November 2026

Where venture wins in AI.
Which parts suit venture capital, and which need other kinds of capital.

The AI stackWho should fund each layer, at seed
Security and governanceVenture
Vertical and process agentsBest risk and return, in my view
Horizontal appsSelective · many markets taken
Infrastructure and toolingSelective · needs a technical edge
Below this line: too capital-heavy for a generalist seed cheque
Frontier modelsStrategic capital
Compute and data centresInfrastructure finance
ChipsStrategic and industrial
Venture fits the top. The rest needs other capital.
01Who is talking

What we see from 3,000 AI companies a year.

We read more than 3,000 AI companies a year and back a handful at pre-seed and seed.
My views line up with where I invest, so check them against the evidence.
Where this view comes fromSure Valley Ventures
Investing in AI since2017
AI companies read a year3,000+
StagePre-seed and seed
Part one
The stack, layer by layer
Which kind of capital each layer needs, and the risk and reward of each.
Part two
Where venture wins
Which industries are still open, and how to find your own.
Part three
What to ask a manager
The risks to screen for, where public capital is additional, and seven questions.
The question is where public venture capital earns its best return in AI.
Where venture wins in AI · November 2026A personal view, for information only: not investment advice or an offer of any fund · every source is in the field guide, Investing in AI
022026 in four numbers

Most venture capital now goes to AI,
and most of that to a few companies.

AI venture · 2026Global and US
AI share of global VC, Q3 202664%$102bn of $159bn
US value in $100m+ rounds, H1 202687.5%Fewer, larger rounds
Forbes AI 50 capital in two labs~80%OpenAI and Anthropic
US share of global AI VC77%Since 2023. Europe: 11%
Late-stage AI funding, 2025
United States
$141bn
Europe
$12bn

Take Europe as the example. It trails the US at every stage, and by most at the late stage. The US figure includes OpenAI’s $40bn and Anthropic’s $16.5bn rounds; without them it is still about seven times Europe’s (my calculation). About three-quarters of European AI funding goes to applications.

Where venture wins in AI · November 2026Crunchbase Q3 2026 · PitchBook-NVCA H1 2026 · Forbes AI 50 2026 · Dealroom/Prosus, Mar 2026
03The agentic shift

AI is moving from helping with the work to doing it.

Most companies now use AI. Very few have changed how the work gets done.
By “agent” I mean software you brief like a colleague: it does the whole job and hands it back for sign-off.
TODAYTHE SHIFT
Same claim: hours become minutes, and the handler checks and approves
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 every source cited, for sign-off.
Who gains
As firms hand over more work, value moves away from the model
AGENT-FIRST COMPANIES
Built to do the whole job. Being formed now, at seed.
THE SYSTEM AROUND THE MODEL
The workflows, controls and memory a firm builds. This is what improves over time.
MODELS
Powerful and getting cheaper. Anyone can rent the same model.
The best model won’t win. The best operating system will.
Where venture wins in AI · November 2026
04The market

Agents get paid out of the wage bill.

Software sold tools and took a small share of what a job costs. Agents sell finished work and are priced against salaries.
The market, to scale
US wages a year · BLS 2024
$11.7T
$450Bcore business software and its upkeep, a year (Bessemer, 2025)
SaaS captured 1–5% of an employee’s cost
tools: the human still does the job
Vertical AI examples capture 25–50%
completed work: drafted contracts, processed claims, closed books (Scale Venture Partners)
That is why vertical AI is sized at ≥10× legacy vertical SaaS (Bessemer, Jan 2026).
What this changes at seed
They can charge per job done
Revenue grows with the work completed, priced against the salary it replaces.
What protects them is built around the model
Their workflow, customer knowledge and records are their own, and get better with each job.
Where venture wins in AI · November 2026BLS QCEW 2024 · Bessemer, AI systems of action, May 2025 · Scale Venture Partners · Bessemer, Vertical AI, Jan 2026
PART ONE

Each layer needs a different kind of capital.

The higher up you go, the less capital a company needs, and the better it suits a venture cheque.

05The stack, at seed and Series A

Venture fits the top of the stack.
The rest needs other capital.

LayerWhat it isCapital that suits it
Security and governance
The checks every agent must pass before it goes live: identity, guardrails, audit trails
Seed and Series A venture
Vertical and process agents
Software that does a whole job: Legora and Harvey in law, Sierra in customer service, Cursor in coding
Seed and Series A venture: best risk and return
Horizontal apps
Assistants for every employee: ChatGPT Enterprise, Microsoft Copilot, Glean
Selective venture, then growth capital
Infra and tooling
The software between models and apps: data, connections, testing
Selective venture, hard technical edge only
The line · what a seed cheque can buy
Frontier models
The large models everything runs on: OpenAI, Anthropic, Google, Mistral
Strategic and late-stage capital, or the state
Compute and data centres
Where models are trained and run: the Microsoft, Amazon and Google clouds; CoreWeave, Nscale
Infrastructure finance and debt
Chips
The processors: Nvidia, and challengers such as Fractile
Strategic and industrial capital
Where venture wins in AI · November 2026SVV view · the risks and the evidence for each layer are in the field guide, Table 2
06Inside the four areas

What sits inside the four areas venture can back.

