AI agents are confined to the digital realm. The agentic AI platform market is $10.75B in 2025, projected to $51.26B by 2030 (36.7% CAGR).1 But agents lack two things to act in the physical world: (1) secure payment rails and (2) trusted physical actuators — the ability to order a service and verify it was completed.
Nobody has solved the actuation and verification layer as a horizontal platform — and there is a graveyard of companies that tried.
The question is not whether agents will need to trigger physical actions. The question is how — horizontal API or vertical-first with verification.
| Layer | Size | Source | Notes |
|---|---|---|---|
| Agentic AI Platforms | $10.75B → $51.26B (2030) | Mordor Intelligence1 | All agent use cases, not just physical |
| Field Service Management | $5.1B → $9.17B (2030) | MarketsAndMarkets2 | 12.5% CAGR. Pest control, HVAC, cleaning, inspections |
| B2B Corporate Gifting | Sendoso $84.8M rev | Latka3 | Proves software→physical API generates revenue |
| Agent Companies | 1,700+ globally | CB Insights4 | Only a fraction need physical-world APIs |
“If each step in an agent workflow has 95% reliability, 20 steps = 36% success rate. This isn’t a prompt engineering problem. This is mathematical reality.” — Utkarsh Kanwat5
“68% of production agents execute at most 10 steps before human intervention. Reliability is the #1 development challenge.” — Study of 306 practitioners6
“The dirty secret of every production agent system is that the AI is doing maybe 30% of the work. The other 70% is tool engineering.” — Kanwat5
“Physical world actions are the hardest. Every tool call that touches real things has a failure mode that LLMs can’t retry.” — HN discourse on agentic systems
| Company | What | Physical Fulfillment? |
|---|---|---|
| OpenAI ACP | Open-source protocol (w/ Stripe) for agent commerce7 | No — merchants handle |
| Stripe Agentic Suite | Single API for merchants to connect with AI agents8 | No — merchants handle |
| Google AP2 + A2A | Cryptographically-signed payment Mandates9 | No — protocol only |
| Zomato MCP | Production MCP for end-to-end food ordering10 | Yes — Zomato’s fleet |
| Company | Model | Revenue / Scale |
|---|---|---|
| Thumbtack × OpenAI | ChatGPT app for home services11 | 300K service pros |
| Sendoso | B2B gifting API + SaaS ($12K–67K/yr)3 | $84.8M rev, $157M raised |
| DoorDash Drive | Delivery-as-a-service API12 | $9.75 base (0–5mi), massive US coverage |
| Florist One API | REST API, 20% commission13 | 10K+ florists, $40M total sales since 2009 |
| Reachdesk | B2B gifting + direct mail API | Sendoso competitor, SaaS + per-item model |
| Instacart Connect | Grocery delivery API for brands | White-label fulfillment, restaurant/retail |
| Company | Model | Risk |
|---|---|---|
| LocalServiceAPI | API for local services | Same graveyard pattern — supplier coordination |
| HumanOps | Human-in-the-loop API | Human cost = margin erosion |
| Callio | AI + human concierge | Unproven at scale |
| Company | Funding | What |
|---|---|---|
| Spot AI | $93M14 | Video AI Agents for the physical world. Camera → proactive problem identification. |
| FYLD | $41M15 | AI for field ops risk + digital work execution. 30K-40K field workers. |
| TechSee | Established | AI visual verification for field services. Reduces manual QA by 95%+. |
| Deepomatic | €10M16 | Visual automation for telecom field workers. Task completion verification. |
| Zeitview | $60M17 | AI inspections for infrastructure. 200K+ assets across 80 countries. |
Bland AI ($65.5M)18 bypasses supplier onboarding entirely by making phone calls to existing booking systems. No API integration, no supplier dashboard. If your agent can just call the plumber, why does it need an API?
| Company | Raised | Why They Died |
|---|---|---|
| Facebook M | FB-funded | Universal request fulfillment (2015–18). Human contractors 24/7 = uneconomical.19 |
| Magic | $12M (Sequoia) | SMS concierge. Broke even on 50%, lost money on 25%. Pivoted to $100/hr.20 |
| GoButler | Significant | Universal text concierge. Pivoted to automated flights. Died.21 |
| Operator | Uber co-founder | Conversational commerce. Human coordination cost too high.22 |
| Saiga | €3M | “Fix the concierge model” with AI/RPA (2020). No news after 2022.19 |
The viability of any agent-to-physical platform hinges on one variable: exception handling rate. When an agent triggers a physical service and the provider ignores, drops, or botches it, a human must intervene. That intervention cost eats the margin.
