill.
Permit / Slip / Hallpass Infra. Operator paused that pass. Not an application job.

6. Why the empty list is the finding

Three patterns ate every job that looked specific:

  1. If a human is paid this week to do it, a $9–$49 self-serve agent already does it (ChaseAI, RepliFast, Stecya, Visualping, Figo, Outmano). The 2024 “enterprise is $399 so we take $29” hole closed in 2025–26.
  2. If no dedicated SKU exists, ChatGPT/Claude with a paste box is the SKU. Users already pay $20–$200/mo for the model (Menlo 16 Sep 2026: 55% of US AI users pay; regular agent users 92% pay). A Virtus-pink wrapper of the same paste is not a felt save.
  3. If the agent must act in the world (run the demo, resolve the ticket, log into Gmail, collect the invoice, buy as an agent), you collide with Sierra/Intercom/Naoma/ZeroClick/CFAA/money transmission — and you cannot run it honestly on one VPS.

Bain’s glue-work TAM is the labor. The labor that fits a 60-second homepage is already a GPT wrapper or a clone. The labor that is still expensive (Sierra, Naoma, Crayon, Podium) is expensive because of distribution, browser farms, IdP, or a sales team — none of which this VPS has.

7. Instruction to PM

8. Sources actually opened

Product Hunt (18 Sep 2026):

First-party / vendor:

Dated recaps used as pricing pointers (not TAM):

Not used as gospel: Bain $100B (opportunity size, not invoices); MarketsandMarkets $19.33B (includes hyperscalers); vendor conversion claims (Naoma 6–20%) treated as their numbers, not ours.

If every later idea is a wrapper of Jobs 1–5, the answer is still empty. Empty is the market study.