When to open:You want to build B2B vertical AI / want to know how much funding a track has raised / want to see how representative companies operate. This wiki organizes 66+ B2B vertical AI cases from Shensi Circle, categorized by 8 major tracks (sales/customer service/recruitment/legal/healthcare/insurance/HR/finance).
One-sentence core:Unicorns have emerged across each of the 8 vertical tracks(Legal $700M / Customer service $1.5B / Sales $52M / Healthcare $24M / Insurance $30M / HR $15M / Finance $50M).Key model: Stop selling tools to lawyers, rebuild law firms directly with AI(Sequoia US).
0. Funding landscape across the 8 tracks (2024-2026)
| Track | Representative funding amount | Representative company type |
|---|---|---|
| Sales | $50M-$35M-$100M | Salesforce alternative / AI SDR / Sales operating system |
| Customer service | $200M+ ($1.5B valuation) / $50M / $21M | Chinese-founded company / Intercom / Decagon / Cresta |
| Recruitment | $50M-$20M-$15M | Recruitment operating system / All-in-one / Generalist |
| Legal | ~$700M (unicorn) / $60M / $20M | Harvey / Eve / Draftwise |
| Healthcare | $42M-$24M-$2.2M | Copilot / Family doctor / Shimmer mental health |
| Insurance | $30M-$15M | Insurance pain points / Agent insurance |
| HR | $15M | AI HR generalist + vertical model |
| Finance | $50M-$43.5M | Next-generation infrastructure / Finance-specific AI |
1. Sales (Sales Agent)
1.1 Flagship case: Peter Thiel's $35M bet to end Salesforce
"Peter Thiel invested $35 million in a Claude Code for the sales domain, claiming it will end the Salesforce era"
Implication:Agents end the SaaS tool era —— It is no longer about CRM tools for salespeople to use, but salespeople having AI do the work on their behalf.
1.2 Sequoia US $50M
"Sequoia US led a $50 million round: how AI Agents are redefining sales software"
1.3 Other sales cases
| Company | Data |
|---|---|
| AI sales company (2 rounds in 9 months) | $52 million |
| AI sales training (extremely niche) | $50 million |
| AI sales (former Stripe executive) | $22 million |
| 40+ Silicon Valley big names rushed to invest | $10 million / World's first AI operating system for salespeople |
| 23-year-old sales prodigy | 0 → $800 million (customer acquisition code) |
| AI sales $5.5M | Is using AI to help sales really feasible? |
| Two sisters left Meta to start a business | $10.6M / Turn social media into a sales channel |
1.4 a16z + Sequoia AI sales landscape overview
"In-depth | Overview of AI+sales products, Sequoia is betting on this track"
2. Customer Service (Customer Service Agent)
2.1 Chinese-founded giant ($1.5B valuation)
"Raised $200M at a $1.5 billion valuation, this Chinese-founded AI customer service company is incredibly aggressive"
2.2 $50M mega-round
"Another $50 million mega-round in the AI customer service track, from tools to platform ecosystem"
2.3 Intercom's self-reinvention
"In-depth analysis | Amid the rise of AI customer service, how legacy giant Intercom completed an AI-First self-reinvention" "Customer service software giant starts selling AI models, what does Intercom's move mean?"
