How do organizations test whether their AI works?

“Any incidents are now an eval. The team collected past PRs that caused incidents, with their fix-forwards and reverts, and had engineers who know that code confirm each bug and where it lived. The reviewer runs against that set daily.”

AnthropicPublished by the organization itself (post), no named author

“What we did was we built a system that actually has agents constantly running and creating BAML programs. … We then look at the entire … transcripts, see what tools they use, see what happened, and obviously we as humans can inspect them. But more importantly, we can have agents go inspect them. And agents find what was good, what was bad. … And we can have humans collaborate with these issues to figure out which ones are real, which ones are hallucinations, which ones don't have taste as much as I hate to use that word. And then we can have agents go ahead and create fixes to these problems and go to address them.”

Vaibhav GuptaCo-founder and Chief Executive · Executive, Boundary (BoundaryML)

“And that's why we have Cursor Bench. We have this private eval set that is mostly made up of things that happen in our code base, which is held out from the … So we ensure that the models aren't trained on it. And it's based on those real world engineering tasks.”

Lee RobinsonMachine Learning Engineer, Model Behavior, Cursor · Practitioner, Cursor (Anysphere)

“Regression test. Those live in ci, those have to pass a certain percent, you know, within some stochastic error rate. Then we have, as you’re building a product, evals of these aren’t passing right now, and this is launch quality. So we have a report card and we need to, on these categories, you know, be it 80 or 90% of all of these user journeys to launch, and then what we have what we call frontier or headroom evals, where we actively wanna be at 30% pass rate.”

Sarah SachsManages Notion's core AI capabilities and infrastructure team · Leader, Notion

“what can I actually compartmentalize in a way that we can understand and evaluate? Because that's what I'm hill climbing against, right? Otherwise, I'm just tuning these things.”

Frank LongWorks at Coatue · Practitioner, Coatue

“We had a GTM engineer, as I mentioned, shadow one of our best SDRs for a few days. … Watching every single tab she opened, … The engineer converted each of those data sources into a step in a tool calling workflow. … After building our first version, … ran in production in shadow mode for six weeks. Our best SDR reviewed every single output to create a feedback flywheel. And then finally, she couldn't improve it anymore. The architecture reflects exactly what she was doing manually. But now, it performs like the 90th percentile rep, 100% of the time. We applied the same framework to 30 different SDR workflows, … SDR quotas increased by 30% this quarter.”

Jeanne DeWitt GrosserChief Operating Officer, Vercel · Executive, Vercel

“Fine-tuning Gemma4-E2B on hundreds of thousands of domain-specific examples lifts its F1 from 0.563 to 0.902, on par with the best proprietary models, while median latency stays at 35ms.”

AirbnbPublished by the organization itself (post), no named author

“We built a benchmark of 200 credit analyst-reviewed income statements and balance sheets, with the analyst's figures recorded for every metric and time period.”

RampPublished by the organization itself (post), no named author

“In a later randomized split, Opus 5.5 at medium showed no measurable change in review outcomes or critical findings compared with Opus 5, so we switched.”

Ryan SkidmoreAuthor of Cloudflare's "Orchestrating AI Code Review at scale" blog post · Practitioner, Cloudflare

“Before switching all reviews over from v1 to v2, we ran every review in every canary repo on both v1 and v2 in a comparative shadow mode. The two agreed on the approval decision in 98% of the merge requests we tested, so we moved everyone over in one go in early October and removed the v1 code.”

Ryan SkidmoreAuthor of Cloudflare's "Orchestrating AI Code Review at scale" blog post · Practitioner, Cloudflare

“we saw that against like a very naive implementation, our costs were down by 60%, and our quality and our accuracy was up by around 70%.”

Ameya KetkarSpeaker, AI Engineer World's Fair 2026 (job title not stated) · Practitioner, Uber

“once you thought you were going to save time by writing actually your context instead of all your code, but if you take this rigorously, you're going to spend time … writing the right evals. And that's kind of like a lot of work … because now you don't only have one prompt that you're trying to get right, it's like all the prompts of the evals.”

Patrick DeboisProduct DevRel Lead · Practitioner, Tessl

“Think about this, like, you run it five times, and out of five, how many times does it succeed? And … maybe in several cases it hits 100% all the time, which is great, but in others not. And depending on how you change your context, it will influence which tests actually work or not. I find it personally helpful to think about this as error budgets, I give a set of tests … that I really care about, so it's only allowed … to fail minimally, and other pieces are okay. So that's how you have to think about testing context, you cannot do, like, exact testing all the time.”

Patrick DeboisProduct DevRel Lead · Practitioner, Tessl

“we can't look at any traces basically. We can't look at any data because it's all super confidential. And that's, like, extremely frustrating to work with because you you basically don't know what people are doing.”

Jacob LauritzenCTO · Executive, Legora

“So we have different, basically, like, worlds, like RL environments, essentially, which is like, here's a… Because because lawyers work in projects always, more or less, so it's like, here's 10,000 file data room, you're doing your due diligence”

Jacob LauritzenCTO · Executive, Legora

“we’ll all sit down every Friday and look at the P 99 of like the most token exhaustive custom agent transcript and just look at why it didn’t do well and cut a bunch of tasks.”

Sarah SachsManages Notion's core AI capabilities and infrastructure team · Leader, Notion

“Notions last exam is a big thing inside the company and we have people, full-time staff to it exclusively. Mm. We have a data scientist, a model behavior engineer, and an full-time, um, evals engineer just dedicated to the evals that we pass 30% of the time.”

Sarah SachsManages Notion's core AI capabilities and infrastructure team · Leader, Notion

“And so we were like, okay, let's handle that part of it and do kind of like the last mile work of taking stuff from your local machine and putting it somewhere. And what we saw was that people were using Sites for sites where yes, building websites, but then also building plugins and hosting data and running evals because it does. It is very powerful.”

Kath KorevecProduct lead for Sites, OpenAI · Leader, OpenAI

“Early on we realized that none of the existing tracing solutions are 100% matching what we want to see and code is cheap, so we’ve build a custom admin panel that evolved in a full solution.”

Sam KronickAuthor of Ramp's "Building Laminar" engineering blog post · Practitioner, Ramp