10 AI Startups to Watch: Agentic AI and Robotics (and What Each One Has Actually Proven)

By Tom Morgan · Research current as of October 27, 2026 · Every figure below carries a confidence label and a source link.

In November 2025, a humanoid robot company did something unusual: it showed its retired workers. Figure AI pulled its Figure 02 robots off a BMW assembly line in South Carolina after about eleven months, and its CEO posted photos of machines covered in scratches and grime. The company said the fleet had helped build more than 30,000 BMW X3s and loaded over 90,000 sheet-metal parts (CB Insights summary of the announcement).

That scuffed fleet is a better “startup to watch” signal than most funding headlines. It is evidence of work done in a real building, and it came with a documented cost: the robots wore out, and the next model had to fix what failed.

Most “startups to watch” lists skip that distinction. They rank companies by valuation and adjectives. This one asks a plainer question for each company: what has it actually proven, and what is still a promise?

The short answer: the 10 startups

Valuations are not like-for-like: some are post-money, one is a pre-money SPAC figure, some are press-reported. Details are in each profile.

First, two definitions (so we’re arguing about the same thing)

Agentic AI is software that doesn’t just answer a question: it plans, picks tools, and carries out multi-step work with limited human prompting. A chatbot tells you your refund policy. An agent checks the order, issues the refund, and updates the record.

Physical AI (embodied AI) is the same idea in the physical world: models that let robots perceive, decide, and act on tasks they weren’t hand-programmed for.

Why put them in one article? Because they fail differently, and the contrast is instructive. Software agents fail through wrong decisions and unclear ROI. Robots fail through wear, safety, and the brutal cost of scaling hardware. Both categories, though, are fighting the same battle: moving from impressive demo to repeatable deployment.

How to read this article

I only used claims I could trace to a source, and I labeled each one:

• High: company press release, regulatory filing, or Bloomberg/Reuters/WSJ-class reporting.
• Medium: reputable secondary reporting or a data aggregator that cites a primary source.
• Low: a single secondary source I couldn’t confirm. Treat as a lead, not a fact.

Where sources disagreed, I say so, and the full list of conflicts is in the correction ledger at the end.

The money picture at a glance

Latest reported valuations, in billions of US dollars (one block ≈ $2B). Not strictly comparable. See notes above.

text

Figure AI          ████████████████████  $39BCognition          █████████████         $26BSierra             ████████              $15.8BHarvey             ████████              $15.5BSkild AI           ███████               $14BPhys. Intelligence ██████                ~$11BApptronik          ███                   ~$5.5Bn8n                ███                   $5.2BDecagon            ██                    $4.5BAgility Robotics   █                     $2.5B (pre-money, SPAC)

A sobering companion chart: how many times annual revenue investors are paying, where a revenue figure exists. This is my own arithmetic from the numbers cited in the profiles, so treat it as a rough scale, not a precise ratio.

text

Skild AI   ██████████████████████████████████████████████  ~467x  ($14B / ~$30M)Sierra     ██████████                                       ~105x  ($15.8B / ~$150M)n8n        ██████                                           ~62x   ($2.5B / ~$40M, Oct 2025)Cognition  █████                                            ~53x   ($26B / ~$492M)

One block ≈ 10x. Skild’s figure is the “roughly $30M” revenue the company cited for 2025. n8n’s ARR is a third-party estimate.

Part 1: Agentic software

1. Sierra: the customer-service agent that grew up fast

Sierra was co-founded by Bret Taylor (the OpenAI board chair, formerly Salesforce co-CEO) and ex-Google executive Clay Bavor. In May 2026, it reportedly raised $950M at a $15.8B post-money valuation, led by Tiger Global and GV, per CNBC reporting carried by Quartz (High).

What it has shown: annual recurring revenue of about $150M after eight quarters. Customers named in that reporting include Prudential, Cigna, Blue Cross Blue Shield, and Rocket Mortgage (High). Taylor claims penetration of more than 40% of the Fortune 50 (Medium: company claim). A secondary outlet reports $200M ARR by May (Low: single secondary source). The path before that: $4.5B in October 2024 and $10B in September 2025 (Medium).

The honest caveat comes from Taylor himself: he told CNBC he expects a significant AI-sector correction before the decade’s midpoint (High). A company priced at roughly 105x revenue is the kind that feels any correction first.

