Matt Fitzpatrick on Enterprise AI, Expert Data and Deploying What Works

Harry Stebbings speaks with Matt Fitzpatrick, CEO of Invisible Technologies, about enterprise AI deployment, expert data, recruiting and the company's business model. Fitzpatrick sells both AI training and enterprise implementation services, a commercial interest that informs his views on the value of human feedback and deployment engineers.

Fitzpatrick says he spent 12 years at McKinsey and led QuantumBlack Labs, a technology development group. He recalls the firm's engineering workforce growing from roughly 100 to 7,000 during his time there and says he oversaw about a fifth. Those are his recollections, not independently verified workforce figures.

He met Invisible founder Francis Pedraza socially and later accepted an invitation to become CEO. Leaving a stable job was difficult, he says, but he wanted to build a company around the subject he had studied for a decade. A mentor urged him to consider the regret of not trying.

Invisible combines model training and an enterprise software platform, Fitzpatrick says. He argues that model capabilities and consumer use have advanced faster than enterprise adoption, because deployment requires usable data, redesigned workflows, accountable operators, monitoring and trust—not just a capable model.

He cites several reports to support his argument, including a claim about only 5% of generative AI deployments working and a forecast that 40% of enterprise projects may be cancelled by 2027. The interview does not establish the reports' precise populations, definitions or methods; the figures should not be read as verified rates for all enterprise AI.

In his view, risk management, validation and integration will make enterprise adoption a five-to-ten-year process rather than a one-to-two-year transition. This is his forecast, not a measured adoption timetable.

Stebbings asks whether large banks will buy outside tools or build internally. Fitzpatrick says the answer varies by sector and that banks often favor internal development. He cites a reported comparison suggesting externally driven builds are more effective, but the interview does not establish how those projects were selected or what effectiveness measured.

Fitzpatrick argues that outside suppliers tend to be held to explicit milestones and returns, while internal initiatives can lack the same discipline. He also says experienced AI talent is scarce. His company competes to provide external deployment, so those assessments come with a direct business interest.

He recounts an unnamed retailer that allegedly spent $25 million on a returns agent and later returned to deterministic workflows. Its evaluation reportedly emphasized speed and sentiment but overlooked costly incorrect answers. The anecdote is his account and does not establish a typical cost or failure rate.

For executives and finance leaders, he recommends identifying the data needed for a specific initiative, assigning an accountable business owner, setting milestones and paying against demonstrated outcomes. A finance leader need not personally be an AI specialist to ask those questions, he argues.

Instead of funding hundreds of experiments, Fitzpatrick recommends selecting three or four operational priorities and assigning business leaders measurable objectives. In a contact center, those might include resolution quality, cost per call and routing—not merely whether a model produces a response.

Stebbings asks how a buyer can choose among a flood of vendors whose claims sound similar. Fitzpatrick cites an agent evaluation with weak single- and multi-turn results, but the interview does not specify enough about its benchmarks to generalize the percentages to every agent.

His advice is to start with a proof of concept and not pay until the technology proves useful. He says Invisible offers customers an eight-week free solution sprint. This is a description of his company's sales approach, not an independently verified comparison of vendor offers.

Fitzpatrick describes five components of Invisible's platform: data integration, agent building, workflow tools, an expert marketplace and evaluation. He says the same architecture can be configured across sectors, and cites approximately 1.3 million experts entering its marketplace annually. The company figures are self-reported and the precise definition of the expert count is not clarified.

As an example, he describes work for concierge healthcare company Lifespan MD: combining fragmented records and business data, building reporting and adding tasks such as scheduling. His timeline and capability claims are company descriptions, not evidence that the system has been clinically validated or that any specific health intervention works.

He says forward-deployed engineers are essential when a customer must change its workflows and connect systems. Invisible has opened offices and assembled a deployment team; his broad claim that enterprise deployments cannot work without this model is his judgment as a seller of such services.

Fitzpatrick distinguishes people who merely answer product questions from engineers who build a specific customer workflow. He says Invisible typically customizes a deployment over roughly three months and expects ongoing adjustment as business conditions and models change. Those timing claims are not a universal benchmark.

