SoftBank Vision Fund 2. Temasek. Salesforce Ventures. Mercedes-Benz. A NZ AI company that generated the most impressive demos in the industry, landed partnerships with the brands every founder wants on their reference list, and built zero sustainable revenue. A forensic teardown of the most expensive case study in NZ tech history of confusing demand signals with distribution readiness.
“All three marquee early customers — ANZ, Air New Zealand, and Mercedes-Benz — abandoned the technology. ANZ ceased using it in early 2022. Air NZ replaced its ‘Sophie’ avatar with an in-house-built ‘Oscar.’ Mercedes-Benz reduced its stake from 4.4% to 1.6%.”
— NZ Herald, February 2026
Founded in Auckland in 2016 by Greg Cross and Dr. Mark Sagar — an Oscar-winning visual effects scientist from Weta Digital — Soul Machines set out to build autonomous, emotionally responsive AI-powered digital humans. The pitch was genuinely extraordinary: photorealistic avatars that could hold conversations, read emotional cues, and serve as customer-facing interfaces for enterprise brands. Not a chatbot. Not an IVR menu. An entity that looked like a person, responded like a person, and could be deployed at scale without the cost, variability, or fatigue of human agents.
The technology was real. The demos were extraordinary — in a room, watching a Soul Machines avatar respond to questions with what appeared to be emotional intelligence, the product felt genuinely category-defining. Mercedes-Benz deployed it as an in-car assistant. ANZ Bank deployed it as a customer service agent. Air New Zealand deployed Sophie, a digital crew member who could answer questions across the carrier’s website. The company raised US$135 million across four rounds from investors including SoftBank Vision Fund 2, Temasek, Horizons Ventures, Salesforce Ventures, and Mercedes-Benz itself. At its peak, Soul Machines had 253 employees and offices across Auckland, San Francisco, London, and Tokyo.
In February 2026, KPMG’s Leon Bowker and Luke Norman were appointed as receivers. All three marquee customers had abandoned the technology. Headcount had collapsed 82% from 253 to 45 before receivership. Both co-founders had departed — Cross in September 2023, Sagar in June 2024. The UK entity had been deregistered. Annual cash burn at the UK entity alone was US$38.1 million as of March 2023, with no publicly disclosed revenue to set against it. The total capital destruction across the company’s life: approximately US$135 million.
Soul Machines’ US market thesis was built on a specific moment: the convergence of AI maturity, enterprise appetite for digital transformation, and a genuine gap in the market for customer-facing interfaces that felt human without being human. The company positioned itself not as an AI vendor but as a platform — the infrastructure layer on which brands could build emotionally intelligent customer relationships at scale.
The ambition was not unreasonable. US enterprise spending on customer experience technology was and remains enormous. The incumbent solutions — IVR systems, rule-based chatbots, live agent centres — were universally disliked by customers and expensive for companies. A product that could replace or augment those systems with something customers actually preferred would command significant pricing power. The total addressable market was real.
The structural error was not the ambition. It was the failure to distinguish between enterprise excitement about the technology and enterprise willingness to deploy it at scale, pay for it on a recurring basis, and integrate it deeply enough into their operations to make switching costs real. Those are different things. The first produces marquee pilots and press releases. The second produces a business.
2016: Company founded in Auckland. Technology development begins. Early pilots with global enterprises generate significant press coverage and investor interest. 2019–2021: Four funding rounds raise US$135 million total. SoftBank Vision Fund 2, Temasek, Horizons Ventures, Salesforce Ventures, and Mercedes-Benz all invest. ANZ, Air New Zealand, and Mercedes-Benz deploy pilot implementations. Peak headcount reaches 253. Offices across four countries. November 2022: ChatGPT launches. Conversational AI is commoditised overnight. The premium positioning of Soul Machines’ bespoke digital human technology is structurally undermined. 2022–2023: ANZ ceases using the technology. Air NZ replaces Sophie with an in-house AI. Mercedes-Benz reduces its equity stake from 4.4% to 1.6%. September 2023: Co-founder Greg Cross resigns as CEO. June 2024: Co-founder Mark Sagar quits as director and chief scientist. UK entity deregistered September 2024. Headcount collapses from 253 to 45. February 5, 2026: KPMG appointed as receivers.
