Audit agent vendor end-to-end claims with three checks

A vendor can show an agent closing a ticket and still be selling a supervised workflow. The pitch has shifted from copilots to agent teams that own back-office processes end to end.

Ema secured $77 million in a Series B led by Creaegis, while Accel, Section 32, and Prosus added more investment. The fresh capital will expand automation of HR, IT, and finance processes. Total funding reached $140 million after the round, and the valuation rose by more than four times from the 2024 round. The raise makes the end-to-end claim worth testing.

Ema frames its AI Employees as orchestrated teams of agents that finish multi-step workflows inside a customer's existing software stack, rather than isolated one-off tasks. Its platform connects to more than 250 business applications and systems of record. Its customer list spans NTT DATA, Hitachi, ADP, PwC, Google, KPMG, Wipro, and Microsoft. Ema reports 50-plus live enterprise accounts, 1 million-plus enterprise users, and 5 million-plus actions and queries.

Funding does not prove control

Capital does not show which steps run without a person. It shows investor confidence, not operational control. Ema reported 50-fold revenue growth over two years and bookings above $150 million, counted as multiyear contract value rather than annual recurring revenue. It charges for completed tasks and business outcomes, not for software seats or AI token usage. Outcome pricing makes verification a buyer problem. The buyer should treat completion as a measurable event, not a status label.

Ema presents its AI Employees as operating under human oversight, with people holding final decision power. The phrase is useful, but it is not an audit. A human approval point can be a formality, a final decision, or a hidden escalation route. The buyer needs to see where the approval point sits in the workflow.

Autonomy is the first thing to map

End-to-end means different things in a demo and in production. A demo may show an agent drafting a request, updating a ticket, and sending a reply. In production, a person may still approve the spend, edit the draft, or reroute the case. The difference is where the claim fails.

Three checks separate real end-to-end ownership from demo-level autonomy. They are short enough to run in a vendor call and specific enough to expose a weakness.

• Mark each step on a one-page process map as autonomous or human-gated. A missing human approval point leaves the end-to-end claim unproven.
• Compare the vendor's claimed action count with a short sample of audited logs for one workflow. A count without timestamps, user IDs, and application names is a marketing number.
• Use a minute in a sandbox or vendor demo to inject a failed handoff state. Silent passage of the bad state shows a weak coordination claim.

Logs and handoffs are where claims fail

Multi-agent process automation is harder than a chatbot. Agents may read policies, update records, and notify managers. When the middle step fails, the operator needs to know who detects it, how fast, and what state the process leaves behind.

Audited logs are the minimum evidence. A log that says an agent completed a task should let the buyer trace the task to a record in the system of record. Success-only logs are thin.

Coordination failures are the real test. Agents can both believe the job is done. An agent may write a note, and another may mark the case closed. The system should detect the mismatch, pause the workflow, and route it to a person. Failure to do that leaves the vendor making an unverified claim.

Ema has nearly 200 employees and offices in Bengaluru, London, and Vancouver. Put the three checks in the vendor evaluation form and require written answers now.