software changesinterfaces changemodels improvepolicies changeworkflows evolveRoger keeps re-validating the work
The models become interchangeable. The company's learned way of working does not.
We don't depend on owning the model; we own the loop around it. Cross-customer learning only where permitted, aggregated, and appropriately governed.
The customer owns what Roger learns. Skillcast keeps making what it learned more useful.
Teamwho owns what
A small operating core, with company-builders around it.
Company-building partners + governance
Brandon Ward
Strategic operating partner · product + distribution
Founder + CEO, Keysha.ai
Product simplification, distribution strategy, digital growth systems, operating-pattern transfer from building a live AI product. Keysha and Roger share a thesis: AI should adapt to the person. Keysha explores it for the individual; Roger applies it to company work. Separate companies, separate IP.
Terrance Jones
Board nominee · capital strategy + governance
Equity, not salary
Financing architecture, dilution and ownership planning, diligence readiness, board evolution, sequencing seed to A to B. Terrance builds the capital system. Chris leads the raise. Professional engineer with finance operating experience.
Operating core · full-time
Chris Cahill
Product, category + distribution
owns
Product vision, the category and the story, founder-led sales and partnerships, the capital narrative and this raise
proof
Built OAISIS through real AI implementations for operational businesses; Roger came out of that implementation pain
Logan Choi
Autonomy systems + AI architecture
owns
Roger's learning system, autonomy architecture, execution engine, model strategy, technical direction
proof
Built the core structured-capture and execution architecture
Eddie Quiñones
AI platform engineering · reliability + integrations
owns
Production architecture, backend systems, integrations, reliability, verification infrastructure, deployment engineering
proof
Backend, microservices and systems architecture focus
First seed hiressenior agent and computer-use systems engineerevaluation and verification leadforward-deployed onboarding leadEvery person owns an outcome.
The game planand the ask
$6M to prove the first leg. The rest is earned.
Gate 1 · prove learning
·Real workflows, real operators
·Interventions per accepted run, measured
·First repeated autonomous runs
target: 0 to 60 days
Gate 2 · prove use
·Day-30 return, multiple workflows per account
·Repeated autonomous work
·Cost per run, measured
target: 2 to 4 months
Gate 3 · prove economics
·Customers pay, teams expand
·Measured path to 60%+ gross margin at scale
·One channel repeatable, CAC and payback visible
target: 4 to 9 months
Gate 4 · prove enterprise
·Governance, audit, security
·Private, model-portable deployment
·First implementation, then expansion
target: 6 to 12 months
Use of funds · capital raised upfront, spend accelerates as evidence appears
People: current team, senior hires, recruiting · $3.2M54%
Runway is 18 months. The learning is not: gates 1 and 2 are designed to resolve inside the first four. Gate success accelerates spend: learning unlocks systems hiring, use unlocks distribution, enterprise pull unlocks security and deployment. Larger rounds are earned by proof.
Day 100 becomes Roger’s Day 1.Chris Cahill · Skillcast
Capture a human demonstration, run it as an agentclaimedscripted flowsnot the modelbuilt
Structured UI state first, vision when requirednot publicly claimedselectorsvision-firstbuilt
Spoken rationale and corrections bound to actionsnot publicly claimednot publicly claimedprompt-levelbuilt
Verification of the learned method, not just the outcomeoutcome checksaudit logsnot publicly claimedin development
Customer-owned operating recordlocal Mac dataenterprise-heldprovider-helddesigned in
Model portabilitynot publicly claimedpartialsingle vendordesigned in
Private deployment and enterprise governancenot publicly claimedyescloudplanned, gate 4
Infragrid: infragrid.ai, August 2026, a16z Speedrun. Frontier computer-use: OpenAI and Anthropic products and APIs. “Not publicly claimed” means we could not find the claim on public pages, not that the capability is absent. Roger tags: built, in development, planned.
Appendix Bunit economics model
appendix
Cost of a run, and what moves it. all values modeled, replaced by telemetry
Variable COGS per verified run, by complexity
Low: short workflow, structured state only$0.02 to $0.08
Medium: 8 to 20 steps, 1 to 3 frontier decisions$0.08 to $0.25
High: vision steps, retries, long context$0.25 to $1.00+
Includes model inference, vision and tool calls, retries, verification pass, infrastructure and storage. Attributable support added when material. Built on official OpenAI and Anthropic list pricing, August 2026, and OSWorld-style step counts (human trajectories average 5 to 14 steps).
Account gross margin, distinct from run COGS
Account revenue − account COGSthe account metric
Self-serve launch pricing under test$20 Pro, $100 Max
Bring-your-own-model shifts inference off our COGSplatform and service costs remain
Benchmarks: AI-native median GM 2026about 52%; durable near 60%
Gate 3 is a measured path to 60%+ gross margin at scale, driven by structured capture, routing, caching, smaller models, local execution and falling inference prices. We will not trade product-market fit for margin in year one.
Try itruns per month 80cost $4.00margin on a $20 plan 80%healthy margin
Sources: OpenAI and Anthropic official pricing pages, August 2026. Margin benchmarks: ICONIQ and Bessemer state-of-AI reports, 2026. Step counts: OSWorld benchmark literature. Calculator uses central cost values per complexity band.
Appendix Centerprise trust architecture
appendix
What has to be true before a company hands over the work.
built
Control
·Intentional observation: capture only while the key is held
·Confirmation before external writes
·The Record: every run, every reason, every correction
gate 4
Governance
·Permissions per workflow, system, person
·Identity and SSO
·Audit export, retention, policy
·Data controls: excluded apps, fields, redaction
gate 4
Deployment
·Laptop, team cloud, private cloud, company servers
·Model portability: frontier, open, private, local
·Verification runs wherever the model runs
Gate-4 items are the definition of the enterprise proof gate, built against the first serious implementation.
Appendix Dcapital strategy
appendix
Every round has a job.
now
Seed · proof
·Risk retired: does judgment transfer into accepted autonomous work, do users return, do the economics hold
·Capability unlocked: measured learning loop, first repeatable channel, enterprise trust foundation
·Capital accelerates a loop that is already working
earned
B and growth · category scale
·Risk retired: category leadership contested on distribution, not on whether the loop works
·Capability unlocked: private and global deployment, ecosystem, the standard control layer for autonomous work
Monetized leverage: how a seed dollar compounds
Usage revenue scales with work performed, not headcount. Every accepted run is billable capacity with no salesperson attachedruns
Expansion compounds inside the account: operator to workflows to team to business unit, designed for net revenue expansion design targetaccounts
Verified outcomes unlock outcome pricing: conformance turns usage into billable, auditable completed workpricing power
Proof reprices capital: gate-based spend converts dollars into evidence, evidence into cheaper future dollars, less dilution for the same plancapital
The seed buys proof. Proof buys leverage. Leverage buys the category.
Round sizes are set when the proof exists, not before. Roger is the product. Skillcast is the operating company being financed. OAISIS is the origin studio. Entity, IP assignment and cap table detail live in the data room.
01 / 15
Roger
by Skillcast
Show Roger how you work. Then let it work.
Do the job normally. Explain what matters. Correct Roger when it's wrong. It picks up your way of working and starts taking the work from there.
Full self-driving for work. One workflow at a time.
Raising a $6M seedConfidential · Sept 2026
scroll
The discovery
Every company has someone who just knows.
Building AI for a law firm, we depended on one experienced paralegal.
Every few cases she would stop us.
“No. This one’s different.”
We asked where that rule was written down. It wasn’t. She just knew.
·Skip this one. It was handled last week.
·Check the account first for this client.
·Hold it if the totals don’t match. Don’t guess.
·Before sending, run it past the partner.
Day 1, you learn the rules. Day 100, you know when they don’t apply.
The best way to do the work lives in the people doing it.
The model isn’t missing intelligence. It’s missing your company.
The breakthrough
Show it. Correct it. Let it run.
Action
She skips a step while doing real work. Roger reads structured controls first, vision when the interface requires it.
+
Reason
“We handled this last week.” Said out loud, bound to that step.
+
Correction
“No. For this customer, check the account first.” Roger keeps it.
=
Verified outcome
Did the run match the learned standard? Not just: did it finish.
Your corrections teach it what right means here.
The productone moment
Before it acts, Roger shows its reason, in your words.
·Read the queue ran clean
·Match lines to the PO ran clean
·Line 12 · freight holding
Hold line 12 and send it back?
Because you told me: “Mercer disputes the freight line every time. Hold it.”from your demonstration · Aug 12
↵ approveN correct
demonstrationscorrectionsaccepted run
The first scoreboard. Launch establishes the baseline.
Trust is built one visible reason at a time.
Where we standSept 2026
Roger works today. Now we prove convergence.
working now
·Mac app, held-key intentional observation
·Structured UI capture via the Accessibility API
·Spoken rationale and corrections bound to steps
·Workflow replay and computer execution
being hardened
·Execution reliability across changing interfaces
·Browser reliability, onboarding
·Verification harness, cost instrumentation
the proof
How quickly can a real person transfer a real workflow into accepted autonomous work?
Real workflows. Interventions counted per run. A second operator. Repeated runs against live systems.
interventions per accepted rundemos to first clean runrepeated verified work
The first scoreboard starts at launch.
Why nowand how big
AI learned to reason. Then to use computers. Now it has to learn your company.
01Models
Rapidly improving intelligence
02Agents
Rapidly improving action
03?
No standard yet exists for transferring company judgment into verified autonomous work
12% → 66%
agents on real computer tasks, 2024 to 2026
0%
orgs using AI, vs single-digit agent deployment
0
US firms, 20 to 499 employees: the initial universe
$0B
agentic spend by 2030: category tailwind, not our TAM
Capability is arriving faster than organizations can safely absorb it.
Stanford HAI AI Index 2026 · US Census SUSB · Gartner
Business model
We make money when Roger does work.
Self-serve · distribution
Subscription plus a refilling execution allowance. Capture is free. launch pricing in appendix
Team · expansion
Shared workflows, pooled execution, controls for the manager. priced through pilots
Enterprise · economic upside
Annual commitment, governance, deployment choice. Outcome pricing as verified work becomes measurable. after first implementation
1 operator5 workflowsteambusiness unitenterprise control layer
Two numbers we run on: COGS per verified run, and account revenue minus account COGS. Modeled today, telemetry at launch.
Customer value compounds with work completed. Our revenue can compound with it.
Go to market
One workflow is the wedge. Expansion is the business.
Land with the operator. Expand through the work: more runs, more workflows, more people, more governance, larger accounts.
Now
Founder-led design partners · OAISIS customer relationships · direct introductions
Employer acquisition through eligible Texas workforce-training partners, where qualifying training can be externally funded · institutional introductions in development
Later
Enterprise field sales · PE portfolio distribution · strategic partnerships
Channels listed in order of maturity, not equal weight.
The moat
Models change. Operating memory compounds.
The customer owns
·Demonstrations, reasons, corrections
·Accepted outcomes, audit history, permissions
·The record of how work is done right
FrontierOpenPrivateLocalinterchangeable
Skillcast owns
·Capture architecture, skill representation
·Execution runtime, interface adaptation
·Verification, model routing, continuous re-validation
The models become interchangeable. The company's learned way of working does not.
The customer owns what Roger learns. Skillcast keeps making what it learned more useful.
Teamwho owns what
A small operating core, with company-builders around it.
Company-building partners + governance
Brandon Ward
Strategic operating partner · product + distribution
Founder + CEO, Keysha.ai. Product simplification, distribution strategy, digital growth, operating-pattern transfer from building a live AI product. Separate companies, separate IP.
Terrance Jones
Board nominee · capital strategy + governance
Financing architecture, dilution planning, diligence readiness, board evolution. Terrance builds the capital system. Chris leads the raise. Equity, not salary.
Operating core · full-time
Chris Cahill
Product, category + distribution
Product vision, the story, founder-led sales and partnerships, the capital narrative and this raise. Built OAISIS through real AI implementations; Roger came out of that pain.
Logan Choi
Autonomy systems + AI architecture
Roger's learning system, execution engine, model strategy, technical direction. Built the core structured-capture and execution architecture.
