An $85 million seed round for a company that is four months old is not a sign that funding discipline has collapsed. It is a sign that the people writing checks have quietly stopped believing the cheap version of the agent story — the one where you wrap a frontier model in a prompt, point it at an enterprise, and watch the work happen.
Hang Ten Systems, founded by former Infosys CEO Vishal Sikka, just added $53 million to its seed, bringing the total to $85 million. Temasek’s Xora led, with Mayfield and Aramco Ventures joining. The stated focus is enterprise AI services. The round closed weeks after the first one. That is roughly all we know, and the mainstream reading of it is already circulating: famous services executive returns to disrupt the services industry he used to run.
I think that reading is backwards, and the reason has to do with where agent systems actually break.
Enterprise services is an architecture problem wearing a business-model costume
The usual pitch for AI in IT services is substitution. Thousands of engineers do repetitive integration, migration, testing, and support work; models do some fraction of it; margins improve. That framing treats the work as text generation with extra steps.
Anyone who has instrumented an agent inside a real enterprise knows the failure modes are not linguistic. They are structural. An agent asked to modify a claims-processing pipeline does not fail because it cannot write Java. It fails because it cannot see the seventeen upstream systems that feed the pipeline, cannot tell which of four conflicting schema definitions is authoritative, cannot distinguish a deprecated internal API from a live one, and has no mechanism for knowing that its change will violate a compliance constraint encoded in a 2011 PDF nobody has read since.
None of that is fixed by a better model. It is fixed by state management, provenance, verification, and a memory layer that reflects how a specific organization actually works rather than how organizations work in general.
What $85 million actually buys
This is why the size of the round is more interesting than the founder’s rĂ©sumĂ©. If you are building a thin orchestration layer over someone else’s model, you do not need $85 million in seed capital. You need a few million and a design partner.
Capital at this scale, this early, tends to be consumed by a small set of expensive things:
- Evaluation infrastructure. Measuring whether an agent did enterprise work correctly is harder than doing the work. Correctness is contextual, delayed, and often only visible in production. Building that measurement apparatus is unglamorous and slow.
- Ingestion and grounding. Turning a client’s tickets, runbooks, repos, and tribal knowledge into something an agent can reason over reliably is engineering-heavy and does not generalize cleanly across customers.
- Senior talent. People who have shipped systems inside regulated enterprises are scarce and expensive, and they are the ones who know which constraints are real.
- Long sales and deployment cycles. Enterprise trust is bought with time, and time costs money before revenue arrives.
I want to be precise here: I have no visibility into Hang Ten’s internal architecture. Nothing has been disclosed publicly beyond the funding and a broad focus on enterprise AI services. What I am offering is a read on what problems anyone raising this much, this early, in this category is almost certainly signing up to solve.
The investor mix is a hint
Xora sits inside Temasek and tends toward deep-tech, longer-horizon bets. Aramco Ventures brings a strategic industrial appetite — the kind of operating environment where an agent’s mistake has physical and regulatory consequences, not just a bad support ticket. Mayfield brings conventional venture patience.
That combination does not read like a syndicate expecting a quick software-margin flip. It reads like one underwriting a multi-year build in a domain where verification matters more than fluency.
What I would want to know
The questions that would tell us whether this is a genuinely different approach or a well-funded orchestration wrapper are architectural, and they are the ones nobody answers in a funding announcement:
- Where does long-lived state live, and how is it reconciled when the underlying enterprise systems change beneath it?
- How are agent actions attributed and audited well enough that a regulated client will accept them?
- Is the verification loop model-based, deterministic, or human-gated — and does it degrade gracefully when the agent is wrong in a way it cannot detect?
- Does the system get better per-client, or does improvement pool across clients without leaking data?
Those answers determine whether the company is building a compounding asset or a very expensive consulting practice with autocomplete. The distinction will not show up in press coverage. It will show up two years from now in renewal rates.
For a field that has spent three years treating model capability as the frontier, the more useful signal in this round is where the capital is going instead. Somebody with a long memory of how enterprise software actually gets deployed raised $85 million to work on the parts that are not the model. That is the bet worth watching.
đź•’ Published: