See pricing
The Robots Already Working in American Factories | Tutor Intelligence
The a16z ShowRobotics

The Robots Already Working in American Factories | Tutor Intelligence

This episode is a lesson in why robotics is not just an AI problem. Josh Gruenstein explains how Tutor uses paid deployments, human fallback, and fast full-stack iteration to turn factory robots into a compounding data and economics flywheel.

November 2, 2026 · 46 min listen · 13 min read · Josh Gruenstein, Martin Casado, Raghu Raghuram
0:00
–:––

Context

Josh Gruenstein, co-founder and CEO of Tutor Intelligence, explains how Tutor is trying to make factory and warehouse robotics commercially real. Tutor has robots deployed across the United States, and Josh says those robots have helped manufacture and ship over a quarter of a billion products. The company's core idea is to put robots into real customer workflows, use human tutors when the robots need help, and turn that work into training data that improves the robots over time. For PMs, the episode is a strong case study in building for a greenfield market where the product is inseparable from operations, manufacturing, and go-to-market.

The Big Idea

Robotics will not be won by model quality alone. It will be won by the company that can connect real deployment, human fallback, and learning into one economic system.

Tutor's thesis is that factory robotics only starts compounding once robots are doing paid work in the field. That is what creates the right data, proves customer value, and justifies owning more of the stack.

Key Insights

Automation gap is enormous

The public image of manufacturing is far more automated than the real thing. Josh describes million-square-foot facilities where people still stand at conveyor belts and folding tables moving cases, packing boxes, and doing light assembly by hand. He also points to labor conditions that make the problem urgent: some of this work is backbreaking, hard to hire for, has about three months of turnover, and is often done by workers with an average age around 55. Josh also cites about 38,000 industrial robot deployments in the United States in 2025 versus roughly 1 million in China, which is why he sees much of the U.S. market as still greenfield.

  • Tutor's main bet is that the best training data comes from robots doing work customers already pay for.
  • Josh's argument is economic, not just technical. If a customer is paying for a task to be done, the data from that task is already aligned with real market value.
  • He contrasts that with generic data collection efforts, which may gather motion data but have no guarantee that the data maps to work anybody will buy.
  • Tutor supplements customer deployments with Data Factory 1, a Boston fleet of 100 Sunny robots used to push real-world reinforcement learning at a scale most labs do not have.

Spatial reasoning changed robotics

Josh calls recent advances in spatial reasoning a sea change. Frontier labs are post-training models on 3D data such as CAD files and 3D modeling assets because there is a real market for those capabilities. Tutor benefits because the same spatial reasoning helps connect general world knowledge to robotic action. The important nuance is that Josh does not see better models as reducing Tutor's value. He sees them as accelerating a system whose durable advantage is still the flow of real deployment data.

Robotics is full-stack by necessity

Tutor started from the view that even perfect remote human control would not solve robotics by itself. A company would still need to manufacture robots, build a supply chain, sign customers, deploy the robots, and support them in production. Josh says that if you refuse to touch one layer of the stack, you give up most of the trade-off space that would let you make progress. For PMs, the deeper lesson is that in immature markets, product design and business design are often the same problem.

Sell labor, not robotics

  • Most Tutor customers have never bought a robot before, so leading with model architecture is a poor sales motion.
  • Tutor opens with a simple offer: what if a robot could do the job for about $16 an hour.
  • That framing lets customers evaluate the offer against labor cost, not against science fiction.
  • Josh says the company averages about 40 to 60 days from first call to signed contract, about 40 days from contract to robots on site, and often gets robots doing production work within the first day.

Diverse customers force generality

Tutor sells to small businesses, mid-market companies, and large enterprises, including Fortune 50 and Fortune 100 firms. Josh argues that this diversity is not only a go-to-market choice. It is a product discipline. If a robot can move from one customer type to another without custom engineering, the team is probably building a real platform. He contrasts that with a common robotics failure mode: one large customer funds a bespoke build, the startup spends years on it, and the product never becomes general enough to scale.

Hardware must iterate like software

Tutor tries to compress physical product iteration until it looks more like software. Josh says the company ships code to robots daily, ships major releases every two weeks, and pushes manufacturing changes on a two-week cadence as well. He also says the product function sits right next to customer success and implementation, operating as one large team, while research stays close to real field problems. That org design matters because in robotics, deployment feedback is not a side channel. It is the main input to the roadmap.

