In 2017, while I was still a junior data analyst, I spent four months cross-referencing the GitHub commit logs of a dozen high-profile ICO projects against their Telegram sentiment spikes. The math was almost always the same: marketing velocity outpaced developer velocity by a factor I could plot on a single chart, and three tokens I flagged collapsed weeks before the broader Ethereum market did. The lesson was never that founders lie. It was that narrative and settlement run on different clocks, and the gap between them is where all the real risk hides.
I thought about that chart this month when Ema, an enterprise AI agent startup, closed a $77 million Series B, bringing its total raise to roughly $140 million. The pitch is seductive and familiar: replace seat-based SaaS, price by outcome rather than access, and let a hundred-plus models orchestrate workflows across HR, IT, and finance. It reads like a 2017 token whitepaper with the ticker removed โ and the same verification gap has quietly reappeared, now wearing a suit and calling itself enterprise software.
Let me establish what Ema actually is, because the category label matters more than the funding headline. Ema positions itself as an enterprise multi-agent orchestration layer โ not a foundation-model lab. Its disclosed architecture coordinates workflows across "150+ frontier and open models," which is a precise way of saying it routes, decomposes, and assembles tasks across other companies' models. There is no signal of proprietary model architecture, no disclosed training methodology, no data-engineering moat. The innovation is compositional and engineering-level, which is not an insult โ but it is a ceiling.
The pricing model is the actual story. Ema charges by result or completed task rather than per seat. CEO-sourced figures claim 50x revenue growth over two years, roughly $150 million in "bookings," an 80% gross margin, 180% net dollar retention, 50 active enterprise deals, and about one million active users. Every one of those numbers is company self-reported and independently unaudited. The CEO explicitly declined to disclose ARR.
The customer list โ NTT DATA, Hitachi, PwC, KPMG, Wipro, and reportedly Microsoft and Google โ is where the narrative gets interesting. These are IT services and consulting firms. They are simultaneously potential customers, implementation partners, and, per Ema's own framing, the labor-heavy incumbents it claims to displace.
Here is where the data needs to be stress-tested, and here is the information gain I want readers to walk away with: Ema's most-cited growth metric, the $150 million "bookings" figure, is a multi-year total contract value โ not annual recurring revenue. Confusing the two is the single most reliable tell of a valuation built on narrative rather than settlement.
Do the arithmetic. If $150 million represents a two-to-three year TCV, the annualized equivalent lands between $50 and $75 million โ and that is signed, not recognized, revenue. Contracted value that has not been delivered is a promise, not a balance. On top of that, the "50x growth" claim is only meaningful against a base, and the base was almost certainly tiny; a jump from $500K to $25M and a jump from $5M to $250M both read as "50x" while implying wildly different businesses. When a company discloses a multiple but withholds the denominator, treat the multiple as decoration. This is the algorithmic truth behind the token narrative, restated for enterprise software: a ratio without a base is a mood, not a measurement.
The 80% gross margin deserves the same scrutiny. In a model-agnostic architecture, margin equals pricing minus the cost of multiple inference calls per task โ planning, execution, and verification often chain five to twenty model invocations. Failure retries, coordination overhead, and human escalation support are usually excluded from the headline figure. The $150 million in bookings is unaudited. The 50x growth is unaudited. The 180% NDR is unaudited. The ratio of users per deal โ roughly 20,000 โ suggests company-wide deployment rather than the replacement of any single critical system.
This is the structural tension I keep returning to: outcome-based pricing transfers the risk of agent failure from the buyer to the seller. If a task succeeds X percent of the time, the effective unit economics only hold when X is high and the sample pool is large. No success rate is disclosed. No dispute mechanism is disclosed. Under a performance-priced contract, the vendor absorbs the cost of every miss โ which means the model is either disciplined unit economics or a slow leak, and nothing in the reporting tells us which.
There is a genuine crypto parallel worth naming, and it is not decorative. Outcome-based settlement is precisely what smart contracts promised in 2017: payment conditional on verifiable delivery, enforced without discretion. Ema's model is that idea ported into a legal wrapper. The difference is that a smart contract's "outcome" is machine-verifiable, while an enterprise agent's "outcome" is disputed by a human buyer. The arbitration layer โ how "done" gets defined and defended โ is the product, and it is entirely undisclosed.
The competitive field compounds this. Sierra, Decagon, Glean, Writer, Cognition, plus every hyperscaler's in-house platform, compete for the same enterprise budget. Ema's cumulative $140 million is unremarkable in this capital-intensive category, and its investor syndicate โ Accel, Prosus, Section 32, Creaegis โ contains no cloud strategic: no compute discount, no distribution lock. That absence is a signal, not a rounding error. The platforms have not yet decided Ema is an asset they must own.
There is a second blind spot the funding narrative skips entirely: agency. An agent with write access to HR, IT, and finance systems converts a hallucination into a payroll error, a permission grant, or a wire. Outcome-based pricing may even create a perverse incentive โ if revenue depends on "task completion," the system is nudged toward aggressive execution rather than cautious confirmation. No permission model, no audit trail, and no third-party certification is disclosed, which is a hard constraint on any claim of replacing "critical systems."
Now the counter-intuitive part, the angle most coverage misses. The prevailing narrative frames Ema as a challenger attacking SaaS's seat-based model. But outcome-based pricing is not Ema's invention โ it is already the standard in the enterprise agent category, pioneered and scaled by competitors like Sierra. Ema is a trend-follower dressed as a heretic, and the "challenger" framing flatters a positioning that is crowded on all sides. Tracing the sentiment pivot from 2017 to today, the crypto industry already ran this experiment: performance-priced protocols sounded revolutionary until everyone realized the hard part was verifying the performance. The same reckoning is queued up here.
More importantly, look at what the customer list actually contains. Microsoft and Google โ both listed as Ema customers โ also build and sell their own agent platforms. Salesforce is integrating agent capabilities directly into its existing ecosystem, choosing to absorb AI rather than be replaced by it. When your customers are also your platform-layer competitors, "adoption" may mean a light pilot or a co-marketing line item, not production replacement. The IT services firms present the same ambiguity: they are simultaneously the "customers" and the "disrupted," which means the industry impact is less likely to be clean substitution and more likely to be services firms re-selling agent implementation โ extending their own lifecycles rather than shortening them.
And the industry's own baseline undercuts the grand narrative. Only 5-8% of GenAI pilots reportedly reach measurable, scaled ROI. Set that against a replacement story and you get the gap I was charting in 2017: loud roadmaps, quiet delivery. A model-agnostic orchestration layer holds no lock-in on its suppliers and no exclusive data moat, which means the moment a foundation-model provider extends downward into enterprise orchestration โ already happening โ the middle layer gets squeezed from both directions. Rewriting the ledger of crypto's lost legends taught me that middleware rarely survives the platforms it depends on.
So where does this leave the $77 million? Not as validation, and not as fraud โ as an option, priced on narrative, held by investors betting that a category eventually materializes around Ema rather than under it.
The signals I will actually watch are unglamorous. Does Ema disclose audited ARR before its next raise? Does a mainstream vendor โ Salesforce, ServiceNow โ adopt outcome-based pricing as a default, which would confirm the category rather than the company? Does any customer verifiably cancel a SaaS or outsourcing contract because of an agent? And does the 5-8% pilot-ROI figure move?
Until those resolve, remember the 2017 lesson: a whitepaper is not a ledger, a booking is not revenue, and a narrative is not a settlement. The most honest number in the entire round may be the one nobody published โ the success rate.