Eddie Quiñones
AI platform engineering · reliability + integrations
Production architecture, backend, integrations, verification infrastructure, deployment engineering.
+ senior agent-systems engineer+ evaluation and verification lead+ forward-deployed onboarding lead
The game planand the ask
$6M to prove the first leg. The rest is earned.
Gate 1 · prove learning · 0 to 60 days
Real workflows, interventions per accepted run measured, first repeated autonomous runs
Gate 2 · prove use · 2 to 4 months
Day-30 return, multiple workflows per account, cost per run measured
Gate 3 · prove economics · 4 to 9 months
Customers pay, teams expand, measured path to 60%+ gross margin at scale, one repeatable channel
Gate 4 · prove enterprise · 6 to 12 months
Governance, audit, private model-portable deployment, first implementation then expansion
Use of funds · $6M · ~18 months
People: team, senior hires54%
GTM and deployment17%
Infrastructure and evaluation12%
Enterprise trust7%
Reserve10%
Runway is 18 months. The learning is not: gates 1 and 2 resolve inside the first four. Gate success accelerates spend. Larger rounds are earned by proof.
Day 100 becomes Roger’s Day 1.
Chris Cahill · Skillcast
Appendix
A · competitive landscape
InfragridRogerDemonstration to agentclaimedbuiltStructured state firstnot publicbuiltReasons bound to actionsnot publicbuiltMethod verificationoutcome onlyin devModel portabilitynot publicdesigned inPrivate deploymentlocal Macgate 4
“Not public” means not claimed on public pages, not absent. Full table vs UiPath and frontier computer-use APIs in the desktop deck.
B · unit economicsall modeled · try it
The economics of a run, in your hands.
Workflow complexity
Runs this person uses per month: 80
Cost to serve
$4.00
On a $20 plan
80%
healthy margin
Cost per run: low $0.02 to $0.08, medium $0.08 to $0.25, high $0.25 to $1.00+. Central values shown. Built on official OpenAI and Anthropic list pricing, August 2026; replaced by telemetry. Heavy users can bring their own model, shifting inference off our COGS.
Every lever above pushes the cost line down while pricing holds. Gate 3 is a measured path to 60%+ gross margin at scale. Self-serve launch pricing under test: $20 Pro, $100 Max.
C · enterprise trusttap each pillar
What has to be true before a company hands over the work.
Controlbuilt+
·Intentional observation: capture only while the key is held
·Confirmation before external writes
·The Record: every run, every reason, every correction
Governancegate 4+
·Permissions per workflow, per system, per person
·Identity and SSO
·Audit export, retention, policy controls
·Data controls: excluded apps, fields, redaction
Deploymentgate 4+
·Laptop, team cloud, private cloud, company servers
·Model portability: frontier, open, private, local
·Verification runs wherever the model runs
Gate-4 items are the definition of the enterprise proof gate, built against the first serious implementation.
D · capital strategymonetized leverage
The seed buys proof. Proof buys leverage. Leverage buys the category.
How a seed dollar compounds
proofusage revenueaccount expansionoutcome pricingcheaper next dollar
Tap the chain, or watch it run.
Four leverage points
1Usage revenue scales with work performed, not headcount. Every accepted run is billable capacity with no salesperson attached.
2Expansion compounds inside the account: operator to workflows to team to business unit. The expansion ladder is the revenue plan, designed for net revenue expansion. design target
3Verified outcomes unlock outcome pricing: conformance turns AI usage into billable, auditable completed work, the pricing enterprise buyers accept.
4Proof reprices capital: gate-based spend converts each dollar into evidence, and evidence into cheaper future dollars. Less dilution for the same plan.
Seed · proof
Does judgment transfer, do users return, do the economics hold. Unlocks the measured loop and the first repeatable channel.
Series A · repeatability
Revenue repeatability, account expansion, enterprise readiness. Capital accelerates a loop that already works.
B and growth · category scale
Private and global deployment, ecosystem, the standard control layer for autonomous work.
Larger capital is earned by proof. Round sizes are set when the proof exists, not before.
Roger is the product · Skillcast is the operating company being financed · OAISIS is the origin studio