Mental Models & Frameworks

Work divided by cost

Josh's value equation is simple: the value of a robot equals how much work it can do divided by its total cost.

  • Total cost includes upfront hardware, ongoing inference, teleoperation, maintenance, and break-fix work.
  • Josh notes that ongoing cost matters more than many outsiders expect.
  • Use this lens when a team is tempted to evaluate automation only as a one-time CapEx ROI project. In Tutor's world, the real competition is the current labor market and the full cost to match its output.

Deploy, tutor, learn loop

Tutor's operating model is a reusable product flywheel:

  • Deploy: put robots into real customer workflows quickly.
  • Tutor: let humans step in only when the model cannot yet complete the task.
  • Learn: convert those interventions into training data and model improvements.
  • Redeploy: expand capability, lower cost, and place more robots into more sites.

The key idea is that human fallback is not separate from the product. Human fallback is part of the learning system that makes the product better.

Specialize, generalize, respecialize

Josh describes robotics market maturity in stages:

  • Early markets start with narrow, specialized machines because that is all the market can support.
  • As the market matures, more general embodiments become viable. He sees humanoid-like form factors as one version of this general layer.
  • Once the market is large enough, specialization can return on top of a broader platform, much like compute evolved from general CPUs to a mix that includes GPUs and ASICs.

Use this model when deciding whether a new category is ready for a horizontal platform or still needs vertical solutions.

Full-stack trade-off surface

Josh's core systems view is that progress comes from making trade-offs across every layer at once: models, hardware, manufacturing, supply chain, product, and customers.

  • If a team refuses to touch one layer, such as hardware design or go-to-market, it loses a large share of the options that could unlock performance or economics.
  • If a team owns more layers, it can co-design them. Cassie and Sunny are examples of hardware shaped around the work Tutor wanted to automate and the models Tutor believed it could train.

Use this model when a product category looks stuck. The bottleneck may not be in the layer everyone is arguing about.

Decision Principles

Principle: Lead with economics

  • When: a product is new enough that buyers cannot easily judge the underlying technology.
  • Why: first-time robot buyers can compare labor cost, risk, and speed to value much faster than they can compare model architecture. Tutor got meetings by selling labor economics, not by explaining robotics research.

Principle: Use heterogeneity as test

  • When: picking early customers for a platform that aspires to generality.
  • Why: a product that works across small manufacturers and large enterprises without custom engineering is learning the right lessons. A product that works only for one flagship account may be learning the wrong ones.

Principle: Verticalize until interfaces mature

  • When: suppliers are too immature, too small, or optimized for an older version of the problem.
  • Why: Tutor argues that many robotics components were built for traditional industrial automation, not for generally capable robots. Owning more of the stack lets the company move faster until stable abstractions and reliable suppliers exist.

Trade-offs & Nuance

Tutors versus pure autonomy

Josh does not frame human tutors as a philosophical defeat. He treats them as an economic variable.

  • Fewer tutors means lower ongoing cost and less operational complexity.
  • More tutor coverage means broader task coverage sooner and less engineering time spent chasing the last bit of autonomy.
  • The right choice depends on the cost of human intelligence, the cost of model inference, and how much revenue wider task coverage unlocks.

Generality versus peak optimization

Cassie and Sunny are designed to cover broad swaths of manufacturing and logistics. That generality helps Tutor scale faster and gather more reusable data. The trade-off is that a general robot will not be the best possible machine for every environment. Josh is explicit that the current robots do not handle everything, especially outdoor work and home settings.

Factories before homes

Josh says Tutor will eventually go anywhere, but factories are the better starting point.

  • Factory work has clearer economics because the robot replaces costly, hard-to-hire, high-turnover labor.
  • Factory environments are more controlled than homes.
  • Home robotics demands both harder technical performance and a much lower price point before the economics work.

Integration versus ecosystem leverage

Tutor still buys some components from OEMs, but Josh expects the company to keep verticalizing as volume rises.

  • Buying components uses existing scale and keeps the company lighter.
  • Building more in-house gives Tutor more control over reliability, iteration speed, and domestic supply capacity.
  • The decision changes as Tutor becomes a meaningful share of supplier demand, because at that point some suppliers start behaving more like contract manufacturers than like broad-market platforms.

Common Mistakes

Mistake: Starting with the model

Josh says most robot pitches start with an insight about model architecture or the shape of the data. Very few start with a clear view of what economically valuable work the robot should actually do. That reversal matters because even a technically impressive system can fail if it is not tied to a painful, paid job.

