Harmony's $34M seed: the context graph is the product

Harmony raised a $34 million seed to point AI agents at the least glamorous work in a company: password resets, laptop requests, software access, onboarding checklists. Led by Lightspeed, announced July 28. What makes it worth your attention isn't the round size — it's what they built to make the agents work, and the fact that the same thing is buildable at your scale.

What actually happened

Per the funding announcement, the round included Hitachi Ventures, Fin Capital, Mercer Ventures, and Operator Partners, plus angels including Wiz co-founder Assaf Rappaport and Eon.io CEO Ofir Ehrlich. Founders Nitzan Shapira and Ran Ribenzaft previously built Epsagon, the cloud observability company Cisco acquired for roughly $500 million.

The product lives inside Slack and Microsoft Teams and handles requests across IT, HR, finance, procurement, and legal. Harmony reports 100+ prebuilt agents, a 70% no-touch resolution rate, and customer deflection of 48% within two weeks of deployment, rising past 75% at three months. Named customers include n8n, eToro, Cyera, and KITH.

The number that actually explains the system is buried in the breakdown: deflection by function runs HR 68%, procurement 58%, DevOps 47%, security 42%, finance 36%. Those aren't different models. They're different amounts of structured context.

Why internal support agents matter for your business

Underneath the agents is what Harmony calls a context graph — every employee joined to their identity, devices, applications, and work history. That's the whole trick. An agent that knows who is asking, what they already have access to, what device they're on, and what they did last week can resolve a request. An agent that only knows the text of the message can write a nice paragraph about resolving it.

Read the deflection spread again with that lens. HR wins because the answer lives in a policy doc and an HRIS record. Finance lags because the answer depends on a purchase order, an approval chain, and a vendor contract that lives in three systems and one person's head. The bottleneck was never the model.

So the sequence for a small operator is the inverse of how most teams do it. Don't start by picking an AI vendor. Start by picking your two highest-volume internal requests, and write down what a person has to look at to answer them. If that list is "one system," you can automate it this month. If it's "four systems and Dave," you have an integration problem wearing an AI costume — and the honest fix is connecting the systems first.

The agents are the cheap part. The graph is the work, and it's yours to own.

Sources: ACCESS Newswire, Business Insider via Techmeme.