Vendor Selection for Humanoids: The 10 Questions You Must Ask

When Figure AI's Figure 02 units completed their run at BMW's Spartanburg plant in early 2025, the automotive press covered it as a humanoid milestone. What got less coverage was the due diligence BMW's procurement and engineering teams had conducted before the first robot arrived: site visits to Figure's engineering facility, detailed review of the safety certification status, a defined pilot scope that BMW controlled, and contractual terms that protected BMW's operational data. BMW didn't just say yes to a vendor pitch. They ran a procurement process appropriate for bringing an experimental technology onto an active production line.

Most organizations evaluating humanoid vendors in 2026 are not running BMW-caliber procurement processes. They're evaluating based on demos, spec sheets, and press coverage — all of which are controlled by vendors and optimized to show the technology at its best. The 10 questions below are designed to surface what the vendor controls don't reveal.

Question 1: How many deployments have operated continuously for 90+ days without vendor personnel on site?

This is the single most diagnostic question in humanoid vendor evaluation, and it is almost never asked in initial sales conversations.

"Deployment" in vendor usage covers a wide range of commercial relationships — from paid production contracts to unpaid POC loans with friendly customers. A deployment that has been running for 90 continuous days without vendor support on site is the minimum evidence of production readiness. Below that threshold, the vendor cannot tell you what the robot's operational profile looks like when it is not being actively managed by people who built it.

Ask for a specific number. Ask for reference contact names at those deployments. If the vendor cannot provide a single reference who will take a call from you and answer specific operational questions without the vendor on the line, the platform is not production-ready.

What the answer reveals: Vendors with genuine production deployments can answer this immediately and specifically. Vendors in demo or early pilot phase will deflect to "we have multiple customer engagements" and offer to put you in touch with a customer success manager (who is a vendor employee), not the customer.

Question 2: What is your current software autonomy percentage in production deployments?

Autonomy percentage is the fraction of robot tasks executed without teleop intervention in a production environment. It is the most meaningful indicator of where the platform sits on the lab-to-production continuum, and it is almost never disclosed in vendor materials.

Ask specifically: "In your current production deployments — not pilots, not demos — what percentage of task cycles complete without requiring human teleop intervention?" Then ask: "What was that number 90 days ago, and what is your trajectory?"

A vendor with genuine production experience will have this number readily available and will be comfortable discussing the improvement trajectory. A vendor who is primarily in demo mode will not have this number, will give you a target rather than an actual, or will conflate it with a different metric (like "successful demo completion rate").

Acceptable answer for 2026: 75%+ autonomy in a constrained, structured task (tote handling, defined-path logistics) is achievable and has been demonstrated. Sub-60% autonomy means significant ongoing teleop costs that should be explicitly modeled in your TCO.

Question 3: What is your SLA for on-site support, and where is your nearest field technician?

Robots break. Software updates misfire. Obstacle maps need recalibration after facility layout changes. The question is not whether you will need on-site support during your deployment — you will — but whether the vendor can provide it at a speed and cost that works for your operation.

A vendor whose nearest field technician is a 6-hour flight away is not a vendor you can rely on for production operations. A hardware failure that takes 72 hours to address on a production line is not a maintenance event — it is an operational crisis.

Ask for the SLA in writing: maximum response time for critical failures (robot down, cannot operate), maximum response time for non-critical issues, and the financial remedy if the SLA is breached. Then confirm the nearest service technician location against public information — not the vendor's claim.

Red flag: An SLA that covers remote support but has no binding commitment for on-site response. Remote support can resolve software issues but cannot replace a failed actuator or a broken sensor assembly.

Question 4: Who owns the training data generated in my deployment?

This question has significant long-term strategic implications and is rarely addressed in early-stage vendor conversations.

Humanoid robots learn from operational experience. The autonomy improvements the robot makes while operating in your facility are driven by data generated in your environment — your specific floor layout, your specific product dimensions, your specific obstacle patterns. That data has value.

In many humanoid vendor contracts in 2026, the default position is that the vendor retains rights to all data generated on their hardware, including the autonomy training data produced during your deployment. This means if you decide to switch vendors or terminate the contract, you cannot take the training data with you. A competitor evaluating the same platform after you will benefit from improvements your deployment funded.

Ask directly: "Who owns the data generated during our deployment? If we terminate the contract, can we export and retain the training data from our environment?" Get the answer in writing.

Question 5: What happens to the contract if you raise your next funding round — or if you don't?

The humanoid vendor landscape in 2026 is venture-funded, early-stage, and consolidating. Several platforms that are currently viable commercial options will be acquired, pivoted, or shut down within 36 months. Several more will raise new capital that comes with investor pressure to change the commercial model — moving from hardware sales to RaaS-only, changing pricing tiers, or pivoting to a different vertical.

Buyers who sign multi-year pilot or deployment agreements with humanoid vendors need to understand what protections they have if the vendor's circumstances change materially.

Ask: "What protections does our contract include if you are acquired by a competitor?" (Your data, your integration, your operational continuity.) "If you pivot your commercial model — say, from purchase to RaaS-only — what is our position under the existing agreement?" "If operations are suspended due to funding issues, what is the contract's wind-down provision?"

These questions are uncomfortable to raise in a sales conversation. Vendors who are thinking seriously about long-term customer relationships will engage with them. Vendors who deflect or treat them as offensive are telling you something important about how they think about the customer relationship.