Mistake: Building custom science projects

A recurring robotics failure mode is to sign one large customer, promise a highly specific solution, and then spend years building something that only works for that account. Josh argues that real product work needs iteration across many deployments. Without that, the company can mistake a services business for a platform business.

Mistake: Treating deployment as handoff

Tutor's org design assumes deployment is where the learning starts, not where product work ends. When implementation, customer success, research, and product sit too far apart, field problems arrive as anecdotes instead of structured insight. In robotics, that delay slows both the roadmap and the training loop.

Practical Application

Define the job first

Before approving an automation or AI initiative, write down the exact job being replaced or assisted. Include the environment, failure modes, throughput target, and current human cost. Tutor's central lesson is that a technology thesis without a sharply defined job is usually just an expensive demo.

Turn escalations into data

Audit the last month of human interventions in your product, such as manual QA, support overrides, implementation fixes, or operator takeovers. Tag what triggered the intervention and what the human changed. Tutor turned human tutoring into a data asset. Most product teams still treat similar interventions as invisible labor.

Price against current labor

If a product category is new, translate value into the buyer's current baseline rather than your feature list. That baseline might be hourly labor, overtime reduction, or throughput improvement. Tutor led with roughly $16 an hour robotic labor because that is easier for a first-time buyer to evaluate than robot intelligence.

Join product with implementation

Set up a shared weekly review for PMs, implementation leads, and customer success managers. Look at time to first value, manual workarounds, failed handoffs, and requests that repeated across accounts. Tutor's org design suggests that early product truth often lives in deployment friction, not in strategy decks.

Choose varied early customers

Deliberately mix early customers by size, complexity, and workflow. A varied customer set exposes whether a product is genuinely reusable or only survives through custom work. Tutor used customer diversity as a forcing function for building a more general platform.

Questions to Consider

  • Where is your team starting with a technology thesis or feature idea before writing down the exact job to be done, the environment, and the cost of the current manual workflow?
  • What human fallback step in your product, such as support intervention, manual QA, or assisted onboarding, could be logged and reused as training data instead of treated as invisible labor?
  • Are your early lighthouse customers helping you discover a general product, or are they pulling the team into bespoke work that will not travel to the next account?
  • If a first-time buyer ignored your architecture and judged only speed to value, ongoing cost, and implementation risk, would your product still look compelling?
  • Where would tighter contact between product management and implementation teams reveal constraints that are currently showing up too late in the roadmap?

Bottom Line

Tutor's argument is that robots become commercially real only when the product, the learning loop, and the business model reinforce each other. Paid deployments create the right data, human tutors cover the gaps, and full-stack control lets the company improve fast enough for robotics to compete directly with labor in real factories.

Resources Mentioned

ResourceTypeWhy it was mentioned
Josh Gruenstein on XSocial profileThe feed description points listeners here to follow Tutor Intelligence and Josh's thinking on robotics.
Martin Casado on XSocial profileThe feed description links to Martin, one of the episode's interviewers and a16z general partners.
Raghu Raghuram on XSocial profileThe feed description links to Raghu, one of the episode's interviewers and a16z general partners.
Tutor IntelligenceWebsiteThe company website is the main reference point for Tutor's robotics platform and commercial offering.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
Tutor IntelligenceBuilds, deploys, and operates robots for factories and warehousesThe company discussed throughout the episode and the source of the deployment, data, and business model lessons.
CassieHeavy-duty mobile robot with one arm and multiple suction cupsJosh describes Cassie as Tutor's workhorse for pallet, case, and other physically demanding logistics tasks.
SunnyIndustrial semi-humanoid robotJosh says Tutor has a 100-robot Sunny fleet in Boston for data collection and a growing fleet in the field for more dexterous tasks.

People to Follow

Josh Gruenstein

Josh Gruenstein is the co-founder and CEO of Tutor Intelligence. He is relevant because he is building at the intersection of robot learning, manufacturing, and go-to-market, not just talking about robotics as a research problem.

Martin Casado

Martin Casado is an a16z general partner and one of the interviewers in the episode. He is useful to follow for thinking that connects deep infrastructure shifts to product and market design.

Raghu Raghuram

Raghu Raghuram is an a16z general partner and one of the interviewers in the episode. He helps push the conversation toward enterprise adoption, supply chain structure, and how new infrastructure becomes usable by real businesses.

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free