2.4 Other customer service cases
| Company | Data |
|---|---|
| Danish customer service company(Used by Perplexity + Replit) | $21 million |
| Australian AI customer service | $35M / Swipes cards, transfers money, places orders for you |
| Seed round in 6 days → Series B in 1 year | 300% improvement in customer service efficiency |
| European voice customer service(15x growth in 1 year) | $20 million from top Silicon Valley VCs |
3. Recruitment (Hiring / Talent Agent)
3.1 All-in-one recruitment operating system ($50M)
"$50 million bet on 'all-in-one AI recruitment', intelligent recruitment operating system turns every HR into a data expert"
3.2 Other recruitment cases
| Company | Data |
|---|---|
| Gained 50,000 users in 6 months | $20 million seed round |
| Stanford Chinese dropout (100k+ LinkedIn followers) | $3 million |
| Inflection + Anthropic already using it | AI-native recruitment |
| Chinese founder (era of 3x surge in application volume) | $20 million / Let every candidate be seen |
| $20 million chooses a different path | Does not replace recruiters |
3.3 Industry war perspective
- "How AI startups can win the recruitment war"(In-depth long read)
4. Legal (Legal AI)
4.1 Flagship case: 25-year-old esports player builds $700M unicorn
"AI+Legal | 25-year-old esports player pivots to build a nearly $700 million legal AI unicorn"
4.2 Sequoia's new playbook: Don't sell tools, rebuild law firms directly
"Sequoia US invested in the legal track again, but this company no longer sells tools to lawyers; it rebuilds law firms directly with AI"
Key pivot:
- Old model:Harvey / Draftwise → Sell SaaS tools
- New model:Directly build an "AI-native law firm" → Operate as a law firm itself
4.3 Other legal cases
| Company | Data |
|---|---|
| Draftwise(Law firm intelligent system) | $20 million Series A |
| Another AI legal product | $60 million |
| 25-year-old gaming youth dropped out | $100 million (directly challenging OpenAI-affiliated players) |
5. Healthcare (Healthcare AI)
5.1 Cases
| Company | Data |
|---|---|
| AI Copilot for healthcare | $24 million Series B |
| Family doctor AI | $42 million |
| Shimmer(YC / Mental health) | $2.2 million |
| Founded by surgeons (used addictively by 2 million doctors) | $100 million in 18 months |
| 4x in three months → $5M(Harvard MIT top students) | Unlocking healthcare's biggest pain point |
| AI+Pharma clinical drug development | YC W24 |
6. Insurance (Insurance AI)
6.1 Flagship: $30M raised across consecutive rounds in 6 months
"Completed two rounds totaling $30 million in 6 months, this AI company found insurance's most valuable pain point"
6.2 Agent insurance (paying for AI mistakes)
"When AI Agents make mistakes, who pays? Raised $15 million, an insurance company built for Agents"
Implication:AI Agent explosion → New insurance demand → New track.
7. HR
"Raised $15 million, built an AI HR generalist, and also developed a dedicated vertical model"
Model:Generalist Agent + vertically fine-tuned model
8. Finance (Finance AI)
8.1 Cases
| Company | Data |
|---|---|
| Finance-specific AI(Former frontline Google/Meta scientists) | $43.5 million |
| Next-generation financial infrastructure | $50 million |
| AI tax company(Benchmark + YC) | 2 consecutive rounds totaling $53 million in 6 months |
| Anthropic enters finance | Strategic pivot |
8.2 Warning on Yu Zijiang-style wrapper apps
- "How a wrapper app replicating Cal AI for financial analysis reached $40,000 monthly revenue"
9. Cross-track commonalities
9.1 Old SaaS → New Agent transformation
| Old SaaS | New Agent model |
|---|---|
| Salesforce(CRM) | Peter Thiel's $35M "sales Claude Code" |
| Traditional accounting software | Sequoia $25M AI-native ERP |
| Harvey tool | Sequoia directly "rebuilds law firms with AI" |
| Recruitment ATS | Recruitment operating system ($50M) |
| Intercom | Sell AI models directly |
9.2 Key judgments
"Besides code and customer service, where will the next breakout scenario for AI agents be?"(Shensi Circle analysis)
Predicted answers:
- Sales (already taking off)
- Legal (unicorn born)
- Healthcare Copilot
- HR / Recruitment
- Financial infrastructure
9.3 Lessons from 11x's collapse
- AI SDR / digital employees must resolve organizational trust / permissions(per Jaya Gupta)
- Cannot just "automate tasks", must be production-ready
10. How to use —— B2B paths by track
| What you want to build | Primary reference | Starting point |
|---|---|---|
| Sales Agent | Peter Thiel + Sequoia $50M | "End Salesforce" / Sales operating system |
| Customer Service Agent | Chinese-founded $1.5B / Intercom transformation | "Tools → Platform ecosystem" / Service fees replacing software fees |
