Free with a sign-in.
The best 5 min of 29
Why Full AI Automation Fails in Production | Angela McNea
Angela McNeal argues that high-stakes AI needs controlled autonomy, explicit decision logic and human intervention—not full automation.
MTS · Theo Jaffee & Sophia Dew
- Angela McNeal is CEO and cofounder of Thread AI.
The best minutes
- 7:13Model routing depends on explicit business logic3 min
- 25:27AI workflows improve through iteration, not instant automation2 min
Also worth watching
- 4:25Enterprise AI stacks can fail as providers and teams change3 min
- 13:22Clinical AI needs a deliberate choice about which errors matter4 min
- 10:13Thread's Army DevCom research targets manual deconfliction3 min
Skip the Lovable ad, the roughly one-minute mid-show sponsor block, and the closing sponsor reads.
Full analysis · 5 min read
Real Alpha · Nov 8
Angela McNeal says broad AI decisions struggle beyond constraints
- In a hypothetical reconciliation example, mistakes can cost 'like, billions of dollars,' McNeal said.
- Her account puts workflow policy ahead of routing mechanics, but offers no comparative deployment results.
- Start with constrained tasks, record human exchanges, and measure what improves before widening scope.
Free with an account
Read the rest of the analysis
6 sections, the 5 other voices it tests McNeal against, and what it means for you. Every episode we cut gets one.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.
Reconciliation illustrates the downside, not a measured loss
McNeal framed reconciliation as a hypothetical enterprise use, with agents helping a large financial institution. She said mistakes could carry a striking financial cost. The phrasing is illustrative scale, not a reported loss.
Her setup matters because she was describing agents handling a process, not reporting an AI-caused incident. She described huge, like, financial risk, with 'potentially, probably, actually, like, like a legal risk element to it.'
For defense, she called the stakes life or death. The episode supplies no frequency, loss estimate, or measured outcomes for either example. That absence limits what the examples establish, but does not rebut the underlying risk.
“Mistakes there can cost your business, like, billions of dollars.”
Our read
The risk case explains why buyers care. It does not show a specific system's loss rate or effectiveness.
Business rules define what a router should optimize
McNeal said router technology was less interesting than its application. The harder work, in her account, is defining what the business wants to optimize. That work precedes useful model selection.
She asked whether expensive means money, time, or people's time. Latency was another trade-off she had named. Teams therefore need distinct priorities, rather than assuming one universal score.
Thread lets customers build routers natively, she said, while focusing on explicit decision gates. This makes her point more specific than choosing a better router and more dependent on customer judgment.
Woodson Martin of OutSystems spoke on The Cognitive Revolution in October 2026. He said most enterprise workloads do not need frontier models, and described a custom harness and router to cut token costs.
That is a model-fit and cost case, not evidence for McNeal's business-policy point. Harrison Chase of LangChain said on 20VC in October 2026 that vertical-agent routing should be harness-aware, while gateways may suit general chat or coding. His point concerns routing context, not who sets the business objective.
“the hard part is the application around it”
Woodson Martin · The Cognitive Revolution · Oct 7Most enterprise workloads do not need frontier models
Harrison Chase · 20VC · Oct 17Useful model routing needs to understand the agent harness
Our read
Treat routing as an expression of policy, not policy itself. Define the task and business priorities before comparing router performance.
Production strain appears after teams assemble the stack
McNeal described one group buying separate vertical tools for tasks like invoice processing, procurement, marketing, and reconciliation. That could make vendor management go wild, she said, especially when teams later wanted to move useful tools elsewhere.
Her second pattern was different. More people can build systems, but operating assembled stacks becomes difficult. She said she saw stacks fall over when coding providers changed, teams turned over, or load testing increased sharply.
So assembling a demo and operating it over time are separate tests. McNeal said Thread provides building blocks so customers need not rebuild isolation, metering, identity, logging, and governance.
Amjad Masad of Replit said on Lightcone in November 2026 that years of systems work gave its agent a head start. He cited transactional, reversible filesystem engineering. That is a distinct product example, not proof of Thread's platform reliability.
Our read
Test day-to-day operation separately from initial assembly. Provider changes, team turnover, and higher load are concrete failure conditions to exercise.