Agents for one industry, agents for one business function, the tools agents need, and AI security.
Year shown: when we expect a seed company can start winning customers there
01 · Vertical
Vertical agent platforms
Built for one industry: the software does the routine work and the professional makes the decisions.
Public sector
central government opens 2029
Health care
scribes established · trials 2027
Finance & insurance
claims 2026 · banking 2027
Professional services
Opens 2026audit, tax
Information & security
security operations
Transport & logistics
supply chain & logistics 2028
Science & research
Opens 2026lab and discovery
Manufacturing
Opens 2028plant operations
Energy & utilities
Opens 2028grid operations
Construction
Opens 2029estimating, sites
Retail & hospitality
commerce and guest operations
Education & media
learning and media production
02 · Cross-industry
Process agents
Agents that run one business function end to end, across industries. Coding is the largest so far, and already has leaders.
Sales & revenue
pipeline, execution, forecasting
Partner & channel revenue
partner and channel programmes
Marketing & AI search
campaigns and AI search visibility
Customer service & success
service, retention, engagement
Finance & accounting
Opens 2026close, audit, tax
Compliance, risk & onboarding
onboarding and risk checks
Procurement & supply chain
Opens 2028sourcing, ordering
HR & people
hiring, onboarding, workforce
Legal & contracts
drafting, review, obligations
Operations automation
back-office work end to end
03 · Enabling
Agent infrastructure
Connecting agents to systems
APIs, middleware, integration
Agent payments & identity
how agents pay and prove who they act for
Testing, simulation & deployment
before an agent is trusted in production
Efficient models & training
smaller, cheaper, faster models
Data, media & tooling APIs
what agents consume and produce
04 · Security
AI security & trust
needed by every layer above
Encrypted inference
models that never see the data in clear
Enterprise AI security
detecting AI-driven attacks
AI-agent security
identity, permissions, runtime guardrails
AI governance & compliance
EU AI Act, audit trails, model risk
Supply-chain security
model and data provenance, poisoning
Below the line
Foundation-model training · hyperscale computeCapital-intensive, built by large technology firms. The line is set by how much capital it takes.
Back industry agents and the tools they need. Build security into the product from the start.
Where venture wins in AI · November 2026SVV view of the investable layers · opening years from SVV, A Thousand Verticals, Oct 2026
07Two kinds of capital

Sovereign AI and venture AI need different tests.

Below the line · infrastructure and sovereign · Europe as the example
€200bnInvestAI mobilisation target, including €20bn for AI gigafactories. €21bn+Mistral’s valuation at its €3bn Series D, led by Samsung, with the Scaleup Europe Fund co-leading. $800bn+Hyperscaler capital expenditure expected in 2026.
Above the line · venture capital
63%Start-ups’ share of enterprise application-layer AI spend in 2025. <18 moLegora, from $1m to $100m of annual recurring revenue. $3.5bnAll vertical AI spend in 2025, against $11.7T of US wages.
Judge national model programmes on national goals and commercial discipline; hold venture programmes to returns and additionality.
Where venture wins in AI · November 2026European Commission and EuroHPC · Orrick, Sep 2026 · Hargreaves Lansdown · Menlo Ventures · Legora
PART TWO

Most of the market is still open.

We mapped about a thousand industries. In almost all of them, no AI company has pulled ahead yet.

08One square per industry · SVV’s indicative assessment

Of 1,012 industries, nine have a clear AI leader.

One square per industry in the 2022 North American classification (six-digit NAICS). A vertical is an industry and its workflow: claims handling inside insurance, conveyancing inside law.
software engineering · Cursor · Copilot
investment banking · Rogo
legal · Harvey · Legora
contact centres · Sierra · Decagon
Click a category · point at a square
100 are contested and 169 are open today.
Where venture wins in AI · November 2026SVV, A Thousand Verticals, Oct 2026 · NAICS 2022 · open now = a seed company could win paying customers within 12 months
09Your own markets

Start with the industries you already know.

Where the open verticals cluster
Financial-services operationsComplianceInsurance distributionPublic administrationPolicing and justiceBuilding safetyPropertyWholesale distribution

Regulated, paperwork-heavy work bought locally. A fixed regulatory deadline turns a pilot into a contract.