Soul Machines did not fail because the technology was fraudulent or the founders were incompetent. The technology worked. The founders were credible. The investors were sophisticated. It failed because three structural decisions — or three structural failures to decide — created a company that could generate extraordinary demonstrations of its capability without converting those demonstrations into the recurring, integrated, mission-critical deployments that enterprise software revenue requires.
The marquee pilot partnerships with Mercedes-Benz, ANZ, and Air New Zealand were Soul Machines’ most visible asset and, ultimately, its most misleading one. These partnerships generated global press coverage, investor confidence, and the kind of brand association that makes fundraising straightforward. They also created a structural illusion: that enterprise excitement about a technology and enterprise commitment to that technology as a core operational system are the same thing. They are not.
A pilot is an experiment. Its purpose, from the enterprise’s perspective, is to evaluate technology at limited cost, limited integration depth, and limited organisational commitment, before deciding whether to scale. The enterprise retains full optionality — it can deepen the relationship, or it can exit when the pilot ends, or when a better alternative emerges, or when the internal champion who drove the pilot moves to a different role. A pilot partnership that is not converted into a deeply integrated, contractually committed, switching-cost-generating deployment is not a customer. It is a reference that can be withdrawn.
All three of Soul Machines’ marquee pilots were withdrawn. ANZ ceased using the technology in early 2022 without a publicly stated reason. Air New Zealand replaced Sophie with an in-house solution that its own team built. Mercedes-Benz reduced its equity stake — the most direct possible signal that an investor-customer has changed its assessment of the relationship’s value. None of these were failures of the technology in deployment. They were failures to convert pilots into the kind of deep operational integration that makes switching too costly to pursue.
The correct architecture for a platform business selling to enterprises is not to maximise the number of marquee pilot partners. It is to maximise the depth of integration with a small number of committed accounts, building switching costs, expanding the scope of deployment, and generating the reference customer evidence that allows direct sales to new accounts. Soul Machines pursued breadth of partnership at the expense of depth of integration — and when the market shifted, there was nothing sticky enough to hold the customers in place.
Soul Machines never publicly disclosed its revenue figures. That opacity is itself a structural signal: a company that is generating meaningful, growing, recurring revenue from enterprise deployments has every incentive to disclose those numbers, because they validate the investment thesis and justify the valuation. A company that declines to disclose revenue — or cannot, because it is not material — is almost always protecting investors and employees from a number that does not support the narrative.
The business model for Soul Machines’ digital humans was necessarily bespoke. Each deployment required significant customisation: the avatar’s appearance, voice, personality, and knowledge base were built to the specification of the deploying brand. This is not a SaaS business — it is a professional services engagement with a technology layer on top. Bespoke deployments do not scale. They require expensive human capital at every new customer, generate one-time revenue rather than recurring subscription income, and produce a customer acquisition model where the cost of each new deployment is structurally high.
For Soul Machines to have built a durable enterprise software business, it needed a productised, repeatable, subscription-based deployment model where enterprises could deploy digital humans from a catalogue of configurations without significant bespoke engineering. That product did not appear to exist at the scale the capital raise implied. The company was burning US$38.1 million per year at the UK entity alone, with no evidence that its revenue model could ever close the gap between that burn rate and its contract economics.
November 2022 is the date most commonly cited as the inflection point for Soul Machines — the moment ChatGPT launched and conversational AI became commoditised. The narrative runs: Soul Machines built an expensive, bespoke solution to a problem that OpenAI then solved cheaply. The market shifted, the premium evaporated, and the business became unviable.