Clinical eligibility turns metrics into a policy choice
McNeal introduced clinical-trial eligibility as a metaphor she used with her former Palantir team, not a Thread deployment. A classifier may admit an unsuitable patient or exclude a suitable one. She named accuracy, precision, and recall as evaluation metrics.
She then shifted from scores to judgment. Teams and scientists must weigh a patient receiving a harmful mismatch against someone being missed. The system's logic can codify that choice.
For example, prioritizing avoidance of unsafe admissions can mean accepting more exclusions of eligible patients. That is an implication of the trade-off, not McNeal's recommendation.
Metrics make performance visible, but they do not decide which mistake an institution values less. The business and scientific teams have to make that choice before the system can apply it consistently.
Our read
Name the trade-off before validating a model. Otherwise, a metric can report performance without resolving which error matters more.
McNeal expects constrained tasks to improve iteratively
McNeal said she is not in the AGI camp, while affirming that people are building intelligent systems. She said she had seen a lot of attempted decision-making across processes struggle, except where the feature space stays constrained.
Her criticism is narrower than saying AI cannot work. Yet broad decision-making remains difficult in her experience, while finite paths and opportunities are more tractable.
Her proposed path is iterative, not instant. Start building, capture relevant information and the human-AI exchange, then create a data asset the system can learn from. That describes learning inputs, not a specified staffing model or measured reduction in review work.
Diogo Almeida of TypeSafe said on The TWIML AI Podcast in October 2026 that Waymo's decomposed engineering suits current model reliability limits. He emphasized understanding and debugging components, which relates to diagnosability, not McNeal's exchange-capture sequence.
“so you're not going to have the perfect automated system the perfect agent out of the get-go from the beginning”
Our read
Start with a constrained task and measure what the captured exchanges teach the system. Do not treat an iterative data process as proof of lower review costs.
Thread sells control infrastructure, not outcome evidence
Thread provides infrastructure for agents and workflows in high-impact settings. McNeal calls its thesis controlled autonomy, with limits and human intervention. Agents may reach out to people, or policy may insert them.
She said the process is captured in a record. She also distinguished governing individual actions from governing the whole transaction. Those controls are related, but a record and a governance boundary are not interchangeable.
Jake Stauch of Serval said on No Priors in October 2026 that agents should request specific tools and workflows, then receive approval under guardrails. That comparison concerns permissioning, not proof of Thread's performance.
McNeal said Thread was doing some research with, like, Army DevCom, recently discussed publicly. She described it as around, like, manual deconfliction and helping scientists and warfighters get accurate information.
The interview explains the product thesis and an intended setting, but reports no comparative outcome measures. That makes effectiveness an open question, not a reason to assume the controls fail.
Our read
The product proposition is specific, but evidence of outcomes remains unreported. Buyers need deployment measures to judge how its controls work in practice.
What it means for you
- Choose the tolerated error. For one workflow, write down the costs of false acceptance and false rejection before tuning a model. Document how those priorities enter the system logic.
- Log the human exchange. Capture the decision, reason, and outcome for each approval or intervention where policy allows. Measure review time and whether similar unresolved cases recur.
- Exercise a provider change. Test a workflow after changing its coding provider, team ownership, or load. Track failures and recovery rather than assuming components will keep working.
- Compare routing options. For a fixed task, compare model options and a deterministic path against the business's cost, latency, and error priorities. Keep the task and acceptance criteria fixed.
What to watch
- Whether Thread reports measured results from its Army DevCom research, if it chooses to publish them.
- Whether deployed workflows show lower error rates without overwhelming review queues or losing useful transaction records.
Send this to product and risk leaders setting AI workflow policy.
Watch the minute McNeal gives her hedged reconciliation example
Written by Real Alpha from the episode's transcript and the other moments it links, then checked against the transcript. Quotes are the speaker's exact words.
Minute by minute
- 0:49Thread builds controlled autonomy
McNeal says Thread AI provides infrastructure for agents and workflows in settings where a wrong action or system failure can have serious consequences. Its platform lets organizations limit what systems can do, add human intervention points and record the whole transaction, rather than govern only individual actions.
- 3:21Production systems need operability
McNeal says she moved from an FDE role at Palantir to product work, while cofounder Maya worked as a software engineer on the product side. She says companies were buying separate tools for specific business functions or assembling their own systems, which could be hard to operate as providers changed, teams turned over or workloads grew.
