The Palo Alto company building the control plane for AI agents. They watch the models nobody else is watching - and the enterprises buying them are starting to insist on it.
Walk into a Fortune 500 deployment review in 2026 and you will hear the same nervous question. The model is live. The agent is taking actions. Who is watching it? More often than not, the answer involves a Palo Alto company with eighty-three vowels in its tagline and a quiet conviction that AI should be auditable by default. That company is Fiddler.
Fiddler does not build the model. It does not write the prompt. It does not ship the agent that emails your customers at 2 a.m. It does something less glamorous and considerably more useful - it sits next to all of that and keeps the receipts.
The category they sit in has a slightly clinical name: AI observability. The job is simpler than the name suggests. Watch the model. Catch the drift. Block the jailbreak. Explain the decision when an auditor asks. Do it in real time, at enterprise scale, across thousands of models and increasingly across agents that chain calls together like a tipsy improv troupe.
Here is the inconvenient truth the AI industry would rather not print on a coffee mug. Almost every model deployed in production is, to some degree, a black box. It works. Until it doesn't. And when it doesn't, the people who deployed it cannot always tell you why.
This is fine when the stakes are small - a slightly worse movie recommendation, a clumsy autocomplete. It becomes considerably less fine when the model decides who gets a loan, which radar return is a missile, or which support ticket gets escalated. The model still works. Until it doesn't. The cost of not knowing why goes up sharply.
Krishna Gade had seen this movie before. He spent years building Facebook's News Feed infrastructure - which is to say, he spent years watching a ranking model influence what billions of people read for breakfast. The lesson was not subtle. Models in production are not science experiments. They are infrastructure. And infrastructure needs monitoring.
In 2018, "explainable AI" was a research paper title, not a budget line. Gade, joined by co-founders Amit Paka and Manoj Cheenath, started Fiddler Labs anyway. The thesis was straightforward and only slightly contrarian - enterprises will eventually be forced to operate AI the way they operate the rest of their stack. With telemetry. With dashboards. With on-call rotations. With audit logs that survive a regulator's questioning.
The first investors were Lightspeed and Lux Capital. The first customers were data science teams in banks and ad platforms who had quietly discovered that their models were drifting and they did not have a good word for it yet. Fiddler gave them the word, and then sold them the dashboard.
"Build trust into AI."
- Fiddler's tagline. Five words. Surprisingly hard.Founded in Palo Alto. Seed round from Lightspeed and Lux.
$10.2M Series A. ML monitoring platform goes GA.
$32M Series B led by Insight Partners.
Mozilla Ventures invests - trustworthy AI thesis bet.
Fiddler Guardrails ships. LLM observability gets teeth.
$30M Series C. Control Plane for AI Agents launches.
Fiddler's platform has gotten broader as AI has gotten weirder. What started as model monitoring for tabular ML now reaches into LLM applications and the multi-step agents that have started doing real work inside enterprises. The interface stays roughly the same. The blast radius does not.
Drift, performance, bias, and root cause - in real time, across ML and LLM workloads.
Low-latency safety checks for LLM apps. Hallucination, jailbreak, and PII screening at request time.
End-to-end visibility for compound systems. Tool calls, trajectories, policy enforcement.
Continuous evaluation and audit trails for governance, risk, and compliance teams.
The customer list is the kind of list that makes a board meeting easier. Nielsen runs unified observability across agents and predictive models on Fiddler. Integral Ad Science uses the platform across its measurement stack. The U.S. Navy uses it - which is a sentence that, on its own, tells you the platform has cleared bars most startups do not bother applying for.
Source: company press releases, Crunchbase. Bars sized to round amount. Total disclosed: north of $100M.
The investor list rhymes with the customer list - careful, mostly enterprise-savvy, occasionally surprising. Lightspeed and Lux were there from the start. Insight Partners came in for Series B. Mozilla Ventures wrote one of its earliest cheques here, on the explicit thesis that the world needs an infrastructure layer for trustworthy AI. In January 2026, RPS Ventures led a $30M Series C with most existing investors plus LG, Benhamou, and LDV joining. Total funding crossed the $100M mark - quietly, the way most things at Fiddler happen.
The tagline is short enough to fit on a hoodie. The work behind it is not. Fiddler's argument, made for years before it was fashionable, is that AI cannot scale into the parts of the economy where it matters most without a layer of accountability underneath it. Healthcare. Finance. Defense. Government. The places where a wrong answer is not just embarrassing but expensive, regulated, and occasionally dangerous.
The company has been remarkably consistent about this. The product has shifted - from explainable ML in 2019, to model monitoring at scale in 2021, to LLM observability in 2024, to a full control plane for agents in 2026. The mission has not. The mission is that production AI should be observable, explainable, and accountable. Everything else is implementation detail.
The agents are coming faster than the policies are. Every CIO in a regulated industry is now negotiating with that fact in real time. They want to ship. Their legal team wants to sleep. Their auditors want logs. Their CISO wants a kill switch. Their CFO wants the model not to refund the wrong customer twice. Somebody has to sit between all of those people and the model. Fiddler is making a serious case that the somebody is them.
If they are right, AI observability stops being a niche line item and becomes what monitoring became for cloud infrastructure - boring, mandatory, and a sign that your company is taking itself seriously. If they are wrong, they will at least have built the most thorough dashboard for it. That is not a bad consolation prize.
So return to that deployment review. The model is live. The agent is taking actions. The same nervous question gets asked. The difference now is that somebody in the room has an answer, and the answer involves a dashboard, a guardrail, a policy log, and a vendor in Palo Alto that has been quietly arguing for almost a decade that this was always going to be the boring, important part.
Fiddler still does not build the model. It still does not write the prompt. It still does not ship the agent. It keeps the receipts. In an industry that has spent two years sprinting ahead of its own paperwork, that turns out to be the part the grown-ups want.