For every $1 spent on software, $6 is spent on services.
Everyone is competing for the $1. We capture the $6.
Framework: Sequoia Capital — "Services: The New Software" (March 2026)
If you sell the tool, you're in a race against the model. Every improvement commoditizes you. If you sell the work, every improvement makes you faster, cheaper, and harder to compete with.
Sells the tool. Makes the professional more productive. Competes with the next model release. Captures the software budget.
Software for accountants.
Software for lawyers.
Software for insurance teams.
Sells the work. Every AI improvement compounds. Captures the labor budget — 6× larger than software.
The company that closes the books.
The company that reviews the contracts.
The company that handles the claims.
"The next $1T company will be a software company masquerading as a services firm."
— Julien Bek, Partner at Sequoia Capital
The total addressable market for autopilots isn't the software budget. It's all labor spend in a category — insourced and outsourced combined.
Where every AI startup is competing. QuickBooks charges $10K/year. Shrinking margin, model-dependent moat.
Where the actual spend is. That same company pays $120K for an accountant to close the books. The autopilot just closes the books.
Our competitors sell one layer. We use all seven to deliver the outcome.
Take the decision-making DNA of your best leader. Scale it across every operation simultaneously. Not a tool that helps the CEO decide — the system that executes with the CEO's judgment encoded.
Encode leadership DNA from coaching transcripts, decision patterns, and institutional knowledge into an autonomous delivery partner. One leader's judgment, infinite surface area.
1–3 humans + AI agents = autonomous product unit. Don't sell project management software. Deliver the project. Done.
95% of AI deployment fails because it's bolted onto dysfunctional structures. Restructure first (150→35 people), then operate with AI. The framework, not the feature.
Start with intelligence-heavy work (rules, patterns, process). As the system compounds domain data, the frontier shifts — today's judgment becomes tomorrow's intelligence.
7-day Deep Sessions with the leader (direction, judgment calls only they can make) → 11 weeks of autonomous execution. The leader does 20% (judgment). The system does 80% (intelligence).
Gain-share model. The client keeps their gross margin. We capture the delta.
"If what you're saying is that with your model, you are accelerating holding periods to get value — yeah, they would build a fund around you."
— Trent Johnson, CSO & SVP Corporate Ventures, Cie Digital Labs
Professional venture evaluator. Builds startups for PE firms.
The playbook: start with outsourced, intelligence-heavy tasks (vendor swap, not reorg). Expand toward insourced, judgment-heavy work as AI compounds.
TAM data: Sequoia Capital analysis, March 2026. US market estimates.
If a task is already outsourced, three things are true: the company accepted external delivery, the budget line exists, and the buyer already purchases outcomes. Replacing an outsourcing contract with an AI-native provider is a vendor swap. Replacing headcount is a reorg.
NDAs, compliance filings, billing codes, IT patching. Clear scope, verifiable output, existing budget. Vendor swap — no organizational change required.
PE operating partners → portfolio companies. Systems integrators → white-label for margin expansion. Managed services advisors → next-gen outsourcing. Three validated channels.
As the system compounds proprietary data about what good judgment looks like in a domain, the frontier shifts. Today's judgment becomes tomorrow's intelligence. The moat deepens with every engagement.
"It is hard to detect a weakness in the model other than time and your ability to scale quickly."
— Trent Johnson, after evaluating the model as a professional venture builder
In 2025, the fastest-growing AI companies were copilots. In 2026, many are trying to become autopilots. But they face the innovator's dilemma: selling the work means cutting their own customers out of doing it.
Have the product and customer knowledge. But every step toward autopilot threatens their existing customer base. The accountant who pays for the tool doesn't want to be replaced by it.
No installed base to protect. No innovator's dilemma. Sell to the company that needs the outcome, not to the professional who does the work. Start compounding domain data from day one.
The outcome package isn't just a product. It's a learning machine. Every client engagement makes the system smarter. Every rep deepens the moat.
Client gets a dedicated AI copilot instance. It serves their users, handles their workflows, saves their data. Not a tool they learn — an outcome they receive.
The system does the work — closes books, reviews contracts, handles claims, ships software. Each task is a rep. Each rep generates signal about what "good" looks like in that domain.
Human gate: ✅ accurate, ⚠️ partially correct, ❌ wrong. Not AI judging AI — human judgment as the only valid eval. Each review is a labeled training signal.
The system doesn't add rules — it subtracts what caused ⚠️ and ❌. Fewer wrong defaults = better output. The improvement model is subtraction, not accumulation.
Today's judgment becomes tomorrow's intelligence. The system accumulates proprietary data about what good judgment looks like — the moat Sequoia says creates trillion-dollar companies.
This isn't a thesis we're planning to test. It's a description of what's already running. Here's what's proven, and what's next.
Warren delivers software — not assists developers. Ships features, manages PRs, runs CI, handles multi-client delivery simultaneously. 4 active clients, concurrent execution.
Human review via #warren-review (✅/⚠️/❌) is the only eval. AI judging AI was killed. QA manager reviews output. Deterministic quality gates catch obvious violations. Real reps with real humans.
"Warren gives pointed answers, not lots of different solutions." — Steve Ward. PE observers "floored." Professional venture evaluator found only one weakness: "time and ability to scale."
7-day deep sessions → 11 weeks autonomous execution. The leader does 20% (judgment), the system does 80% (intelligence). Running in production. Mini-CEO pods active.
"So gain share. Yeah." — Trent Johnson validated the pricing without pushback. Client keeps gross margin, we capture the delta. Aligned with outcomes, not seats.
Trent mapped three distribution channels unprompted: PE operating partners → portcos, Systems Integrators → white-label, Managed Services Advisors → next-gen outsourcing. Customer mapped the GTM.
Need 1 engagement with published numbers (anonymized is fine for PE pitch). Before/after: team size, delivery cost, cycle time, quality metrics. The deck needs a number, not a narrative.
Today's RLHF is artisanal — humans correct, patterns stay in memory files. Scale requires: automated pattern extraction from reviews, domain-specific judgment capture, structured eval per outcome (not per message).
Current cap: 3 concurrent programs before human review gates max out. Trent offered two Slalom PMs (Igor, Philip). The bottleneck isn't AI — it's the review layer. Scaling the gate scales the flywheel.
Prove the flywheel empirically: measure if Warren's quality on client N+1 is measurably better than client N. Track ✅/⚠️/❌ ratios over time. The thesis says it compounds — the data needs to show it.
4 clients. Human review loop. Subtraction-based improvement. Digital Twin validated. Gain-share pricing validated. The flywheel turns manually — humans close every loop.
Automated pattern extraction. Eval per engagement (not per message). Domain-specific judgment rules generated from human corrections. The flywheel starts to self-feed.
Each vertical accumulates enough proprietary judgment data that new entrants can't replicate the quality. Today's judgment becomes tomorrow's intelligence. The convergence Sequoia describes.