Practical takeaway: Sierra is the cleanest example of agents earning money in a domain with clear unit economics (support tickets). If you’re evaluating any support agent, ask Sierra-style questions: what share of conversations resolve without a human, and what happens to the rest?

2. Cognition: from viral demo to real revenue (and real burn)

Cognition’s Devin launched in 2024 to a wave of skepticism over whether the demo was overhyped (CO/AI, Medium). The numbers since are hard to dismiss. Devin’s ARR reportedly went from about $1M in September 2024 to $73M in June 2025 (Sacra, Medium). The July 2025 Windsurf acquisition added an estimated $82M ARR and 350+ enterprise customers (Medium). Cognition raised $400M at $10.2B in September 2025 (Built In, High) and, in May 2026, $1B at $26B (Dealroom, Medium), with ARR around $492M (Medium).

Be careful with what came next. One crypto-news outlet reports ARR of about $900M in August 2026 and a plan to burn $800M in cash this year (Low: single, unverified source). I’d treat it as a signal worth checking against the company’s own statements, not a number to repeat.

Practical takeaway: Among the five agentic companies here, Cognition’s reported revenue growth is the steepest, though the figures are secondary-sourced. The lesson for buyers (my suggestion): start with narrow, reviewable tasks such as bug fixes or test writing, where a human can check the output in minutes.

3. Decagon: fast growth, customer-reported results

Founded in 2023 by Jesse Zhang and Ashwin Sreenivas, Decagon raised $250M at a $4.5B valuation in January 2026, led by Coatue and Index Ventures, according to Bloomberg (High). The same report says it added more than 100 corporate clients in 2025, including Avis Budget Group and Deutsche Telekom (High).

The most quoted outcome is Chime’s: a reported reduction of more than 60% in contact-center operating costs and a doubled Net Promoter Score (CMSWire, Medium: customer-reported, not independently audited). I did not find a disclosed ARR figure.

Practical takeaway: Compare Decagon with Sierra, and you see a pattern: in support, the evidence is mostly vendor- or customer-reported. Insist on a pilot with your own tickets and your own success metric.

4. n8n: the unglamorous layer that might matter most

n8n is a Berlin workflow-automation company founded in 2019 by Jan Oberhauser. After pivoting toward AI, it raised about $60M in March 2025 at a reported valuation near €250M (TechCrunch, High), then $180M at $2.5B in October 2025, led by Accel (Pulse 2.0, Medium). In May 2026, SAP reportedly made a strategic investment that doubled the valuation to $5.2B (Orchestra summary, Medium).

Its pitch is a useful counterweight to full autonomy: teams decide how much freedom the AI gets and where rules and human review sit. ARR of “$40M+” circulates (Low: third-party estimates).

Practical takeaway: Gartner’s warning about unclear value and weak risk controls (see “What goes wrong” below) suggests autonomy should be a dial, not a switch. n8n’s model, with some steps AI-decided, some rule-based and some human-approved, is one concrete way to build that. That is my reading, not a finding from any source.

5. Harvey: agents in the most error-intolerant profession

Harvey said in March 2026 that it raised $200M at an $11B valuation, co-led by GIC and Sequoia. The same announcement reports 25,000+ custom agents running on its platform, 1,300+ customers across 60+ countries, and a majority of the AmLaw 100 as partners (Harvey press release, High, company claims). In September 2026, it reportedly raised $550M at $15.5B (Unite.AI, Medium; another aggregator says $16B, see ledger). Sacra estimates ARR of about $300M in May 2026 (Sacra, Medium: estimate).

A nice detail for the skeptics: Harvey open-sourced a Legal Agent Benchmark in May 2026 so others can measure long-horizon legal agents (Medium).

Practical takeaway: Legal is the best stress test for agent reliability, because a confident wrong answer is expensive. Look for vendors who publish benchmarks, as Harvey now does, instead of only publishing customer logos.

Part 2: Robotics and physical AI

6. Physical Intelligence: the lab that hasn’t shipped, and is valued as it will

Founded in 2024 by AI academics and former Google DeepMind researchers, Physical Intelligence builds general-purpose AI models meant to run many different robots. It raised $600M at $5.6B in late 2025 (Medium) and was reported in March 2026 to be in talks for about $1B at more than $11B (Bloomberg Law, High for the talks; I could not confirm a closing from a primary source, though an aggregator lists a May 2026 close, Low).