Question 6: What is your ISO functional safety certification status?

For humanoid robots operating near human workers in manufacturing or logistics environments, functional safety certification is not optional — it's a prerequisite for regulatory compliance and insurance in most jurisdictions.

Agility Robotics is targeting ISO functional safety certification for Digit in mid-to-late 2026. That timeline means Digit deployments launched before that date are operating under a preliminary safety regime, with full certification pending. This is a real consideration for regulated industries (pharmaceutical, food processing, automotive with specific OEM safety standards).

Ask: "What is your current functional safety certification status?" "Are you certified under ISO 10218, ISO/TS 15066, or equivalent?" "If you are not yet certified, what is the certification timeline and what interim safety protocol governs our deployment?"

The answer to the interim safety protocol question is the one that actually matters for your risk management: what risk assessment has been conducted for your specific deployment environment, and who conducted it — the vendor, an independent assessor, or nobody?

Question 7: What does the software update protocol look like, and who controls the update schedule?

Humanoid robots receive frequent software updates — autonomy model improvements, sensor calibration updates, task policy changes. Some of these updates are beneficial and improve performance. Some introduce regressions that affect task success rates or behavioral profiles.

Ask the vendor: "Who controls the update schedule — you or us?" "Can we delay or roll back a software update if it affects our production operations?" "What is your protocol for notifying us before a major model update, and what is the rollback path if the update degrades performance?"

In 2026, many humanoid vendors push updates without customer notification because the robots are cloud-connected and the vendor controls the deployment pipeline. This is fine for consumer software; it is not acceptable for production automation where an unexpected behavioral change can cause operational disruption, safety incidents, or line stoppages.

Minimum acceptable answer: 48-hour advance notification before any significant model update, with a documented rollback procedure that the vendor will execute on your request within 24 hours of a reported regression.

Question 8: What is your teleop model — who provides operators, what is the expected ratio, and what is the cost structure?

Teleop is the operational dependency that humanoid vendors most consistently underemphasize in sales conversations. Every humanoid in production deployment in 2026 requires teleop coverage for exception handling — the question is who provides it and what it costs.

Some vendors provide teleop operators as part of their service package (typically bundled into a RaaS fee). Others require the customer to designate and train in-house teleop operators. Others use a hybrid model with remote vendor support for complex exceptions and on-site staff for routine interventions.

Ask: "Who provides teleop coverage during our deployment?" "What is the expected teleop ratio — teleop hours per robot operating hours — at deployment start, and what is your contractual commitment for the ratio at 90 days and 180 days?" "If the ratio does not improve on the committed schedule, what is the remedy?"

The answers reveal how seriously the vendor has thought about the operational reality of their platform versus the demo-optimized narrative.

Question 9: Can we speak directly with operations contacts at two of your current production deployments?

Not the vendor's customer success team. Not a joint call that the vendor joins. A direct conversation with the operations manager or plant director at a facility that has been running the platform in production for at least 90 days.

The questions to ask that reference contact:

• "What was the biggest implementation problem you encountered?"
• "What does your teleop coverage look like today versus at launch?"
• "Has the vendor delivered on their support SLAs? Give me a specific example of a failure and how they responded."
• "What would you do differently if you were starting this deployment today?"
• "Would you recommend this vendor to a peer at a similar facility?"

Reference contacts arranged through the vendor have been pre-briefed and selected for positive feedback. Ask for the reference contact list and then independently verify that the named person is actually an employee at the named facility (LinkedIn, the facility's public directory) before taking the call. If the vendor resists providing direct reference contacts without vendor presence, that resistance is itself informative.

Question 10: What does your hardware refresh look like during a 3-year contract period?

The iteration velocity in humanoid hardware is unlike anything in traditional industrial automation. Figure 02 was replaced by Figure 03 in approximately 18 months. Boston Dynamics moved from research Atlas to production Atlas across a major architectural revision. Buyers signing 3-year agreements should expect at least one significant platform generation change during the contract period.

Ask: "If you release a new hardware generation during our contract period, what is our upgrade path?" "Is the upgrade at additional cost, at cost, or included in our service agreement?" "If the current platform is no longer actively supported 18 months from now, what are our options?"

For RaaS customers, hardware refresh terms are particularly important: the RaaS model's primary advantage is flexibility, and a RaaS contract that locks you into a deprecated platform generation without a cost-effective upgrade path eliminates that advantage.

The Red Flag Summary

Any vendor engagement that produces the following responses warrants significant caution:

• Cannot provide reference contacts at 90+ day production deployments (not pilots)
• Will not disclose software autonomy percentage in production
• No binding on-site support SLA
• Customer-unfavorable data ownership default with no negotiation position
• No advance notification or rollback capability for software updates
• Teleop model is undefined or entirely the customer's responsibility with no ratio commitments
• Deflects the hardware refresh question to "we'll cross that bridge when we come to it"

The humanoid market in 2026 has real, commercially deployed platforms — Digit, Figure, and Apollo are all demonstrably operating in industrial environments. It also has a significant number of vendors who are in demo and early-pilot phase presenting as commercial-ready. The 10 questions above are designed to make that distinction clear before you commit capital, staff time, and operational disruption to a deployment.

The vendors who are genuinely ready for production deployments can answer all 10 questions specifically, in writing, with reference contacts who will take your call. That's the standard. Everything else is a demo.