What changed
Two capability shifts landed together: (1) Google shipped computer use in Gemini 3.5 Flash, a low-cost, low-latency tier, making screen/GUI-driving agents economically viable per-task (FACT, per https://deepmind.google/blog/introducing-computer-use-in-gemini-3-5-flash/); (2) OfficeCLI provides headless, single-binary Office file manipulation on any server, removing the Microsoft Office dependency from agent pipelines (FACT, per https://github.com/iOfficeAI/OfficeCLI). A venture-backed startup, Coasty, claims to automate GUI-only legacy software without APIs or RPA scripting — a market-validation signal, though reliability is a vendor claim and unverified (FACT that the claim exists; HYPOTHESIS that it works).
Why now
Until now, driving a legacy GUI cost frontier-model prices per screenshot loop, killing unit economics for low-value SMB tasks. Flash-tier computer use changes the cost curve; Coasty's funded launch suggests investors believe the timing is right (HYPOTHESIS: their entry validates rather than saturates the market).
Converging signals
Cheap computer-use model tier (Gemini 3.5 Flash) + headless Office document tooling (OfficeCLI) + a funded competitor validating the 'automate API-less legacy software' wedge (Coasty). All three are capability/supply-side signals; the input contains NO demand-side signals.
Customer pain
HYPOTHESIS ONLY: SMBs running legacy desktop software without APIs pay staff to hand-key data, re-type reports, and assemble Office documents. This is a widely believed pain, but the provided input includes no demand_evidence array — no complaints, no job postings, no mandate. I cannot cite a single piece of evidence that a specific buyer is in pain right now.
Who pays
HYPOTHESIS: owner-operators of SMBs in verticals stuck on legacy Windows software — e.g., scrap/recycling yards on legacy yard-management systems, trucking back offices, property managers, medical/dental fronts (avoid regulated data). Charles's recycling and fire-service background suggests scrap yards and fire departments as niches where he has credibility, but no payer is evidenced in this input.
Solved today
Manual data entry by staff; traditional RPA (UiPath, Power Automate, Robocorp) which requires scripting, breaks on UI changes, and is priced/complex beyond small SMBs; or offshore VA services. (Industry-general knowledge — HYPOTHESIS relative to this input, which provides no spend evidence.)
Why current solutions are bad
RPA needs a developer to script and maintain selectors; SMBs can't afford implementation consultants. VAs are slow and error-prone. Computer-use agents promise 'describe the task, the agent drives the screen' — but agent reliability on real legacy GUIs is exactly the unverified claim in the Coasty signal, so 'why current is bad' cuts both ways: the new approach may also be too brittle to charge for.
Proposed product
NOT a horizontal platform (that's Coasty's fight, with VC money). Instead: a productized service in ONE niche Charles knows — e.g., 'we auto-key your scale tickets / inbound loads into your legacy yard-management system and generate your end-of-day Office reports, priced per document/task.' Charles operates the agents (Gemini Flash computer use + OfficeCLI for document output) on his own infrastructure; the customer just emails or drops files. Per-task pricing mirrors his proven ELDT per-upload model, except here no regulation forces anyone to buy — a critical difference.
MVP version
One niche, one workflow: pick a single legacy application common in scrap/recycling or another vertical he can access, record 5 real task examples from 1 design partner, build the agent loop (screenshot → Flash computer-use action → verify → OfficeCLI output), and deliver a week of completed tasks with an accuracy log. MVP is the service delivered manually-supervised, not software the customer installs.
30-day build
Days 1-7: demand test BEFORE building — post in 2-3 niche operator communities (scrap yard owners, trucking back office) and cold-message 20 operators he can credibly reach: 'what do you re-type into old software every day?' Success gate: 3 operators describe a concrete recurring task and agree to a paid pilot. If the gate fails, kill. Days 8-30: build the agent loop for the single winning workflow with one design partner; measure error rate honestly.
60-day build
Convert design partner to paid ($200-500/mo or per-task), harden the loop (retry logic, human-review queue for low-confidence actions), onboard 2-3 more customers on the SAME workflow/same software product. Publish before/after time-savings numbers as the sales asset.
90-day revenue plan
Target: 4-6 SMBs at $200-600/mo = $1k-3k MRR (HYPOTHESIS — no pricing evidence in input). Realistic only if the day-7 demand gate passed and the agent hits >95% task accuracy; below that, human-review labor eats the margin and this becomes a low-paid BPO job.
Distribution path
Weakest link. No forced buyer, no registry of affected parties (unlike ELDT). Distribution = direct outreach in niches where he has operational credibility (recycling/scrap networks, fire service) plus demonstrated-value demos (screen recording of the agent doing the actual task). No enterprise sales, which fits, but SMB direct outreach is slow and churn-prone.
Pricing hypothesis
Per-task/per-document (e.g., $0.50-2 per record keyed, $5-20 per generated report) or flat monthly per workflow. Per-task matches his ELDT muscle memory and matches Flash-tier unit economics; margin depends entirely on agent accuracy (HYPOTHESIS).
Technical difficulty
Moderate. The pieces exist (Gemini Flash computer use API, OfficeCLI, VM/Windows sandbox for the legacy app). Hard parts: hosting the customer's legacy app or accessing their machine safely, agent reliability on idiosyncratic old UIs, and verification/rollback when the agent mis-keys data into a system of record. Mis-keyed production data is a real liability.
Legal / regulatory risk
Low-moderate: no regulator, but customers hand over business data and system access; a data-entry error the agent commits into their books is Charles's fault contractually. Avoid any workflow touching PHI/PII-heavy or financial-filing systems at first. Also check legacy software EULAs — some prohibit automated access (HYPOTHESIS, must verify per vendor).
Platform dependency
High on Google's computer-use API pricing/availability and on OfficeCLI (a single GitHub project of unverified maturity). Mitigable — Anthropic/OpenAI have computer-use alternatives, LibreOffice headless is a fallback — but the cost advantage that makes this 'newly economical' belongs to a model vendor who can reprice it.
Founder fit
Mixed. Fits his strengths (automation, AI workflows, industrial ops credibility, per-transaction monetization, no enterprise sales) and his preferred product types (AI agents, niche operational tools). But it LACKS the proven-edge shape: no regulation compels anyone to buy, no deadline, no forced filer list. This is a convince-the-buyer play, not a they-must-file play, so it does not qualify for the VERY HIGH regulatory-wedge rating.
Breakout potential
If one niche workflow works, the same harness extends to adjacent workflows and adjacent verticals, and could productize into per-seat software. But Coasty and inevitable fast-followers with funding will compress the horizontal opportunity; defensibility is only niche depth and service quality.
Final recommendation
CONDITIONAL PASS — do not build now. The capability convergence is real, but this brief contains no demand evidence, no forced buyer, and an unverified reliability premise, and it lacks the regulatory-wedge shape of Charles's proven edge. Spend max 5-7 days on a demand test in one credibility niche (scrap/recycling back office). Build only if 3+ operators name a concrete recurring task and commit to a paid pilot; otherwise archive and instead hunt for a mandate-driven variant (e.g., a workflow where the legacy-GUI output must be filed with a government system — that hybrid would score far higher).
Next action
This week: message 20 scrap-yard/recycling operators and 2-3 niche communities asking 'what does your office re-type into old software every day, and what would you pay to never do it again?' — collect verbatim answers as the demand evidence this brief currently lacks.