Dealroom notes that the company has no stated timeline for commercialization and calls generalizing across physical tasks an unsolved research problem (Dealroom, Medium).

Practical takeaway: This is a bet on a research result, not a product you can buy. Watch its publications and the hardware makers that adopt its models, not press releases.

7. Skild AI: “one brain, any robot”

Skild AI, based in Pittsburgh, raised about $1.4B in January 2026 at a valuation above $14B, led by SoftBank with NVentures, Macquarie Capital, Bezos Expeditions and 1789 Capital (Dow Jones via MarketScreener, High). That more than tripled its roughly $4.5B valuation of about seven months earlier (High).

Its commercial claim is the interesting part: revenue going from zero to around $30M within a few months in 2025 (High: company-stated), with more than eight customers (Arise, Medium) and no client names disclosed (Medium). Its model, the Skild Brain, is designed to control different robot bodies without knowing their exact shape (The Bridge, Medium).

Practical takeaway: Revenue is a far stronger signal than a demo, but anonymous customers mean you can’t check it. If you are a buyer, ask for a reference call.

8. Figure AI: the best-documented deployment, and the biggest gap

Figure raised $675M at $2.6B in February 2024 (press release, High) and announced more than $1B at a $39B post-money valuation in September 2025 (Reuters, cited by ThomasNet, High).

The proof is real and measurable. BMW said Figure 02 supported production of more than 30,000 X3s over about ten months at its Spartanburg plant, and that Figure 03 would move next into a logistics sequencing project at the same site (RuntimeWire, Medium). One industry profile reports 1,250+ runtime hours, 90,000+ parts loaded, and a failure source Figure 03 was designed to fix: forearm electronics and cabling (Robotics.press, Medium).

Now the counterweights. In spring 2025, Fortune questioned how much work Figure’s robots were really doing at BMW, which Figure’s CEO disputed; I could only read a secondary summary of that exchange (Low). The same industry profile says BMW is piloting a different company’s wheeled humanoid in Germany, which suggests automakers are running parallel trials, not picking winners (Medium). And I found no disclosed revenue to set against a $39B valuation.

Practical takeaway: One factory, one carmaker, eleven months of public numbers. It’s a strong pilot, not yet a repeatable business. Ask how a pilot becomes a multi-site contract.

9. Agility Robotics: the only one with the paperwork

Agility is the outlier because it has been forced to put numbers into securities filings. On June 24, 2026, it announced a merger with the SPAC Churchill Capital Corp XI at a $2.5B pre-money equity value (SEC exhibit, High). The filings and press release state that its Digit robot is commercially deployed with Schaeffler, GXO, Toyota Motor Manufacturing Canada, and Mercado Libre, with 65,000+ operating hours across nine customer facilities, and more than $300M of multi-year Digit v5 orders (High: company statements in filings). Gross proceeds could exceed $620M, including $420M held in the SPAC trust assuming no redemptions (Verdict, Medium). The SEC declared the registration statement effective on October 13, 2026 (Unite.AI, Medium). I did not find confirmation that the deal has closed.

Read the fine print: the $300M+ in orders is described as subject to achieving certain contractual milestones (High). And the company itself says humanoid deployment today still requires robots and people to work in segregated areas (High).

Practical takeaway: Orders contingent on milestones, a SPAC that depends on low redemptions, and robots fenced off from people: that’s what “early commercial” really looks like. It’s also more disclosure than any private rival offers.

10. Apptronik: heavy backing, light disclosure

Austin-based Apptronik, which grew out of a University of Texas robotics lab, announced a $520M Series A-X extension in February 2026, bringing its total Series A to more than $935M and total capital raised to nearly $1B. Investors include Google, B Capital, Mercedes-Benz, PEAK6, AT&T Ventures, John Deere, and the Qatar Investment Authority; partners include Mercedes-Benz, GXO, and Jabil (Apptronik press release, High). It also collaborates with Google DeepMind on next-generation robots (Medium). The company did not state a valuation in its release; press reports put it at roughly $5B to $5.5B (Dealroom, Medium).

I found announced partnerships, pilots, and a production-ramp plan, but no public figure for hours worked or units deployed.

Practical takeaway: A partnership announcement is not a deployment. When a vendor lists logos, ask which are pilots, which are paid, and how many robots are working today.