- 7:13Routing depends on decision logic
McNeal says the difficult part of model routing is defining the business logic: which models to choose and what to optimize. Teams must specify what counts as a hard problem or an expensive one, including whether the cost is money, time or people’s time. Thread lets customers build routers and decision gates into its platform.
- 10:13Defense work raises the stakes
McNeal says Thread was designed as a dual-use company and describes research with Army DevCom on manual deconfliction, helping scientists and warfighters get accurate information at the right time. She contrasts this with financial reconciliation, where mistakes can cost a business billions of dollars and create legal risk, and says defense errors can be life or death.
- 13:22People design the trade-offs
McNeal says engineers at large institutions are shifting toward managing, reviewing and designing systems rather than implementing them. She uses a clinical-trial model to show how people must weigh the risk of admitting a bad-fit candidate against the risk of excluding someone who could qualify. She says AI does not remove the legal, regulatory and human factors that shaped existing processes.
- 18:35Workflow design is the hard part
McNeal says Thread’s building blocks modernize established distributed-systems security and software practices rather than discard them. She sees workflow design—such as deciding who receives triage alerts and how much alert volume is useful—as harder than assembling the blocks. Thread’s platform is self-service, with an applied team that works with customers; she says spending on consultants to assemble AI stacks points to a gap a platform can fill.
- 20:57Enterprise APIs need infrastructure
McNeal says Meta’s Toolformer paper showed language models using simple APIs, but enterprise systems involve thousands of software providers, APIs, parameters, protocols and authentication schemes. She says the idea of flexible decision-making in software inspired Thread’s vision for horizontal infrastructure. Her problem-solving approach combines scientific method and hypothesis testing with empathy for customer experts, lessons she connects to her upbringing, studies and Palantir experience.
- 25:27Full automation fails in open-ended work
McNeal says decision-making is difficult to automate across a whole process unless the possible paths are tightly constrained. She argues that AI workflows should start iteratively: use building blocks, capture the human-AI exchange and build data the system can learn from. She cautions against expecting to automate every decision from the outset.
Published Nov 8, 2026.
Transcript · 21 min read
MTS, hosted by Theo Jaffee and Sophia Dew, with Angela McNeal.
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
I believe in like intelligent systems and that we're building intelligent systems and using intelligent systems. But I think a lot of the decision making that I've seen attempted to be automated across the stack, across the process, be kind of unsuccessful, except when the feature space is so constrained. But when you like open up the aperture, I've seen things struggle still. Hello, welcome back. I'm here with Angela McNeal, who is the CEO and co-founder of Thread AI, which builds infrastructure for enterprise AI workflows. She also previously spent six years at Palantir and writes the Handshake Protocol about how people and AI systems work together.
So welcome, Angela. Great to have you on. It's a pleasure to be here. Thanks for having me. Yeah, so we were talking about a few things before we went live, but I guess, first of all, wanted to hear about Thread AI and how you guys are thinking differently about making progress with AI. So at Thread, we provide the infrastructure that AI agents and AI workflows can run on when an action or a wrong action or a system failure or a bad decision can have critical impact. So financial risk, reputational risk, legal risk. So we built our platform on the thesis of controlled autonomy. So how do my agents and workflows reason and act in a system that you control, that you limit, where you can inject humans at any point in time of that process.
Either the agents will reach out to them or you've injected in by policy and then everything is over, is captured in a record. So a lot of companies govern individual actions, we govern sort of the whole transaction. Sure. Okay, okay. So what are some of the, like, failure modes or some of the problems that you were seeing that you wanted to address? Yeah. So we saw a couple of different patterns of companies sort of rebuilding and reassembling stacks when they're thinking about putting AI into production. Maya and I talk about this a little bit about how we were, you know, my co-founder, we were inspired to start the company when we, you know, read this interesting paper. But when we saw, okay, like there's this bottlenecks we're shifting from building models and training models to, okay, you have models, like now what do you do with them? And when you look at an enterprise and when you look at the real world, you're like, okay, a real enterprise has, you know, tens, hundreds of different softwares. You have different identity providers. You have different kinds of data you have different kinds of risks and rules and regulations um and we saw people sort of again assembling and reassembling all these different building blocks and pieces they needed to take
Read the whole transcript
these models and put them in meaningful useful ways and wasting a lot of time um so we're like okay well how can we provide um sort of the building blocks that are necessary so that you don't have to rebuild isolation and metering and identity and logging and governance and you can and just focus on, you know, what you're trying to do. Right. So I'm curious, did your time previously at Palantir as an FDE, is that when you started thinking about, like, how do we put these things into production or, like, into real-world environments as opposed to just purely having the technology? Yeah, Maya and I joke a little bit.
