A 95% Accurate AI Model, Chained Twenty Steps Deep, Gets the Job Right 30% of the Time

Sasha Haco spent her PhD at Cambridge staring at one equation — the mathematics of what happens to something as it falls into a black hole. Four years, one equation. It is, by her own account, a useful way to learn how niche and narrow deep technical expertise actually is, even in a field that sounds impossibly broad from the outside.

She founded Unitary in 2019, wanted the domain unary.com, found it taken, and settled for unary.ai instead — a decision that at the time felt like a liability. Calling yourself an AI company in 2019 invited skepticism. Six years later, as she put it, everyone wants “.ai” in their name whether or not it’s earned.

Unitary didn’t start in insurance. It started in content moderation — building automation for dating apps and social platforms that had to review flagged text, images, and video against platform policy and take action. That is a genuinely hard AI problem: assessing whether specific content violates a specific policy requires real judgment, not pattern matching. But Sasha and her team noticed the product they’d built, which they call virtual agents, generalized well beyond content moderation to almost any repetitive task an entry-level employee might do. About a year before this conversation, they went looking for an industry with heavy manual process spread across a lot of disconnected legacy systems. Insurance fit immediately.

In Episode 144 of InsurTechTalk, Sasha and I covered why the insurance industry’s definition of “agentic AI” is mostly wrong, why chaining AI reasoning steps together is a math problem before it’s a technology problem, and why she thinks companies attempting three-year AI transformation projects are already the losers.

About Sasha Haco

Sasha Haco is the Founder and CEO of Unitary (unitary.ai), a company that automates manual back-office work — submissions intake, bordereaux processing, loss run requests, and similar operations — through what the company calls virtual agents: AI-and-software hybrids that log into a customer’s existing tools, including legacy systems, and mimic human workflows without requiring API integration or engineering effort. Sasha holds a PhD in physics from the University of Cambridge, where she also completed her undergraduate degree, graduating first in her class at Queens’ College. Unitary began in content moderation for social platforms and marketplaces before moving into insurance operations roughly a year before this conversation.

What “Agentic AI” Actually Means — and Why the Industry Is Using It Wrong

Sasha didn’t hedge on this: the term agentic AI has a specific technical meaning, and the insurance industry has largely stopped using it that way.

• True agentic AI refers to a specific class of models capable of multi-step reasoning and exploratory decision-making — genuinely figuring out an approach, not just executing one
• In insurance, “agentic” has come to mean almost anything that automates work that used to require a human — regardless of whether any actual AI reasoning is involved
• That conflation matters because it obscures a real architectural decision every automation project has to make: which steps genuinely need a reasoning model, and which just need reliable software

The Compounding-Error Problem

This is the core technical argument Sasha makes for why most complex insurance workflows shouldn’t run entirely on AI, even where AI is genuinely useful for parts of them.

• AI models are probabilistic — even a very good model might be 95%, maybe a bit more, accurate on a given step, but not 100%
• Apply 95% accuracy independently across a 20-step process — like a full submissions intake workflow, from receiving a broker’s email to keying structured data into an internal system — and the probability of getting the entire process right compounds down to somewhere around 30%
• The fix isn’t more AI, it’s less: use deterministic software for the steps that have one correct, mechanical answer — copy this field to that field — and reserve AI specifically for the steps that require genuine interpretation or judgment
• Her framing: most workflows need a blend, not a single tool applied uniformly across every step

Virtual Agents vs. RPA

Sasha draws a sharp distinction between what Unitary builds and traditional robotic process automation, a category insurance ops teams already know well.

• RPA is brittle — if a system’s interface changes, or the process touches a system the customer doesn’t fully control, the automation breaks and has to be rebuilt by engineers
• RPA also cannot handle steps that require actual reasoning or interpretation, which is exactly where a meaningful share of insurance operations work lives
• Unitary’s virtual agents combine the robustness needed to survive system changes with an AI reasoning layer for the judgment-based steps RPA simply cannot do
• Deployment requires no API or engineering integration — a virtual agent logs in with user credentials the same way a human would, and can be operational within a couple of weeks

An Accuracy Guarantee Benchmarked Against Humans, Not Perfection

Rather than promising a fixed accuracy number, Unitary commits contractually to match or beat whatever accuracy the customer’s own human team is currently achieving on that task.

• Measuring that baseline first is itself revealing — Sasha has seen human team accuracy range from roughly 70% to 99% depending on the task, with more tedious, repetitive tasks trending toward the lower end
• Humans-in-the-loop supervise the virtual agents early on, and every human correction feeds back into training and improving the automation
• Within roughly two months, the large majority of decisions on a given process typically become fully automated
• Her point to skeptical buyers: demanding 100% accuracy from AI while accepting far less than that from human teams is an inconsistent bar

What Makes a Company an Actual AI Company

Sasha’s test for this is simple and, she argues, mostly missing from how the market talks about itself: is AI a feature bolted onto an existing product, or is it the product?

• Established software tools adding an AI assistant or agent feature — she cites Salesforce’s Agentforce as an example — are still fundamentally what they were before; a CRM with an AI feature is still a CRM
• A genuine AI company is one where the AI is core to what’s being sold, not an add-on layered onto an existing offering
• She’s sympathetic to why everyone wants the AI label right now, but argues it’s a distraction from the actual question buyers should ask: is this the best tool for this specific job, or just the most fashionable one

Adoption: Risk-Averse, But Newly Willing to Try

Insurance’s traditional risk aversion is running headlong into genuine FOMO around AI, and Sasha sees that tension playing out differently across geographies.

• Insurers that would never have taken a meeting with an early-stage automation startup a few years ago are now actively running proofs of concept
• The US market is generally less risk-averse than the London market
• London’s density works in the opposite direction from what you’d expect — because so much of the industry sits within the same square mile, peer awareness and “what is my neighbor doing” pressure spreads adoption faster once it starts, even in a risk-averse market
• Trust ultimately comes down to security fundamentals: SOC 2 and ISO 27001 certification, and firm guardrails ensuring customer data isn’t used to train models for other companies or leaking into consumer-facing tools

Why Multi-Year Transformation Projects Are Already Losing

Sasha’s sharpest prediction was about which companies will be worse off in three years, not better.

• Large-scale, multi-year transformation projects — replacing core systems wholesale, re-architecting everything at once — are, in her view, likely to still be unfinished three years from now
• The winners will be the companies taking small, incremental bites: piloting automation on one process, measuring results, and expanding from proof points rather than attempting to transform everything simultaneously
• Longer term, she expects dramatically smaller operating teams across the industry — gesturing, only half-jokingly, toward a future of near one-person MGAs, carriers, and brokerages orchestrating a stack of AI tools rather than managing large back-office headcount

Key Takeaways

• The insurance industry’s casual use of “agentic AI” mostly describes automated work, not genuine multi-step AI reasoning — the distinction matters when deciding what tool a workflow actually needs
• Chaining AI reasoning steps across a long process compounds error mathematically; a 95%-accurate model applied to every step of a 20-step workflow can produce a process that’s only right about 30% of the time
• The right architecture for most insurance automation is a deliberate blend — deterministic software for mechanical steps, AI specifically for steps requiring judgment
• Benchmarking AI accuracy against a company’s actual current human performance, not an abstract 100%, is a more honest and more useful standard
• Whether a company is a genuine AI company depends on whether AI is the core product or a feature added to something else
• Large multi-year transformation projects are a likely tell for which companies fall behind; incremental, proof-of-concept-driven adoption is outperforming it