Why Ema's round matters to Indian enterprise buyers
On 24 September 2026, Inc42 reported that Ema had raised a $77 million Series B led by Creaegis, with Accel, S32 and Prosus participating, taking its total funding to about $140 million. Ema sells what it calls "AI Employees": agentic software aimed at enterprise workflows in areas such as HR, IT and finance, which connects to 250+ business applications and routes sensitive decisions to a human for approval. Its named customers include Wipro, Hitachi, ADP and PwC. The company has a Bengaluru office, nearly 200 employees, and plans to expand into APAC and EMEA.
The round was not an isolated bet. Inc42 ranked AI as the top funded sector in the 21 to 25 September week, and Entrackr reported that AI led Q3 2026 startup funding at $635.3 million. Capital at that scale funds sales teams, and the obvious next buyers for an agent that processes reimbursements, provisions IT access and reconciles vendor payments are Indian Global Capability Centres (GCCs) and large corporates with shared-services operations.
The productivity pitch is concrete. PhonePe reported, according to Inc42 on 30 September 2026, that one audit process fell from four days to three minutes using AI agents. That kind of compression is exactly why procurement teams will be asked to sign these contracts quickly.
This post is written for the buyer: the CFO, CHRO, CIO, procurement head or risk manager about to put an AI employee into production. The vendor-side question (how an AI company insures its own liability) is covered in our post on agentic AI liability in enterprise operations. Here the question is narrower: when the agent you bought causes a loss, who pays, and what should you demand before signing?
Three losses that land in a grey zone
An AI employee does real work with real credentials. That is the point, and it is also the source of loss. Three scenarios cover most of what a buyer should worry about.
A wrong vendor payment. An accounts-payable agent matches an invoice to the wrong vendor master record, or accepts a changed bank account from a spoofed email, and releases a payment. The money is gone. Your commercial crime or fidelity guarantee policy may respond to funds transfer fraud or social engineering, but many wordings require a fraudulent instruction by a third party or a dishonest employee. A software error with no fraudster may not trigger the insuring clause at all. The vendor's tech E&O policy might respond, but only if your contract gives you a claim against the vendor and the vendor's policy covers the error.
A payroll data leak. An HR agent with read access to payroll and personal records sends salary files to the wrong distribution list, or an integration misconfiguration exposes employee data. Your cyber insurance policy is the natural first responder for breach costs, forensics and notification. Under the Digital Personal Data Protection Act, 2023, you remain the data fiduciary and stay responsible for processing done on your behalf by a data processor, whatever the contract says. The regulatory exposure stays with you even when the vendor's agent made the mistake.
A wrong approval. A finance or HR agent approves a reimbursement, leave encashment or access request that policy should have rejected. Each item is small, but an agent can repeat the same mistake thousands of times before anyone notices. This is pure financial loss with no fraud and no breach, which is the category least likely to be covered by your own policies.
The indemnity and liability terms to negotiate
Standard SaaS terms were written for passive software. They usually cap the vendor's total liability at fees paid in the preceding 12 months and exclude indirect and consequential loss. For an agent that moves money and handles personal data, that default leaves the buyer carrying almost all the risk. Push on these points:
- Separate cap for agent errors. Ask for a higher, specific liability cap for losses caused by the agent's actions (wrong payments, wrong approvals, unauthorised data disclosure). A multiple of annual fees, or a fixed rupee figure tied to the value of transactions the agent touches, is more realistic than 12 months of subscription.
- Data protection super-cap. Carve out breach of data-processing obligations into its own higher cap, sized against your realistic breach-response cost and regulatory exposure.
- Define direct loss. Write into the contract that funds paid to the wrong party, recovery costs, and breach-notification costs are direct loss. Otherwise the vendor will argue they are consequential and excluded.
- Indemnity for third-party claims. The vendor should indemnify you against claims by employees, vendors or regulators arising from the agent's error or the vendor's security failure.
- Human-approval carve-out language. Vendors will argue that once a human clicked approve, the loss is yours. Agree in advance what the approver actually saw. If the agent presented incomplete or wrong information, the approval should not shift liability back to you.
- Subcontractor and model-provider flow-down. Agentic platforms rely on third-party model providers and connectors. The vendor should stay liable for its subprocessors.
A practical shortcut: ask the vendor what enhanced terms it has already accepted from large enterprise customers. A vendor already selling to large enterprises has usually moved beyond its standard paper at least once, and starting from that position saves weeks.
Checking the vendor's tech E&O cover before you rely on it
An indemnity is only as good as the balance sheet or the insurance behind it. A venture-funded vendor with $140 million raised may look solvent today, but a contractual promise to pay is worth far more when a technology errors and omissions (professional indemnity) policy sits behind it. Ask for:
- A certificate of insurance naming the tech E&O insurer, limit, retention and policy period. Ask for renewal certificates every year, not just at signing.