He says Invisible does not separately charge for its forward-deployed engineers and collects payment once software is working. The company shifts spending from sales toward deployment staff, he explains, but those staff still have a cost and the interview does not independently establish the resulting margins.

His distinction for founders is between a relatively standard information product, which may not need embedded engineers, and a workflow transformation that requires integration and customer adoption. Even so, the claim that deployment engineers are the only workable method reflects his own approach.

Fitzpatrick argues that enterprise software has always involved substantial configuration, often supplied by systems integrators. He believes a small SaaS contract can struggle to support intensive customization, while generative AI deployments often require more customer-specific data and validation.

He describes user acceptance testing as the point at which Invisible typically begins receiving payment, often after two to three months. He compares the approach to earlier enterprise machine-learning deployments rather than assuming every AI product can be installed as an identical off-the-shelf box.

Stebbings asks whether Invisible is mainly a talent marketplace. Fitzpatrick prefers the label AI training platform: sourcing specialists is only one step, he says; work also includes assessment, organizing tasks and validating outputs. He does not disclose an exact revenue percentage for the marketplace.

He says AI training represented a material part of the earlier business and that enterprise deployments are now a growth focus. He reports 12 enterprise deals in 45 days, but does not disclose their value or define whether they are bookings or recognized revenue.

A small number of model developers makes customer concentration a feature of the training market, he acknowledges. He expects enterprise customers and specialized model developers to broaden the customer base, but does not provide Invisible's exact concentration figures.

Stebbings asks how a supplier negotiates with customers that account for substantial revenue. Fitzpatrick says good data can prevent wasted compute and claims that customers compare suppliers' results. He disputes the idea that buyers will pay any price and says projects have price boundaries; these are his market observations.

Fitzpatrick expects human expertise to remain important, especially for complex, multilingual and multimodal tasks. He contrasts these with tasks such as mathematics where an answer can sometimes be checked against a clear result. His decade-long demand forecast reflects the interests and outlook of a data provider.

He invokes institutional memory: teams that repeatedly source specialists, organize tasks and evaluate outputs may become difficult to replicate. He mentions about 26,000 experts working in a given week alongside the larger marketplace figure; the interview does not establish that all of the 1.3 million are simultaneously active.

He says data work has moved beyond simple image labels toward narrow expertise, giving the example of a French-speaking specialist in seventeenth-century architecture. His account emphasizes recruiting, assessment and accumulated performance data rather than exclusive ownership of every expert.

Fitzpatrick compares expert pay to marketplace price discovery: the rate depends on the skill, timing and location. Paying more does not guarantee a good expert, he says. The interview does not provide wage distributions, worker conditions or independent evidence that any particular rate is fair.

He does not expect one company to control a fixed pool of experts or dominate the entire market. Needs change, he says, and a marketplace must maintain sufficient demand and supply. His expectation of several providers is a forecast rather than a measured market structure.

As an example of specialized enterprise work, he describes collaboration with SAIC, Vantor and the US Navy on models for underwater unmanned systems. He says the work involves interpreting sensor information and possible responses. The interview supplies no independent assessment of the system's performance or deployment.

Public benchmarks can help people follow model progress, Fitzpatrick says, but they may not predict whether a specific enterprise task meets a customer's requirements. He expects more task-specific testing, validation and risk management, especially where errors carry significant consequences.

Stebbings asks whether automating junior tasks could damage the future talent pipeline. Fitzpatrick expects adoption to be slow enough for workers to adapt and says his company is hiring younger staff. His accounting analogy suggests technology can expand work as well as automate it, but it does not establish what will happen to AI-era employment.

He predicts AI training and enterprise services will each support several providers rather than a single winner. He praises Palantir for investing early in customized deployment, while saying companies such as Surge, Mercor and others compete with different parts of Invisible's business.

Stebbings questions whether large data-labeling figures represent revenue or gross marketplace volume. Fitzpatrick calls them revenue and points to varying project margins and expert rates. That explanation alone does not resolve the accounting distinction between gross and net revenue; the interview does not supply audited recognition details.

