What changed
FACT (per cited sources): Google shipped computer use in Gemini 3.5 Flash, putting agentic screen control into a cheap low-latency model tier; OfficeCLI provides headless single-binary Office-file manipulation without a Microsoft Office install; Coasty markets no-API legacy-GUI automation (vendor claim, reliability unverified per the signal itself).
Why now
The unit economics of vision-driven GUI automation just dropped from 'enterprise RPA licensing' to 'pennies per run' (HYPOTHESIS: actual per-workflow cost is uncited). This makes a solo operator's labor, not software licensing, the main cost of automating a legacy desktop workflow.
Converging signals
Cheap computer-use model tier (Gemini 3.5 Flash) + headless Office document I/O (OfficeCLI) + a market entrant (Coasty) validating that someone believes no-API legacy automation is sellable. Note: Coasty's existence proves competitor interest, not customer payment.
Customer pain
HYPOTHESIS ONLY. The input's demand_evidence is empty — there are zero PAIN complaints, zero HIRING/SPEND postings, and no FORCED BUYER mandate in this convergence. That SMBs suffer with legacy GUI-only software is plausible intuition, but per the scoring rules it cannot be treated as demonstrated demand here.
Who pays
Hypothetically: SMB owners running GUI-only line-of-business apps (dispatch, scale-house, inventory, billing) who re-key data between the app and Office documents. No evidence in the input identifies a specific paying segment or budget.
Solved today
Manual re-keying by staff; traditional RPA (UiPath/Automation Anywhere/Power Automate Desktop) which is licensed per-bot/per-seat and consultant-heavy; local IT consultants writing brittle AutoHotkey/macro scripts. (Industry-knowledge inference, not from cited sources.)
Why current solutions are bad
RPA stacks are priced and designed for enterprises; selector-based scripts break on UI changes; SMBs can't justify the licensing or the consultant retainer. HYPOTHESIS: vision agents are more resilient — but the Coasty signal itself flags reliability as unverified, and legacy-GUI vision automation failure modes (occluded windows, modal dialogs, latency, non-deterministic layouts) are real and undemonstrated at production quality.
Proposed product
A productized service: 'We automate one ugly desktop workflow for a fixed fee + monthly monitoring.' Stack = Gemini 3.5 Flash computer use driving the legacy GUI in a VM, OfficeCLI generating/parsing the Word/Excel artifacts server-side, Python glue with human-in-the-loop review on low-confidence runs.
MVP version
Pick ONE legacy application the founder already knows and can access — strongest candidate: scrap/recycling yard-management software (ScrapWare, ScrapDragon, 21st Century Programming ROM), which is exactly the GUI-only legacy niche where Charles has operational credibility. Record one real workflow (e.g., daily purchase-ticket export → Excel settlement report), automate it end-to-end in a VM, produce a before/after time-and-error demo video.
30-day build
Weeks 1-2: build the one-workflow demo against a scrap-yard package; validate agent reliability over ≥50 consecutive runs and log the true failure rate and per-run model cost. Weeks 3-4: cold-demo (video, not calls-first) to 20 yards/SMBs the founder can reach through recycling-industry contacts; goal = 3 discovery conversations. If nobody bites on a free pilot, kill.
60-day build
Convert 1-2 pilots to paid ($1.5k-3k fixed setup + $200-500/mo monitoring). Harden the runner: screenshots + audit log of every action, human-approval gate for irreversible steps, VM snapshot rollback.
90-day revenue plan
Realistic ceiling: 2-4 paying clients ≈ $4k-10k one-time + $0.5k-2k MRR. This is consulting revenue, not product revenue; it pays bills but does not compound without a repeatable vertical template.
Distribution path
Weakest link. No forced buyer, no filing deadline, no registry of prospects. Path is founder's recycling-industry network + demo videos in niche operator communities. Anything broader ('SMBs with legacy apps') is unreachable solo without ad spend or long outbound cycles.
Pricing hypothesis
Fixed-fee per workflow ($1.5k-3k) + monthly monitoring retainer ($200-500). Avoid hourly (caps income) and avoid per-seat (nothing forces recurring usage).
Technical difficulty
Moderate. The pieces exist, but production reliability of vision agents on idiosyncratic legacy GUIs is the unproven core — the Coasty signal explicitly labels this unverified. Expect substantial per-client tuning; this is the hidden labor cost that kills margins.
Legal / regulatory risk
Meaningful for a service: the agent operates inside the client's business systems with real financial data. An agent misbooking transactions creates liability; needs contracts, action audit logs, and human-approval gates. Also client-data handling obligations (screenshots of their books transit to Google's API — some SMBs will refuse).
Platform dependency
High. The economics rest entirely on Gemini 3.5 Flash computer-use pricing and availability; a price change or capability regression breaks the margin. OfficeCLI is a single third-party GitHub project of unknown maintenance quality.
Founder fit
Mixed. Automation, AI workflows, and industrial-operations credibility fit; the scrap/recycling wedge is genuinely differentiated. BUT this is NOT the proven FMCSA shape: no regulation compels anyone to act, no deadline, no per-filing meter, and it requires SMBs to trust an outsider's AI with desktop control — a trust-building sale, which the founder profile avoids. It is bespoke services, his stated anti-pattern versus productized per-transaction tools.
Breakout potential
If one vertical workflow proves repeatable (same software, same report, many yards), it could productize into a niche micro-SaaS ('automated settlement reports for ScrapDragon users'). That is the only version worth pursuing; generic 'we automate anything' has zero breakout path.
Final recommendation
PASS on the broad version. CONDITIONAL REVISIT only as a narrow wedge: a 2-week reliability spike automating one scrap-yard-software workflow the founder can access through his recycling network. If ≥95% unattended run success and one yard agrees to a paid pilot, continue; otherwise kill permanently. Do not divert time from FMCSA-shaped forced-filer opportunities, which fit the proven edge far better.
Next action
Obtain trial/demo access to one scrap-yard management package (ScrapDragon, ScrapWare, or ROM), record one real report workflow, and run a 50-iteration reliability benchmark with Gemini 3.5 Flash computer use, logging success rate and per-run cost — before writing any sales material.