This narrative is partially true and structurally misleading. ChatGPT did not make Soul Machines’ technology obsolete — the photorealistic, emotionally responsive avatar capability that Soul Machines built is still genuinely differentiated from a large language model chatbot. What ChatGPT did was change enterprise risk tolerance for AI experimentation: it made the baseline so accessible and so cheap that the justification for a six-figure, heavily bespoke digital human deployment became much harder to make internally. The enterprise champion pitching a Soul Machines implementation to their board in December 2022 was suddenly asking for significant budget against a technology landscape that had just become dramatically more competitive.
The deeper structural point is this: if Soul Machines’ customer relationships had been deeply integrated, contractually committed, and operationally mission-critical before November 2022, ChatGPT would have been a threat to manage, not a death blow. ANZ’s decision to exit, Air New Zealand’s decision to build in-house, Mercedes-Benz’s decision to reduce its stake — these decisions reveal that the prior deployments were not mission-critical. They were experiments that could be stopped. ChatGPT gave enterprises a reason to stop them. But the real structural failure was allowing them to remain experiments for long enough that an external market shift could trigger exit without contractual or operational consequence.
THE BURN RATE REALITY
Annual cash burn at Soul Machines’ UK entity as of March 2023 — with no publicly disclosed revenue to set against it
At US$38.1 million annual burn with 253 staff at peak, Soul Machines was spending at a rate that required either very large enterprise contracts or a productised SaaS model to sustain. Neither appears to have existed at the required scale. The capital was burning. The revenue model was not catching up.
FAILURE DIMENSION ANALYSIS — SOUL MACHINES
ChatGPT’s launch in November 2022 was not the cause of Soul Machines’ failure, but it was the moment the company’s structural fragility became impossible to conceal. Before November 2022, the company could argue — credibly — that its bespoke, emotionally intelligent digital humans occupied a genuinely premium position in a market that had no accessible alternative. After November 2022, every enterprise customer’s board and procurement team had a reference point for what AI-powered conversational interfaces could cost and what they could deliver.
The customer departures that followed were not simultaneous — ANZ had already exited before the ChatGPT launch — but they all arrived within a compressed window that turned the narrative from “some customers are rotating out as pilots end” to “the marquee customer base has abandoned the technology.” That distinction matters because enterprise sales is a reference business. A platform that cannot name its active, satisfied, expanding customers cannot sell to new ones. By late 2023, Soul Machines had no active marquee references to sell from.
The co-founder departures in 2023 and 2024 were the final structural signals. When the people who built the technology and led the company exit in sequence, the organisation’s ability to pivot, rebuild, or stabilise is gone. By the time KPMG was appointed in February 2026, there were 45 people remaining from a peak of 253. The company had spent seven years and US$135 million building toward a market that was not there in the form the model required.
Soul Machines is the definitive NZ technology case study in the difference between demand signals and distribution readiness. The company had extraordinary demand signals: world-class investors, Fortune 500 pilot partners, global press coverage, and a technology that genuinely impressed everyone who experienced it. What it did not have was the distribution architecture — the deeply integrated, contractually committed, switching-cost-generating customer relationships — that turns impressive technology into durable enterprise revenue.
The failure is not reducible to ChatGPT. It is reducible to a business model that was built around demo excellence and pilot acquisition rather than integration depth and recurring revenue. A company that burns US$38 million per year needs either very large recurring contracts or a productised deployment model that can be sold at volume. Soul Machines had neither in sufficient quantity. The capital funded the demonstrations. It did not fund the distribution architecture that would have made those demonstrations into a business.
Distinguish between a pilot partner and a customer. A pilot partner is an enterprise that is evaluating your technology. A customer is an enterprise that has integrated your technology into a mission-critical workflow, renewed its contract, and expanded the scope of deployment. Before counting pilot partners as evidence of product-market fit, ask: if this enterprise’s internal champion moves roles tomorrow, does the deployment continue? If the answer is no, the relationship is not a customer. It is an experiment that is still running.
