Angela McNeal Is Building for AI’s Messy Middle

Angela McNeal has spent much of her career where artificial intelligence stops being a captivating demonstration and starts acquiring consequences. A model can produce a striking answer in a clean window. An enterprise has old databases, permissions, auditors, unpredictable edge cases and people whose judgment still matters. Between those two worlds lies what McNeal calls the “messy middle.” It is not glamorous territory. It is also where she has chosen to build.

Her route there begins at Columbia Engineering, where she studied computer science with a minor in applied mathematics and specialized in intelligent systems. The combination feels almost biographically tidy: code, mathematics and an early interest in machines that make decisions. Yet her professional story has been less about elegant abstractions than the awkward business of making technical systems answer to reality.

After Columbia, McNeal worked at Goldman Sachs. She later moved to Palantir, ultimately becoming head of AI product for Foundry. The job placed her close to large organizations and government agencies attempting to use machine learning in operating environments, where a wrong result cannot simply be dismissed as an amusing hallucination. Her work encompassed product strategy, responsible AI, privacy and ethical data handling. The lesson she carried forward was blunt: capability gets attention; deployment earns trust.

A proof needs a lemma

At Palantir, McNeal worked alongside engineering leader Mayada Gonimah. Years spent putting data and AI systems into demanding institutions gave them a shared view of the market’s missing layer. Enterprises wanted to adopt new models, but their software estates were fragmented. Building connective infrastructure from scratch was slow and expensive. Buying a narrow application solved one workflow and stranded the next. The pair believed the useful company sat between those choices.

They founded Thread AI in 2023. Its platform, Lemma, emerged from stealth in October 2024 with a $6 million seed round led by Index Ventures, alongside Greycroft, Homebrew and angel investors. The name is an engineer’s wink. In mathematics, a lemma is a small result used to prove something larger. In Thread’s version, reusable components connect models, data, APIs and business logic into workflows that can be inspected and changed.

The product idea contains a quietly unfashionable proposition: a business should not have to detonate its technology stack to benefit from AI. Lemma is meant to meet companies where they are. It orchestrates event-driven workflows across existing systems while preserving controls, human handoffs and visibility into what happened. Models may change by the month; institutional responsibilities are less forgiving.

The early customer set made the breadth of that integration problem plain. Thread worked across luxury hospitality, digital marketing, public safety and financial services. VaynerMedia used the platform to tailor workflows to its own requirements rather than accept a generic application. BRINC Drones used Lemma to connect AI capabilities with a physical-product environment. One platform had to accommodate very different data, risk and operating habits. The common problem was not access to a model. It was getting the model safely through the front door and teaching it the building’s rules.

The last ten percent

By 2025, generating an AI prototype had become remarkably easy. McNeal’s argument was that easy beginnings can disguise expensive endings. The final stretch contains durability, scalability and reliability. A workflow that fails once in a thousand runs may look excellent on a dashboard and still create a serious operational or reputational problem. Her preferred test is disarmingly practical: run an essential workflow on the platform beside the existing solution and compare speed and stability.

That practical streak shapes how McNeal talks about AI. She resists both the prophecy that machines will replace everyone and the shrug that they are toys unfit for consequential work. Years of deployment led her to a less theatrical position. AI can perform useful work inside important systems, provided its freedom is bounded and its actions remain visible. The hard questions are about permissions, failure, escalation and accountability. They sound less like science fiction than good operations.

McNeal describes the desired result as controlled autonomy. An agent may reason and act, but only inside explicit boundaries. Thread’s framework divides the work into control, governance and reliability. Control limits access and action. Governance supplies oversight and traceability. Reliability keeps the process durable when inputs, services or models misbehave. The sequence is a useful corrective to an industry forever tempted to begin with the magic trick.

This concern has a human edge. McNeal and Gonimah wrote during the Series A announcement that technologists have an obligation to elevate human capabilities as AI changes work. McNeal’s formulation is not merely ergonomic. At Columbia Engineering’s 2025 Women in Science and Engineering conference, she called human involvement a moral stance and urged students to build systems that augment expertise. It is the kind of sentence that sounds philosophical until an automated system reaches payroll, procurement or public safety.

The concern predates Thread. While at Palantir, McNeal co-authored work on enabling responsible use of AI through model-management infrastructure. The premise was operational: responsibility cannot live only in a policy document. Teams need mechanisms to develop, evaluate, deploy and maintain models while preserving the context in which those models act. That continuity helps explain why Thread’s pitch joins powerful new capabilities to decidedly mature words such as constraints, observability and control.

Strong opinions, inspected often

Returning to Columbia as a keynote speaker also revealed McNeal’s working temperament. She encouraged “relentless curiosity” and “strong opinions, weakly held.” The pairing matters. Infrastructure founders must maintain a point of view long enough to build difficult things, then surrender it when the system offers contrary evidence. Confidence without inspection produces brittle software. Humility without conviction produces no software at all.

Thread’s own chronology shows that rhythm. The company came out of stealth in 2024, raised a $20 million Series A in June 2025 and expanded its focus across enterprise and public-sector work. McNeal wrote at the end of 2025 that customers had moved beyond novelty. They wanted authentication, relationship-based access control, provisioning and dependable human-in-the-loop systems. The “unsexy” security layer, in her word, became the roadmap.

In 2026 she took that argument to HumanX, joining a session on moving prototypes into production and presenting Thread’s framework for critical environments. She also began publishing The Handshake Protocol, a newsletter named for the deliberate design of how people and machines work together. The title is apt. A handshake is an agreement, a boundary and a small ceremony of mutual recognition. It assumes two parties remain present.

The newsletter gives McNeal a venue to report from inside deployments rather than comment from the grandstand. She has promised to examine what works, what fails, which architecture choices matter and which questions about governance and trust are being rushed. Its premise is that the industry’s two loudest positions, total replacement and total dismissal, both miss the lived texture of enterprise AI. McNeal’s more hopeful view comes with chores attached. Human-machine collaboration has to be designed deliberately, one consequential handshake at a time.

There is commercial ambition here, naturally. Thread wants Lemma to become the observability and orchestration layer for agents running across an organization. Its customers have included businesses in digital marketing, public safety, financial services, hospitality and other complex sectors. The promise is not one miraculous app. It is a place where many models and systems can cooperate without turning the enterprise into a swamp of new dependencies.

Making the invisible dependable

McNeal has said the best tools should feel invisible, meeting workers inside the software they already use rather than adding cognitive load. That invisibility is earned through a great deal of visible engineering. Authentication has to work. Data must stay governed. A human needs to know when to intervene. The system must explain what it did after the moment has passed. Reliability, like good stagecraft, succeeds partly by concealing the labor that made it possible.

Her story is therefore not a tale of chasing each new model release. It is a story about taking new capability seriously enough to impose discipline on it. Columbia supplied the fundamentals. Goldman Sachs exposed institutional scale. Palantir offered years inside operational AI. Thread AI is the wager that those lessons can become reusable infrastructure rather than bespoke experience.

The AI business likes clean before-and-after pictures. McNeal keeps pointing to the middle: the databases that disagree, the rules that cannot be hand-waved, the expert who must retain judgment, the workflow that needs to survive Friday night. It is messier than a demo and more useful than a prophecy. For a founder, it is also a generous supply of problems worth owning.