So my co-founder and our CTO, she and I were at Palantir together for four years. But, so I joined as an FDE and then switched to the product side after about a year, I spent the rest of my time there. Maya joined exclusively as a software engineer on the product side. And it's funny. We joked that we were spoiled by solving hard problems. Right, okay. Because of the types of things, the kinds of, again, problems that we were exposed to and that we were working with. When we started our company together, that sort of same ethos, those tenets of, okay, how do we provide the right tools to solve critical problems and power mission-critical workflows and agents?
That was in our DNA.
“this is just going to be compounded when builder ecosystems grow, model providers expand and grow.”
“the actual hard part, the technology behind a router I don't think is as interesting as like the application”
weekend too? With Lovable, you get software you can ship today. That includes an editable codebase
you can inspect and change, plus two-way GitHub sync. Lovable also handles the backend and infrastructure, including managed Postgres, auth, storage, hosting, payments, and more, so you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable.dev. Now onto the episode. So I think like, you know, the topic of agent
“a lot of that work is around, like, manual deconfliction, right? So how do you help people get accurate information at the right time”
So we're growing, sort of, quite a bit. We have a bunch of open roles around some of our public sector work in particular. So you'll be seeing more from us on that element. And then one of the things I'm pretty excited about is that sort of the investments that we've been making around our building blocks.
So for companies who are selling products or building products that they're selling to their customers and want those products to be more adaptable and agentic and AI forward, like we've been providing sort of building blocks for companies building products. So keep an eye on some of the stuff we're putting out there. Yeah, awesome.
“you as a business person with the scientists have to decide what's more important”
sponsors 11 labs ai that communicates at human level across every channel and modality 11 labs dot io slash mts scale your startup on neon millions of developers and startups have already chosen Neon for their backend. Start on the free plan or get up to $100,000 in credits for your
startup at neon.com slash MTS. Special thanks to our sponsor Kong, the AI connectivity platform. Connect APIs, LLMs, agents, and systems with serious security and governance. KongHQ.com. Finally prove what your AI investment is worth with Lairdyn. Head to Lairdyn.com today to start measuring your AI's bottom line. MongoDB, trusted to run critical applications by 75% of the Fortune 100, the intelligent data platform for the era. Apply today for access to up to $15,000 in Atlas credits, dedicated technical support, and exclusive events and go-to-market opportunities. MongoDB.com. Chargebee, billing and monetization infrastructure for the AI era. Chargebee.com.
I wonder when that will like the dust will begin to settle there yeah um yeah so I'm curious also is there anything that you find like what is the hardest to automate or even with the building blocks that you were mentioning like um what is difficult about designing building blocks you kind of have to create something that is that will work across many different like teams or whatnot I'd say that is mine and Maya's superpower in that, like, our primitives that we built sort of from the beginning have been able to adapt to all of the sort of technological changes in the past sort of four years.
Um, uh, and, um, we were very explicit from the beginning, um, to not throw away, right, like 20 some years, 50 years of distributed system security, um, right, software best practices, but to kind of like modernize and make them more adaptable. also learn from best practices. And I think the hard part is often the piece, again, that comes after of like, you know, you can build a triaging workflow, right? But then you as the builder of this workflow, how do you decide who gets these alerts, the volume of those alerts? Are they too noisy? Are they not as important? So the actual workflow design, I think, is the more...
challenging part. Yeah. Um, do you find that you can transfer these across industries as well? Um, we, so we built our platform to be self-service. Um, so our customers are a lot of our builders, but we do have like our applied arm too, we'll work with our customers and we'll share best practices and knowledge and sort of construct with them. Um, I was thinking about, I saw there was that right enterprise AI spend is raw allegedly rocketing or booming. Um, I think what we're seeing probably a big piece of that is there's this been huge investment we're seeing in like consulting services where 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 like a gap for a platform to fill yeah and that but but people do you know want help Yeah. So tell us more about the platform and how you went about designing it. Did you design this from the get-go? Was the vision always, we're going to do a platform?