| Recruitment Agent | All-in-one recruitment OS | Recruitment operating system / Turn HR into data experts |
| Legal AI | Sequoia new model | "Don't sell tools, rebuild law firms directly" |
| Healthcare AI | Copilot $24M model | Clinical scenarios / Serving doctors |
| Insurance AI | Agent insurance $15M | Paying for new types of Agent risks |
| Finance AI | Finance-specific $43.5M | Founded by former Google/Meta scientists |
Universal disciplines:
- Unicorns have emerged across each of the 8 tracks, the tracks are not crowded
- Key pivot: Don't sell tools, rebuild industries directly with AI(Sequoia legal template)
- Resolve organizational trust + permissions(11x lesson)
- Agent + vertically fine-tuned model(HR generalist + vertical model paradigm)
- Cross-track commonality: Old SaaS → New Agent
11. Reference list
- Shensi Circle author complete works portal —— Complete works portal
- Main sources:
sources/learning/深思圈/Including ~66+ pieces covering "legal" / "healthcare" / "insurance" / "HR" / "recruitment" / "sales" / "customer service" / "finance" etc. - AI Investment and Valuation (cross-author topic synthesis) —— Synthesized VC perspective (this wiki supplements vertical industry data)
- Demand Analysis / JTBD / Product Core (cross-author topic synthesis) —— Demand analysis (methodology for selecting vertical tracks)
- Shensi Circle / AI Agent Company Case Library —— Agent company cases (vertical Agent details)
- Shensi Circle / AI Primary Market VC Perspective (a16z + Sequoia + YC) —— VC investment perspective
Sources and related materials
2026-10-08 Addendum: Hardware R&D Agent, first turn change impact into traceable data
Compiled based on Shensi Circle's 10/7 full interpretation of Flow Engineering. Sources cite company materials, investor conversations, and secondary research; funding, valuations, customer seats, efficiency, and compliance claims have not been independently verified this time. Working mechanisms and acceptance questions are retained below; self-reported company figures are not treated as actual returns.
The entry point for vertical products can be cross-tool coordination
A single parameter change in hardware R&D can simultaneously affect requirements, design, simulation, testing, and certification. Each specialized tool records only a portion of this information, yet teams need to know whether these portions remain consistent. The systems graph described in the Flow case explicitly connects the dependencies between requirements, interfaces, budgets, tests, and evidence, enabling changes to be traced along these relationships.
This differs from the product approach of simply adding a chat box to every tool. Its assessable value lies in reducing the time spent manually identifying impact scope and lowering the risk of missing associated changes. If the relationships themselves are not recorded or have become outdated, no matter how adept an agent is at generating text, it cannot guarantee reliable engineering judgments on that basis.
Propose changes first in an independent version, then decide whether to incorporate them into the formal plan
The demonstration in the article shows how a change in motor torque drives adjustments to the housing and downstream requirements. It can inspire a review workflow: propose a requirement change → list affected objects → generate candidate modifications and rationales → review by specialists → merge into the formal version. The few-second response time in the demonstration does not mean all designs have passed physical validation.
| What to check | Why it is needed | Deliverables that can be required |
|---|---|---|
| Whether dependencies are complete and materials are of the same version | Missing relationships can lead to mistaking local correctness for overall correctness | Changed objects, affected objects, and uncovered scope |
| Whether recommendations are traceable | Engineers need to understand why modifications are made, rather than only receiving new files | Original values, recommended values, justifications, and the input version used at generation time |
| Whether changes have undergone necessary checks | Being able to generate CAD files does not mean strength, thermal performance, or safety compliance are qualified | Record the status of simulation, testing, and manual review separately |
| Whether the system remains usable after tool integration changes | Changes to CAD, simulation, or data interfaces can invalidate the relationship graph | Synchronization failure alerts, deactivation conditions, and recovery verification |
The above are pilot requirements distilled from the case, and should not be treated as a product checklist confirming that Flow already possesses all these capabilities.
