What changed
Three enabling pieces landed together: Google shipped computer use in Gemini 3.5 Flash, a cheap low-latency tier (FACT, per DeepMind blog); OfficeCLI provides headless single-binary Office file manipulation with no Microsoft Office dependency (FACT, per GitHub repo description); and Coasty launched claiming no-API GUI automation of legacy software (FACT that the claim exists; the capability's reliability is explicitly unverified in the source).
Why now
The economic floor for screen-driving agents just dropped: Flash-tier pricing makes per-run costs low enough that a solo builder can sell automations without enterprise RPA licensing economics. HYPOTHESIS: RPA incumbents (UiPath-class) remain priced and packaged for mid-market/enterprise, leaving SMB legacy-GUI automation underserved — plausible but not evidenced in the provided sources.
Converging signals
(1) Gemini 3.5 Flash computer use = cheap screen/browser control; (2) OfficeCLI = agents can emit real .docx/.xlsx deliverables server-side; (3) Coasty's launch = at least one vendor believes GUI-only legacy automation demand exists. Together: an agent that drives an old desktop/GUI app and outputs Office files, at near-zero marginal cost.
Customer pain
HYPOTHESIS: SMBs with legacy desktop software (dispatch, scrap-yard, dental, title, property-management systems) re-key data by hand and pay staff hours weekly for it. The input's convergence description asserts 'validated demand' but NO demand_evidence array (PAIN/HIRING/FORCED BUYER items) was provided, so this pain is unproven in this brief and scored accordingly.
Who pays
Owner-operators of SMBs running one legacy system, paying for a specific recurring workflow to disappear. HYPOTHESIS — no buyer evidence supplied.
Solved today
Manual re-keying by admin staff; occasional VBA/macros; RPA tools (UiPath, Automation Anywhere, Power Automate Desktop) that require licenses and scripting expertise SMBs don't have. (Competitive landscape is general knowledge, not from provided sources.)
Why current solutions are bad
RPA is seat-licensed, brittle, and needs a consultant to script; macros break; manual entry costs real payroll hours. Computer-use agents promise setup-by-demonstration, but reliability on arbitrary legacy GUIs is the open question — the Coasty source itself flags reliability as unverified.
Proposed product
A productized service, not a platform: 'Pick one workflow in your legacy system; I automate it end-to-end for a fixed setup fee + monthly fee.' Stack: Gemini 3.5 Flash computer use for GUI driving, OfficeCLI for deliverable generation, a supervision/retry harness, and human-review checkpoints on anything that writes data. Later, harden the 2-3 workflows that recur across customers into a self-serve micro-SaaS.
MVP version
One vertical, one workflow: e.g. scrap/recycling yards (founder's domain credibility) — agent pulls daily transactions from a legacy yard-management GUI and emits a formatted Excel/Word compliance or settlement report. Build the harness on a Windows VM, record a 3-minute demo video of the agent doing the job, and cold-outreach 20 yards Charles already understands operationally.
30-day build
Week 1-2: build the harness (VM + Flash computer-use loop + OfficeCLI output + logging/screenshot audit trail) against one real or demo legacy app; measure actual reliability across 50 runs — this is the kill/no-kill gate. Week 3-4: demo video + direct outreach to 20-30 scrap/recycling operators; goal: 3 paid pilot commitments at $500-1,500 setup.
60-day build
Deliver 2-3 pilots, instrument failure modes, add human-in-the-loop confirmation for writes. Standardize the top workflow into a repeatable package with fixed pricing. Collect before/after hours-saved numbers for proof.
90-day revenue plan
Target: 3-5 customers at $500-1,500 setup + $200-500/mo monitoring/maintenance = roughly $3-8k booked in 90 days. HYPOTHESIS — depends entirely on the reliability gate passing and on unproven demand.
Distribution path
Direct demonstration: screen-recorded before/after videos posted in vertical operator groups/forums and sent cold to named businesses in a niche Charles has credibility in (recycling/scrap). No enterprise sales, no ads. This matches his demonstrated-value selling style.
Pricing hypothesis
$500-1,500 one-time per automated workflow + $200-500/mo per workflow for monitoring, retries, and updates when the legacy UI changes. Anchor against the payroll hours it replaces, not against RPA license prices.
Technical difficulty
Moderate. The models exist off-the-shelf, but the hard 20% is reliability engineering: legacy GUIs with modal dialogs, focus bugs, and timing issues will break naive agents. Windows VM management, audit logging, and safe-write guards are all solo-buildable but real work. Coasty's own positioning concedes reliability is the unsolved part.
Legal / regulatory risk
Low-moderate. Agent operates customer's own software under their credentials — get written authorization per client. Risk concentrates if workflows touch regulated data (payroll, PHI); avoid those verticals initially. Errors that corrupt customer books are a liability exposure — mitigate with dry-run/review modes and E&O-style contract language.
Platform dependency
Real but diversified: dependent on Gemini Flash computer-use pricing/availability, but Anthropic and OpenAI ship equivalent computer-use capability, so no single-vendor lock. OfficeCLI is an open-source single binary (low risk, but repo maturity unverified).
Founder fit
Good but not perfect. Fits: automation, AI workflows, fast prototyping, industrial/recycling operator credibility for the beachhead vertical, demonstrated-value selling. Does NOT fit the proven FMCSA-shape edge: there is no government mandate, no forced filer, no per-filing toll booth — the buyer can always keep doing it manually, which weakens urgency. It is a services-flavored business at first, which caps leverage until workflows standardize.
Breakout potential
Moderate. If 2-3 workflows recur across a vertical, this becomes a niche vertical micro-SaaS ('automated settlement reports for scrap yards') with defensibility from workflow knowledge rather than tech. The tech itself is copyable within weeks; the vertical wedge and reliability harness are the moat.
Final recommendation
CONDITIONAL GO as a narrow productized service, not a platform. The capability convergence is real and cheap to test, but demand is unevidenced in the input. Spend ≤2 weeks and <$100 building the harness and running the reliability gate (≥95% success on 50 runs of one real legacy workflow). If it passes, sell 3 pilots in the recycling/scrap niche where Charles has credibility. If reliability fails or 30 outreach attempts yield zero paid pilots, kill it and keep the harness as infrastructure for a future forced-buyer (government-filing) play, which remains his higher-fit pattern.
Next action
Stand up a Windows VM, wire Gemini 3.5 Flash computer use + OfficeCLI into a supervised loop against one legacy yard-management (or demo) app, and run the 50-iteration reliability test this week — that single number decides everything downstream.