Yeah. We built horizontal infrastructure past life. Wanted to build it now. And it's funny. I think Maya and I talk about this, I think we've talked about this publicly, of the thing that really inspired us was this paper that came out of Meta. Sure. And it was the Toolformer paper where they were like, okay, we have a language model and we're going to train the last couple layers to speak simple API. And there it's like, I think one was like a calculator or like weather or whatever. And if you're a software engineer and you look at that, you're like, well, that's not how an enterprise, like enterprise APIs don't work that way, right?
There are thousands of software providers at an enterprise. Each of these providers might have thousands of APIs with hundreds of different parameters. You have different protocols and different authentication policies and schemes. You have different data types. The ecosystem is so complicated. But the idea that, like, cool, I could put flexible decision-making in software was pretty exciting. And so we're like, okay, how do we take that? And then you play it out N years. You're like, what could it look like best case if you have these again sort of intelligent systems in and across a company you're like well to do that and to make that a reality working backwards from that you're going to need a lot of infrastructure yeah and so that was the vision from the beginning yeah how do how easily do you find it you personally to like extrapolate all these trends when it comes to like enterprises and their use of technology i think that um if you are building in and as like as a product person if you are building in this space and leading a company in this space that is like table stakes that is a skill, a superpower that you have to have if you're selling also to horizontal so across these different patterns
and that also too is the fun part of it the pattern matching I get to see 10 different kinds of workflows a day across different companies and again the fun part is you have all this information how do you pattern match and how do you pull out the interesting pieces and right like fun product work yeah yeah um i'm also curious too because you seem to think about logic a lot and and so like what was your um we talked briefly too about like your background in math computer science but like what is the lens that you view problems in and maybe like to at palantir what was the lens that they view problems and how did that all tie together um i think you know i've i have i grew up You know, I'm the child of two climate scientists, grew up in a household where you're always deconstructing and talking about things at dinner tables.
So there's a lot of the scientific method being applied sort of throughout my life. And then spent, I was at Columbia where I studied computer science and math. And so there is an engineering element to science. There's a scientific element to how I approach problems. So how do you deconstruct problems? How do you determine what the base case is? And how do you hypothesis test? So there's that element to that. But I think there has to be, we don't live in a vacuum. We don't live in like a simulation with like a topical hot take. Yes. With, right, perfect actors applying perfect behavior, right? We live in a deeply messy and human world. And so the way that you approach problems, you have to approach them with empathy and understanding and try to like solve the problem. That's where, again, the scientific method comes of like approaching problems, being data driven, forming hypotheses, like actually challenging those hypotheses but the big picture has to be like so deep yeah and i think that was a great lesson i learned as an fde where and i think that is one of the most important things for me as a founder like you need to have um a balance of like confidence and conviction and humility and and i think that as again you learn that as an fde of like you are confident in
yourself and the things and you're helping this customer and you're solving problems together but you need to be you are talking they're the expert in whatever the thing is not you and so how do you work with them to solve problems together is um i think if you're a good fde that is the
“AI workflows are always iterative so you're not going to have the perfect automated system”
Yeah. Well, I think that's a great place to wrap. Thanks so much, Angela, for talking to us about Thread AI.
Thanks for having me. Great to have you. Pleasure. A huge thanks to MTS sponsors. Autonomous Software Development for Enterprise Codebases Ship 5X faster. Blitzy.com. AdQuick. Make your brand a billboard. Out-of-home advertising as easy to scale as digital. AdQuick.com. Arena. Measuring AI performance in the real world. Arena.ai. Support for the show comes from VCX, the public ticker for private tech. The U.S. stock market started history's greatest wave of wealth creation. From factory workers in Detroit to farmers in Omaha, anyone could own a piece of the great American companies. But today, our most innovative companies are staying private longer, which means everyday Americans are missing out.
Until now. Introducing VCX, a public ticker for private tech. Visit getvcx.com for more info. That's getvcx.com. Carefully consider the investment material before investing, including objectives, risk, charges, and expenses. This and other information can be found in the Innovations Fund prospectus at getvcx.com. This is a paid sponsorship. Thank you.