Measure adoption scale and engineering outcomes separately
Increases in user seats and call volume can indicate expanded usage scope, but they cannot independently demonstrate shorter design cycles, more thorough testing, or fewer errors. When evaluating such tools, first define a specific type of real-world change, and record the time required to find all impacted items, the number of omissions, manual review time, and final rework. The difficulty and acceptance criteria of the before-and-after samples need to be comparable for the results to be explanatory.
Cost is also not just a subscription quote. Organizing relationship data, maintaining interfaces, handling failures, and coordinating with team reviews all require investment. If time saved on generating files is offset by additional verification and coordination work, the overall benefit may be limited.
Human responsibilities focus on specifications, constraints, and accountability
As parts of execution become easier, teams need even clearer definitions of “what must be satisfied, what trade-offs are allowed, and who accepts the risks.” A practical starting point is to ask: after changing one requirement, how long does it take to find out which objects it affects? If teams always rely on ad hoc meetings to fill in gaps, prioritizing the organization of dependencies and responsibilities may be more effective than adding generation capabilities.
It is suitable to start pilots with processes that have clear scope and allow for controlled comparison. Physical testing, professional sign-off, and certification cannot be omitted just because agent outputs are faster; deployment and compliance qualifications in fields such as defense require separate verification. The methodological value of this case lies in making the coordination process inspectable, not in proving that a particular company’s valuation is reasonable or that its product has delivered on all its promises.
Source:Reference articles for this piece.
October 8, 2026 addition: make hardware engineering change impacts traceable
Based on Shensiquan's October 7 analysis of Flow Engineering. It cites company materials, investor conversations and secondary research. Funding, valuation, customer seats, efficiency and compliance claims were not independently verified here. This addition retains mechanisms and acceptance questions, without treating company-reported numbers as demonstrated benefits.
A vertical product can start with coordination across tools
A hardware parameter change may affect requirements, design, simulation, testing and certification. Each specialist tool records only part of that picture, while the team needs to know whether the parts remain consistent. The described systems graph connects dependencies among requirements, interfaces, budgets, tests and evidence so that changes can be traced through them.
The potential value is reduced effort finding affected components and a lower risk of missed dependencies. If relationships are missing or outdated, fluent generated text cannot guarantee sound engineering judgment.
Propose changes in a separate version before accepting them
The article demonstrates a motor torque change affecting the housing and downstream requirements. It suggests a review process: propose a requirement change, identify affected objects, generate candidate changes with reasons, obtain specialist review and incorporate accepted changes into the official version. A response within seconds does not mean every design has passed physical validation.
| Check | Reason | Evidence to request |
|---|---|---|
| Dependency completeness and consistent input versions | Missing relationships can make a locally correct result appear globally correct | Changed objects, affected objects and uncovered scope |
| Traceability of suggestions | Engineers need reasons, not just new files | Original value, proposed value, supporting evidence and input versions |
| Required verification | A generated CAD file does not establish strength, thermal or safety performance | Separate simulation, test and human-review states |
| Reliability after integrations change | CAD, simulation and data-interface changes can invalidate the graph | Sync failure alerts, stop conditions and recovery validation |
These are pilot requirements derived from the case, not a list of capabilities already confirmed in Flow.
Measure adoption separately from engineering results
More seats or calls can indicate wider use but do not independently establish shorter design cycles, better testing or fewer errors. Define a real class of changes and measure the time needed to find all affected items, omissions, human review effort and eventual rework. Before-and-after samples should have comparable difficulty and acceptance conditions.
Cost includes organizing dependency data, maintaining integrations, handling failures and supporting team reviews. Saving file-generation time may provide limited overall benefit if it creates more verification and coordination work.
Focus human responsibility on specifications, constraints and accountability
As execution becomes easier, teams need clearer agreement on requirements, acceptable trade-offs and who may accept risk. Ask how long it takes to identify everything affected by one requirement change. If repeated ad hoc meetings fill that gap, documenting dependencies and responsibilities may help more than additional generation capability.
Begin with a bounded process that supports comparison. Faster agent output does not remove physical testing, professional sign-off or certification. Deployment and compliance eligibility in areas such as defense require separate verification. The useful method is making coordination inspectable, not proving a valuation or every product promise.
Source: Article used for this addition.