SVV
Pitch SVV →
Thesis

Work is becoming a market.

Agentic AI is a step change. Software is starting to do the work, not just speed it up, and we think the value will go to systems built for one profession at a time.

01The shift

The next shift: AI that does the work.

For decades, enterprise software made a person faster. The analyst still built the model and the claims handler still handled the claim. An agent is software briefed like a colleague: it does the whole job and returns the work for sign-off. The human moves from production to review.

Today · AI that assists

“Summarise this claim file.”

Useful. Ten minutes saved. The claim is still the handler's to process.

The shift · AI that does the work

“Process this claim end to end: check the policy, gather the evidence, draft the settlement. Cite everything.”

An agent runs the whole workflow and returns the finished work, with receipts, for your sign-off.

Hours → minutes
↑Where the value accrues

As work is delegated, advantage moves up the stack.

Top · forming nowAgent-first companiesBuilt for delegation from day one. Being formed now, at seed.
Middle · compoundsAgent operating systemsWorkflows, governance, memory. What a company builds around the models.
Base · rentedModelsPowerful, abundant and commoditising. Everyone can rent the same intelligence.
Why now
ReasoningModels break down goals and recover from failure.
Tool useAgents read and write the systems of record.
Context & memoryEvery completed job can improve the next.

Own the layer above the models.

02The money

Agents compete for labour budgets, not IT budgets.

Software sold tools and captured a sliver of value. Agents sell completed work, and price against the payroll line.

AWhat incumbents sell$450B

Seats. The old software market.

BWhat agents sell$11.7T

Work. Annual US wages alone. The labour budget is the market.

Software $450BUS labour $11.7T · drawn to scale
SaaS captured1–5%of an employee's cost. Tools: the human still does the job.Scale Venture Partners
Vertical AI captures25–50%Completed work: drafted contracts, processed claims, closed books.Scale Venture Partners, early examples
Sell work, not seatsPriced on outcomesRevenue scales with jobs completed, and expands as agents take on more of the role.SVV thesis
Software was the beachhead. Work is the market.Bessemer, Vertical AI · Jan 2026
03The field

A thousand verticals. Almost all still open.

The US industry classification lists 1,012 industries. Treat each one as a vertical and only two are close to having a winner: legal, with Harvey and Legora, and software engineering, with Copilot and Cursor. The other thousand are open, and the map is still being drawn.

1,012Verticals
2Close to won
1,000+Still open
Test 1 · design partnersWould a company founded today win design partners this year?Live pilots, AI budgets, buyers who can sign without long procurement, and a record of working with startups.
Test 2 · paid conversionAnd would they pay?A labour cost the product clearly replaces, rules that allow the work to be delegated, and room left by incumbents and funded rivals.
04What to build

Build the system of work, not the model.

Frontier models are powerful, rented and improving fast. Every founder can rent the same intelligence. What compounds is the system of work built around them, for one profession, owning the workflow, the context and the customer.

01ContextProprietary data, procedures and precedents.
02WorkflowOwning the job end to end, with sign-off built in.
03Evaluation dataEvidence of what good looks like, job after job.

Model portability has to be proven, not assumed. Learning rights have to be contractually real.

See how we decide →
Anatomy of a vertical AI operating systemread bottom to top
The learning loopEvery completed job feeds back. The system improves with use.
Collaboration & governanceTeams, permissions, review, accountability
Evidence & auditProvenance, citations, compliance-grade outputs
Workflow orchestrationLong-running jobs with human sign-off gates
Skills libraryReusable domain procedures, specialist sub-agents
Specialist toolsIntegrations to systems of record
Domain contextProprietary data, ontologies, the regulatory corpus
Reasoning coreFrontier models, usually rented. Not a moat on its own.
05Where we invest

Four lines. One network.

We invest where enterprise AI becomes a product. Everything we back runs on one of four lines, and they all meet at SVV. Founders on different lines share customers, hires and introductions. Tap a dot to see the companies there.

Swipe to see all four lines

↓Below the lineFoundation-model training and hyperscale compute. Capital-intensive and built by large technology firms, so not where we put capital. The line is capital intensity, not “models”: efficient models that make agents cheaper to run sit above it.
06The clock

Once a category forms, it closes fast.

Text-native, reviewable work forms first. Adoption broadens as reliability and governance improve, and consolidation follows durable workflow ownership. Legal went from first product to a crowned pair of leaders in about three years. The window is open now.

Timeline · SVV scenario, not a forecast● Refreshed as evidence changes
To 2027SetupTrust and runtime layers mature.Now
2027–28ValidateWorkflow evidence accumulates. Forming categories produce their first hard proof.Next
2028–30ScaleCategories and competitive structures become clear.Later
2030+ConsolidateLeaders mature and acquirers arrive.Later
!Risks

What would prove us wrong

We track six pressures. If they dominate, the thesis is wrong, not just early.

Models climbLabs absorb the workflow.
Incumbents bundleGood enough wins in the installed base.
Review holdsServices consume the margin.
Authority stallsBuyers won't delegate.
Data doesn't compoundThe learning loop is a contractual fiction.
Categories fragmentThe prize stays subscale.

The models will be rented.
The work will be owned.

If you're building the layer above the models, one vertical at a time, we'd like to hear from you.

Pitch SVV →For founders →