Add preferred source
Thread AI CEO Angela McNea argues that most attempts to automate decisions with AI break down outside tightly constrained problems, and that the durable path is “controlled autonomy” — agents acting inside governed systems with humans injectable at any point. Speaking on the MTS podcast, the former Palantir engineer says enterprises keep rebuilding the same infrastructure plumbing (identity, logging, isolation, governance) while spending heavily on consultants to assemble stacks, a gap she reads as a market signal. Her sharpest point is that workflow design, not model selection, is the real intellectual work: someone has to codify which error is worse, who gets alerted, and what “expensive” means before any routing happens. The stakes range from billions in reconciliation losses at banks to life-and-death deconfliction work with Army DEVCOM. She is explicitly not in the AGI camp, and expects iteration rather than instant automation.
Key Elements
Angela McNea has a message for every company that bought an AI agent and expected it to run the business by Friday: the decision-making part is where it falls apart. Across the stack and across the process, she says, automated decision-making has mostly been unsuccessful — “except when the feature space is so constrained,” meaning a finite set of paths and options, like routing a ticket to a person or picking one of a few actions.
“When you open up the aperture, I’ve seen things struggle still,” said McNea, CEO and co-founder of Thread AI, speaking on the MTS podcast. She does not count herself in the AGI camp. That is an unusual thing to hear from someone selling agent infrastructure, and it is the key to understanding what her company actually bets on.
The bottleneck moved, and nobody cleaned up the plumbing
McNea’s diagnosis starts with a shift. The constraint is no longer training models — it is deploying them. A real enterprise runs tens to hundreds of software systems, multiple identity providers, heterogeneous data, and layers of risk and regulatory rules. Faced with that, companies respond by assembling and reassembling stacks from scratch.
She watched two failure patterns before founding Thread AI, and they are worth memorizing because most companies are living one of them right now.
| Failure pattern | What companies did | What broke |
|---|---|---|
| Vertical tool sprawl | Bought point tools for invoice processing, procurement, marketing automation, reconciliation | Vendor management “goes pretty wild”; a build that worked in one business unit could not move elsewhere |
| Everyone-is-a-builder | Non-engineers drew large capability maps and assembled stacks | Operability collapsed when a coding provider changed, a team turned over, or the system needed load testing |
Thread AI’s answer is to supply the missing primitives — isolation, metering, identity, logging, governance — so teams spend their time on the workflow rather than rebuilding plumbing. The platform governs the whole transaction rather than individual actions, built on what McNea calls “controlled autonomy”: agents reason and act inside limits the customer sets, humans are injected either when an agent escalates or by policy, and everything is captured in a record.
McNea and co-founder and CTO Maya both spent six years at Palantir, four of them together, before leaving in spring 2023. They raised a Series A in 2025. The founding thesis was a bet that “the glue, the harness, the control, the governance” would matter — and she says the compounding of frontier, sovereign, and local models has only reinforced it.
The origin story is precise: Meta’s Toolformer paper, in which a language model’s last layers are trained to call simple APIs like a calculator or a weather service. As a software engineer, McNea saw the mismatch immediately. Enterprise APIs mean thousands of providers, each with thousands of endpoints and hundreds of parameters, across different protocols, authentication schemes, and data types. But the idea that flexible decision-making could live in software was compelling, so she worked backward from a best-case future and arrived at an infrastructure company.
Routing is the easy part. Defining “hard” and “expensive” is the wall.
In her essay “every good decision is an ensemble,” McNea argues that the routing debate is aimed at the wrong layer. The router technology itself, she says, is less interesting than the application around it. Someone has to predefine what counts as a hard problem, what counts as expensive — in time, people, and money — and what the system is optimizing for. All of that happens before any routing.
Thread AI lets customers build routers natively, but she says the team spent most of its effort on those explicit decision and logic gates, because “in the real world, those were the actual barriers to putting something in production.”
She extends the ensemble idea beyond model routing to institutions, pointing to how the judicial system distributes decisions rather than vesting them in a single actor. The practical read for buyers: model-selection tooling is commoditizing, while the codified business logic wrapped around it is where both differentiation and risk live.
A clinical trial shows why workflow design is the real work
Her sharpest example is a model deciding who can join a clinical trial. A classification model outputs yes or no; evaluation surfaces accuracy, precision, and recall. But a business person and the scientists must decide which error is worse: admitting someone who is a poor fit and may be harmed by the drugs, or excluding someone who might have benefited.
“It’s never like yes or no. It’s an investment someone makes in the logic as they’re building these flows,” she says.