US wages, open verticals$3.15tn
All vertical AI spend, 2025$3.5bn
Four steps for your own country
1List the large industries at home where you have deep expertise.
2Mark those with a national regulator, procurement system or language a US vendor can’t easily serve.
3Check whether a local company leads, or only foreign players.
4Ask your managers who leads the first rounds there.
Where venture wins in AI · November 2026SVV analysis, Oct 2026 (BLS OEWS May 2023 wages) · Menlo Ventures, Dec 2025
10Where to invest, what to avoid

What to back at seed, and what to leave alone.

Invest · at the first cheque
Vertical and process agents in open industriesA named buyer and budget, a workflow the company owns, a regulator or language that favours a local team.
Security, evaluation and governance for agentsEvery agent deployment has to pass it, whoever built the model.
Infrastructure and tooling, selectivelyOnly with a technical edge the labs and clouds will not copy.
Avoid · or leave to other capital
Frontier models, compute and chip fabrication at seedDifferent capital: strategic, infrastructure finance and industrial.
Verticals that already have a leaderLegal, coding, contact centres: enter only where the leaders do not serve.
Thin wrappers, and services dressed as softwareA chat product on someone else’s model, or people doing the work behind the demo.
DEFENCE AND DUAL-USE
Back the software layer at seed: sensing and analytics, autonomy, secure inference at the edge. Accreditation takes time, so expect longer timelines. Platforms and hardware need other capital.
Back companies that do a customer’s whole job.
Where venture wins in AI · November 2026SVV view, drawn from the stack, the thousand industries and the risk screen in this deck
PART THREE

The risks, the funding gap and what to ask.

Most of these risks can be spotted before you invest. Public capital can help first rounds happen when private investors would otherwise hold back.

11Risks and red flags

Six risks, and what to check for each.

RiskWhat it looks likeWhat to check
Thin wrappers
A generic chat product on someone else’s model
Does it own a workflow and its data?
AI labs move in
The model provider ships the same feature
Does value rise as models get cheaper?
Margin
AI products project ~53% gross margin for 2026; SaaS 75–85%
Margin after inference, per completed job
Revenue quality
Paid pilots reported as recurring revenue
Contracts separated from pilots
Project failure
40%+ of agentic projects cancelled by end-2027
A narrow task, with a known cost when it fails
Your existing software holdings
Seat-based software that agents replace
Which holdings own the workflow?
Margin is the risk most investors miss: running the model is a real cost.
Where venture wins in AI · November 2026ICONIQ State of AI 2026 · Gartner, Jun 2025 · Wallonie Entreprendre AI call, 2026 · IMF GFSR, Oct 2026
12Where public capital is additional · Europe and the UK as the example

The US leads at every stage.
The pipeline starts at the first round.

Large sums are easier to find once a company is clearly winning. One priority for public capital is first rounds that private capital would not otherwise lead.
Under €0.5m

Grants, accelerators, angels

Existing funding routes

Potential first-lead gap
€0.75–3m

First institutional lead

In my experience, generalist funds wait for revenue

Series A–B

The next rounds

US capital often prices them

Series C+

Growth

Short of the US. New: ETCI 2.0, Scaleup Europe Fund

The data at each step · US against UK, venture investment as a share of GDP, 2023–25BBB Small Business Equity Tracker 2026
UK seed-stage deals, 2025−27%704 deals, the sharpest fall of any stage
Seed · US ahead by1.15×Widening; 14.4 months between rounds
Early stage · US ahead by1.31×Series A and B
Late stage · US ahead by1.35×US totals heavily influenced by AI megadeals
A company that never finds its first lead never raises a Series B.
Where venture wins in AI · November 2026BBB Small Business Equity Tracker 2026 (PitchBook; Beauhurst) · EIF, Jul 2026 · EQT, May 2026 · Dealroom/Prosus, Mar 2026
13The takeaway

Seven questions for an AI fund manager.

1Which workflows do your companies own, and who at the customer pays and answers for it?
2Can the agent finish the job, and what happens when it is wrong?
3Does each company get more valuable as models get cheaper?
4What is gross margin after inference, per completed job?
5How do you separate recurring revenue from paid pilots?
6Show us the companies you passed on, and why.
7How do entry prices, ownership, dilution and reserves turn into returns at this fund size, and who leads the next round?
Strong and weak answers to each are in the field guide.
Where venture wins in AI · November 2026

Where venture
wins in AI.

Barry Downesbarry@svv.ai · svv.ai/barry

Every source is in the field guide, Investing in AI.

This presentation sets out my personal view as Managing Partner of Sure Valley Ventures (SVV). SVV invests in early-stage AI companies, some of which may benefit from the developments described here. Nothing in it is an offer of, or an invitation to invest in, any fund. It is for information only and is not investment advice.

Three things to do
  1. Match capital to the layer.
  2. Test the manager’s company and fund economics.
  3. Test additionality in your own market.
→Spacenext step or slide
←previous slide
Ffullscreen
Gslide grid
Pprint or save as PDF
?this help
or use the arrow keys