Build switching costs before you raise the next round. In enterprise software, switching costs are not a commercial tactic — they are the structural foundation of recurring revenue. If your customers can exit without disrupting their core operations, they will exit when a cheaper or more accessible alternative arrives. Before you scale, identify the specific integrations, data dependencies, and workflow embeddings that would make leaving genuinely painful. Build those before you claim the customer as a reference.
Revenue opacity is a signal, not a strategy. A company that does not disclose revenue while burning US$38 million per year is almost certainly protecting itself from a number that contradicts the valuation. Investors who accept revenue opacity in a company at this burn rate are accepting a structural information asymmetry that benefits the founders and harms the investment thesis. If the revenue cannot be disclosed, the model is not working at the required scale.
Soul Machines is the clearest available case study of what the research calls confusing demand signals with distribution readiness — and it is the most expensive version of that mistake in NZ tech history. The demand signals were genuine. The technology worked. The investors were world-class. The pilot partners were the brands every NZ founder wants on their reference list. None of that produced a sustainable business because none of it was the business.
The business is recurring revenue from deeply integrated, contractually committed customers who cannot exit without disruption. Soul Machines never demonstrated that it had built that. The US$135 million funded the demonstrations of what was possible. It did not fund the architectural work of converting those demonstrations into the kind of customer relationships that produce a durable enterprise software company.
A pre-entry US architecture for a company in Soul Machines’ position would have asked three questions before the Series B: What is our conversion rate from pilot to committed contract, and what does that imply about our total addressable market in the near term? What is the integration depth of our current deployments, and what would it cost a customer to exit? And what does our revenue model look like at the scale our burn rate requires — specifically, how many contracts at what ACV do we need to be sustainable, and when do we expect to have them?
Those questions do not require a market shift to answer. They require honesty about what was there. Soul Machines had extraordinary technology, a compelling narrative, and world-class investors. What it needed — and did not have in sufficient quantity — was a business model that could survive contact with the US enterprise procurement reality at scale.
“The capital funded the demonstrations. It did not fund the distribution architecture that would have made those demonstrations into a business.”
— PIVOTAL CATALYST VERDICT
FREQUENTLY ASKED
Did ChatGPT cause Soul Machines to fail?
Not directly. ChatGPT changed enterprise risk tolerance for AI experimentation and made it harder to justify bespoke, expensive deployments. But the structural failures — pilot partnerships that were never converted into deeply integrated, contractually committed relationships, a revenue model that was not designed for scale, and a burn rate that required either very large contracts or a productised deployment model to sustain — were embedded before November 2022. ChatGPT was the accelerant. The structural fragility was the cause.
What is the difference between a pilot partner and a customer?
A pilot partner is evaluating your technology at limited integration depth and limited organisational commitment. A customer has integrated your technology into a mission-critical workflow, renewed its contract, expanded the scope of deployment, and built operational dependencies that make switching genuinely costly. Soul Machines had pilot partners. It needed customers. All three marquee pilot partners exited when their experiments ended — which is what pilot partners do when the experiment does not convert into a committed customer relationship.
What should NZ deep tech founders take from this?
Three things. First, impressive demos and marquee pilot partners are not product-market fit — they are evidence of technology interest, which is necessary but not sufficient for a business. Second, switching costs are not optional for enterprise software — they are the mechanism by which pilot relationships become durable revenue relationships. Build integration depth before you claim references. Third, revenue transparency is a discipline, not a disclosure choice — if you cannot show the revenue trajectory that justifies the burn rate, the model is not working at the required scale, and more capital will not fix a structural design problem.
Go In Knowing
The US Entry Diagnostic delivers a structurally sound plan to enter the US before you invest $500,000 in painful mistakes.
Book the Paid Diagnostic$10,000 NZD. Credited in full toward any engagement.