The same applies to a triage workflow. Building it is tractable; deciding who receives alerts and at what volume — whether they are too noisy or not important enough — is the hard part. Existing enterprise processes were designed deliberately, with legal, regulatory, and human factors baked in, and injecting AI does not remove those factors. It forces a redesign. McNea says enterprises are putting a meaningful share of their IP into exactly that redesign, while individual productivity tools remain “wild west.”
“The construction of the system, the construction of the process, is often the intellectual work.”
She links this to a shift in the developer role at large institutions: engineering is moving toward managing, reviewing, and designing rather than implementing, with a rough parallel emerging for knowledge workers and domain experts.
The stakes: billions in losses, or life and death
Being a dual-use company mattered to McNea and Maya from the start. She calls the U.S. Department of Defense the most demanding customer she can imagine. Thread AI has publicly discussed research with Army DEVCOM focused on manual deconfliction — getting accurate information to the right person at the right time, both for scientists maintaining and innovating control systems and for warfighters in the field. She describes the stakes there as life and death.
On the commercial side, the canonical example is reconciliation flows at a large financial institution, where an agent error can cost billions of dollars and carries legal and reputational exposure. Thread AI sells across financial services, manufacturing, hospitality, and advertising, and McNea says the same primitives serve both defense and enterprise workloads.
That range is the whole argument for horizontal infrastructure: build the governance layer once, apply it everywhere the stakes are high.
They modernized the best practices instead of discarding them
Asked what is hardest to automate, McNea points to the design layer, not the technology. She credits her and Maya’s “superpower” as building primitives that have adapted to four years of technological change, and says they were explicit from day one about not throwing away 20 to 50 years of distributed-systems and security best practice — instead modernizing it for adaptability.
The platform is self-service, paired with an applied arm that works alongside customers to share best practices and co-construct workflows. That services layer is not incidental. McNea reads the surge in enterprise AI spend partly as consulting dollars — companies paying outsiders to assemble stacks — and takes it as a clear signal of a gap a platform should fill.
“People are spending a lot of money asking for help and having those people assemble and build those stacks for them, which to me is a perfect indicator that there is a gap for a platform to fill.”
On the roadmap, she flags hiring around public-sector work and upcoming building blocks for companies that sell products and want them more adaptable, agentic, and AI-forward.
Her method: climate-scientist parents, and humility from the field
McNea’s problem-solving lens comes from two climate-scientist parents and a Columbia degree in computer science and math — deconstruct the problem, establish a base case, hypothesis-test. She pairs that with the central lesson of her forward-deployed engineer years at Palantir: founders need “a balance of confidence and conviction and humility,” because the customer is the expert in their domain, not you.
That humility shows up in her skepticism. Her evidence against broad automation is empirical, not ideological. Automated decision-making across the stack has mostly been unsuccessful except where the feature space is tightly constrained — a finite set of paths and options, such as routing to a person or trying one of a few actions.
She attributes part of the gap to expectations rather than capability. AI workflows are iterative, and the death of innovation is the inability to get started. “You’re not going to have the perfect automated system, the perfect agent out of the get-go.”
Her prescription is to ship the building blocks early, capture the human-AI exchange, and build a data asset the system can learn from and evolve on — rather than assuming every decision can be automated from day one.
There is an unresolved tension running through the episode: between McNea’s skepticism of broad automation and Thread AI’s commercial bet on agents in high-stakes settings. Her answer is governance and human injection points, but she does not quantify how much human involvement customers actually retain once a workflow matures — a question worth pressing as the public-sector work scales. The second question to track: whether the consulting spend she cites as a platform opportunity converts into self-service adoption, or persists as services revenue that keeps enterprises dependent on outside assemblers.
For investors, the episode points at a less glamorous layer of the AI trade than chips and data centers. If McNea is right, the value accrues to whoever owns the identity, logging, governance, and decision logic around agents — the unglamorous middle where an error costs billions and a well-designed workflow compounds. The companies paying consultants to rebuild that layer are the ones whose budgets will eventually move somewhere. Whether they move to a self-service platform, or keep renting the work, is the open question that decides which side of this market wins.
Full content available at:Why Full AI Automation Fails in Production | Angela McNea
- RBA’s Hunter Says AI Productivity Gains Still Elusive as Data Center Boom Fuels Inflation
- ByteDance Tightens External AI Tool Access, Bans L4-Level Code Repositories and Feishu Documents
- Big Tech Rallies as Voters Put the Brakes on AI Data Center Buildout
- Microsoft AI Chief Mustafa Suleyman: Models Have No Consciousness, and We Shouldn’t Build It Into Them