| Scenario | Exception Rate | Avg Fee | Ops Cost / Exception | Gross Margin |
|---|---|---|---|---|
| Optimistic | 10% | $12 | $5 | 96% |
| Realistic | 30% | $10 | $8 | 76% |
| Pessimistic | 50% | $8 | $12 | 25% |
| Death | 60%+ | $8 | $15 | Negative |
| Component | % of Revenue | Who Captures | Note |
|---|---|---|---|
| Platform fee (markup) | 15–25% | API company | Only viable if exceptions <20% |
| Supplier execution | 50–70% | Provider | Pest control, cleaning, delivery |
| Human ops (exceptions) | 10–40% | Burned | Chase providers, resolve disputes |
| Payment rails | 2–3% | Stripe, etc. | Big Tech owns this layer |
At 50% exception rate, human ops exceed platform fee. Margin collapses.
Magic learned this: they broke even on 50% of orders and lost money on 25%.20 Sendoso and Reachdesk solve this by charging $12K–67K/yr in SaaS fees on top of per-item costs — the SaaS fee subsidizes the human coordination layer.3
The agent-to-physical orchestration space has a fatal gap: nobody can verify that the physical task was actually completed correctly. OpenAI ACP handles payment. Thumbtack handles matching. DoorDash Drive handles delivery. But none of them can confirm that the pest control technician actually treated every corner, or that the cleaner actually scrubbed the kitchen. FieldCheck can.
| Problem (Horizontal APIs) | FieldCheck’s Solution |
|---|---|
| Scraped supplier DBs — providers don’t respond | Opted-in providers using the FieldCheck app. Camera verification = accountability. |
| Exception rate >50% kills margin | Visual verification closes the loop. Agent can see the work was done via camera feed. |
| No quality signal — was the job done well? | Gemini/VLM analysis of captured images provides automated QA scoring. |
| Human ops needed to chase providers | Worker app with real-time guidance eliminates coordination overhead. |
Today, if an AI agent managing a building wants to order pest control, it hits a wall — it can find a provider (Thumbtack) and maybe pay (Stripe), but it cannot verify completion. With FieldCheck in the loop:
This is the closed-loop agent-to-physical pipeline that the horizontal API companies couldn’t build. FieldCheck already has steps 2–3 (worker capture + VLM verification). Steps 1 and 4 are API wrappers.
| Graveyard Pattern | Why FieldCheck Is Different |
|---|---|
| Horizontal = infinite categories, no expertise | Vertical-first (pest control, cleaning, inspections). Master one before expanding. |
| No supplier onboarding | Providers actively use the app for AI guidance. They’re opted-in. |
| No verification = no trust = no repeat | Camera verification IS the product. Trust is built into every job. |
| Human ops eat the margin | VLM replaces human QA. Exception rate pushed below 10%. |
| Facebook M: “learning rate slower than hoped” | FieldCheck doesn’t need general AI. Needs domain-specific VLM for specific tasks (pest traps, cleaning zones). |
YC Spring 2026 RFS explicitly calls for “AI Guidance for Physical Work” (David Lieb): “Imagine wearing a small camera while an AI sees what you see and talks you through the job. Instead of needing months or years of training, workers can become effective immediately.”23 This is exactly FieldCheck’s product.
| Phase | What | Agent Integration |
|---|---|---|
| Now | Manager capture + worker guidance for repair.sg (pest control) | None — human-triggered |
| Phase 2 | FieldCheck API: external systems can trigger jobs programmatically | Building management systems, property tech |
| Phase 3 | MCP Server: AI agents can directly order verified physical services | Any AI agent can order + verify pest control, cleaning, inspections |
STRONG for FieldCheck — FATAL for Horizontal APIs
The agent-to-physical orchestration space is real ($10.75B agentic AI, $5.1B field service management) and growing fast. But the startup graveyard is unambiguous: Facebook M, Magic, GoButler, Operator, and Saiga all tried horizontal physical concierge and died. The pattern is always human coordination cost exceeding markup.
FieldCheck inverts the graveyard pattern. Instead of building an API layer hoping suppliers will respond, FieldCheck starts with the hardest part — camera-verified physical task completion — and can later expose it as an API. The verification layer is the moat. Every horizontal API company died because they couldn’t answer one question: “Was the job actually done?” FieldCheck can.
Sequence: (1) Prove repair.sg pest control loop. (2) Generalize to cleaning + NinjaVan. (3) Expose FieldCheck API for building management / proptech. (4) Ship MCP server so any AI agent can order verified physical services. Don’t skip to step 4.