He rejects the claim that synthetic data will eliminate human feedback within a few years. His argument is that context, private information and complex judgments can require expert review. He predicts humans will remain involved for decades; that is a forecast from an operator whose company sells human expertise.

Asked about profitability, Fitzpatrick says Invisible had historically raised relatively little primary capital, then raised approximately $130 million and increased investment. He states that the company will not be profitable that year. These financing and profitability statements are self-reported, with no financial statements presented.

He argues that the current opportunity justifies investing rather than harvesting cash and describes a long-term plan for enterprise and training products. More capital also creates execution risk; the interview provides his rationale rather than a forecast of actual returns.

One area he wants to expand is physical-world data. He describes engineers visiting farms, connecting remote sites and building computer-vision models for herd monitoring. He sees similar opportunities in energy and robotics, while acknowledging that the work requires time and capital.

Fitzpatrick says Invisible had historically kept a low public profile but has increased its attention to brand. He sees visibility as a way to build trust and says the external story should match what customers and employees experience.

Stebbings challenges whether startups must sell ahead of delivery. Fitzpatrick draws a distinction between transparently proposing an unbuilt solution and claiming an agent can already do anything. Nondeterministic systems can fail in ways that polished demonstrations do not reveal, he says.

Reflecting on early work at McKinsey, he says he built trust by explaining why an idea might work and iterating with customers, rather than promising certainty. He describes recruiting capable people as central to building a useful product.

His hiring advice emphasizes retaining and developing strong people as well as recruiting them. He prefers adaptable employees who can take on more than one role, and believes an intellectually engaging workplace helps keep them.

Stebbings argues that demanding cultures can produce winning teams. Fitzpatrick qualifies his earlier answer: an execution-heavy business may organize differently from a research-oriented AI company. He says Invisible also pushes hard on delivery, while trying to leave room for engineers to explore.

He describes a capital-allocation tension between steady growth and a much larger investment in expansion. Invisible intends to spend in the near term while maintaining a path to profitability, he says. That is a management goal, not evidence of future profitability.

Invisible was fully remote for years and now works largely in person while retaining some remote staff, Fitzpatrick says. He believes offices improve collaboration, customer proximity and culture. His assertion of a productivity increase is an observation without a disclosed measurement or controlled comparison.

He separates office location from work hours: in his view, people can collaborate in person most weekdays while retaining flexibility to work from home when needed. He says he does not require everyone to be in the office six days a week and wants to avoid losing flexibility.

On management, he says centralized control becomes less useful as customers and projects multiply. His response is to flatten the organization, share tools and decision frameworks, and empower teams close to customers to act. He draws an analogy to field decision-making in military organizations.

Long-range culture and operating principles still matter, but Fitzpatrick considers detailed five-year AI product plans less useful when technology changes rapidly. He favors stable core beliefs combined with frequent iteration and interoperable systems; this is his operating philosophy.

He describes spending much of his time traveling and says frequent communication and a supportive partner help sustain his relationship. He expects the intensity to reflect a particular growth stage rather than continue indefinitely.

In a quick-fire round, he declines to choose between investments in model providers, praises Databricks' data infrastructure and returns to his CEO priorities: recruit strong people, build a culture in which they work well together and help them benefit from what they create.

He argues that off-the-shelf agents will not automatically fix every business process. His investment interest would include companies using AI to deliver real-world services, not only vendors selling agent software. His comments on software margins and potential returns are forward-looking judgments, not audited comparisons.

Looking ahead, Fitzpatrick is optimistic about energy, healthcare and education. He cites various numbers on electricity, medical spending and errors, but the interview does not provide the definitions or evidence needed to verify them. His claims that AI will yield net environmental benefits or improve clinical outcomes are hopes and predictions, not established results.

In education, he hopes accessible tools can widen learning opportunities and that employers will assess people by skills as well as credentials. He notes that Invisible recruits some people without college degrees. The conversation closes with that aspiration, without establishing how AI will affect educational inequality overall.