The “proof ladder”: who has shown what

Rungs run from weakest evidence (1) to strongest (5):

1. Demo or research paper
2. Funding and named investors
3. Announced partnership or pilot
4. Named customer with measured usage or outcomes
5. Audited or filed numbers (revenue, orders, hours)

What goes wrong (documented, not imagined)

Hype outruns reality. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. It also estimated that only about 130 of the thousands of “agentic” vendors are real, calling the rest “agent washing” (Gartner press release, High). The same release predicts that by 2028 at least 15% of day-to-day work decisions will be made autonomously by agentic AI, and 33% of enterprise software will include it.

Valuation outruns revenue. The multiples chart above is the visual version of this point. Figure’s and Physical Intelligence’s valuations rest on revenue that is undisclosed or not yet there.

Hardware wears out. Figure’s own retirement of the Figure 02 fleet, and its stated redesign to fix a failure point, is the most honest evidence in this article that physical deployment has costs software doesn’t.

Safety separation. Agility’s own filing says today’s humanoids still work apart from people. A robot that can’t share a floor with workers limits where it can be useful.

Concentration risk. Several of these companies share investors (SoftBank, Nvidia, Bezos Expeditions, Sequoia, and Google’s venture arms appear repeatedly). If sentiment turns, funding for the whole category could tighten at once. That’s an inference from the investor lists, not something any source states.

This section is my own synthesis of the evidence above, not a claim from any source.

A practical playbook: how to evaluate an agent or robot vendor

1. Climb the proof ladder. Ask which rung they’re on. A named customer with measured results beats any demo.
2. Separate pilot, contract and deployment. Ask how many units or agents are running in production today, and for how many hours.
3. Read the contingencies. Agility’s disclosure that orders hinge on milestones is the model: look for the same condition in any vendor’s “backlog.”
4. Design the pilot around one measurable task. Support resolution rate, bugs fixed per week, parts moved per shift. If you can’t measure it, don’t automate it yet.
5. Keep a human on the dial. n8n’s approach (choose where AI decides and where people approve) is a pattern that fits most real deployments.
6. Price in the downside. Ask what happens to your costs if the vendor is acquired, merges, or raises prices. Windsurf’s 2025 ownership change is a reminder that vendors move.
7. Check for benchmarks. Prefer vendors who publish evaluations (Harvey’s legal benchmark is an example) over those who publish only testimonials.

What to watch over the next 12 months

• Agility’s SPAC outcome and first earnings. The first public-market test of humanoid economics.
• Figure’s second site. Does BMW expand, or do parallel pilots go to rivals?
• Physical Intelligence’s first commercial signal. A product, a customer, a price.
• Cognition’s confirmed numbers. Whether revenue and burn reports hold up.
• Sierra, Decagon and Harvey’s competitive pressure. Rising competition in customer support and legal AI could squeeze growth and pricing.
• Gartner’s 2027 prediction. Whether the cancellation wave arrives on schedule.

Correction ledger

Where my sources disagreed, here’s what I chose and why.

Coverage limits: I did not find primary announcements for Cognition’s May 2026 round, Physical Intelligence’s Series C, n8n’s SAP investment, or Harvey’s September round. Those figures rely on secondary sources and are labeled accordingly. Funding markets move fast; re-check before relying on any number.

Sources

Primary and high-confidence

• Gartner: Over 40% of agentic AI projects will be canceled by end of 2027
• Agility Robotics / Churchill Capital Corp XI (SEC exhibit)
• Apptronik press release
• Harvey press release
• Figure Series B press release
• Bloomberg on Decagon
• Bloomberg Law on Physical Intelligence
• Dow Jones via MarketScreener on Skild AI
• TechCrunch on n8n
• Reuters on Figure’s $39B round (as cited by ThomasNet)

Secondary reporting and aggregators (medium/low confidence)

• Quartz on Sierra (CNBC-based)
• Dealroom on Cognition · Built In · Sacra
• CMSWire on Decagon
• Pulse 2.0 on n8n · Orchestra on n8n
• Unite.AI on Harvey · Sacra on Harvey
• Dealroom on Physical Intelligence
• The Bridge on Skild AI · Arise
• RuntimeWire on Figure and BMW · Robotics.press Figure profile · CB Insights BMW page
• Verdict on Agility · Unite.AI on Agility S-4
• Dealroom on Apptronik
• CO/AI on Cognition