- The limit and whether it is shared. A vendor serving hundreds of enterprises on one aggregate limit may have little left if several customers claim in the same year. Ask whether the limit is aggregate across all clients.
- Confirmation that AI output is not excluded. Some recent wordings add exclusions or sublimits for losses arising from generative or autonomous AI output. Ask the vendor to confirm in writing that errors by its AI employees are within the insuring clause.
- Cyber and E&O together. A combined tech E&O and cyber policy should respond both to your financial loss from an agent error and to a data breach on the vendor's platform.
- India jurisdiction. If the vendor's policy is placed overseas, check that claims brought in India by Indian customers are covered and that territorial limits do not exclude them.
Set a minimum E&O limit in the contract and make failure to maintain it a breach. This is routine in IT outsourcing contracts and should be routine for agentic software too. Our post on insurance-backed AI agent performance warranties explains how some vendors are wrapping insurance into the product itself.
Human-approval logs and evidence of control
Ema's design routes sensitive decisions to humans for approval. That design is only useful to a buyer if the approvals are recorded in a form that survives a dispute or an insurance claim. When a loss happens, three parties will want to know what the agent did and who approved it: you, the vendor's E&O insurer, and your own crime or cyber insurer.
Write these logging requirements into the contract and the statement of work:
- Action logs for every tool call. Each action the agent takes in a connected application (payment initiated, record updated, access granted) should be logged with timestamp, input data, output and the application touched.
- Approval records. For every human-in-the-loop decision, record who approved, what the agent showed them, and how long they spent. A rubber-stamp approval pattern is evidence your insurer will use against you.
- Thresholds you control. You, not the vendor, should set which actions require approval: payment value limits, new-vendor bank changes, bulk payroll exports, privileged access grants.
- Retention and export. Logs should be retained for at least the limitation period for claims and be exportable to your own systems, not held only inside the vendor's platform.
- Change notification. The vendor should notify you before changing models, approval logic or connectors in ways that alter agent behaviour.
These logs do double duty. They let you prove the vendor's agent caused the loss, which supports your indemnity claim, and they show your own insurer that controls existed, which supports your crime or cyber claim.
DPDP data-processing terms for an agent with HR and finance access
An AI employee working in HR and finance will process some of the most sensitive personal data a company holds: salaries, bank details, identity documents, and health-related leave records. Under the Digital Personal Data Protection Act, 2023, your company is the data fiduciary and the vendor is a data processor. The Act places responsibility for compliance on the fiduciary, including for processing done by its processor, and the penalty schedule runs up to INR 250 crore for failure to take reasonable security safeguards.
Your data-processing agreement with an agentic vendor should cover:
- Purpose limitation. The agent processes personal data only for the workflows you configure, and is not used to train shared models without your written consent.
- Access scoping. The agent gets the least access it needs in each of the connected applications. An agent that reconciles invoices does not need payroll read access.
- Breach notification timelines. The vendor notifies you fast enough for you to meet your own obligations to the Data Protection Board and affected individuals.
- Data location and subprocessors. A list of subprocessors, including model providers, and notice before any change.
- Deletion on exit. Personal data, logs containing personal data, and any fine-tuned artefacts are deleted or returned when the contract ends.
For a deeper look at governing AI vendors under DPDP, see DPDP and AI vendor governance. The practical point for procurement is that the DPDP exposure cannot be contracted away, so the indemnity and the vendor's cyber cover are your only route to recovering the cost.
Evidencing controls to your own insurer at renewal
Your own crime, cyber and D&O insurers will start asking about AI agents in renewal proposals, if they have not already. Treat that as an opportunity: a buyer that can show disciplined deployment is in a stronger position on terms and on claims.
What to tell your crime insurer
Disclose that an AI agent can initiate or release payments, and describe the controls: dual approval above set thresholds, call-back verification for any bank-detail change, and segregation between the agent that creates a payment and the human who releases it. Ask the insurer to confirm, ideally by endorsement, that funds transfer fraud and social engineering cover applies where a fraudulent instruction is processed by an automated system. Our post on deepfake payment fraud and crime cover covers the related social-engineering wording issues.
What to tell your cyber insurer
List the agentic platforms with access to personal data, the data categories, and the logging and access controls above. Confirm that a breach caused by a vendor's platform is covered as a breach of your data, and check how the cyber insurer's subrogation rights against the vendor interact with your contract's limitation of liability. A contractual waiver of claims against the vendor can prejudice your insurer and your claim.
A procurement checklist for AI employees
Use this list before signing with Ema or any agentic AI vendor that will touch HR, IT or finance systems:
- Map every connected application and the actions the agent can take in each. Mark any action that moves money or exports personal data.
- Set approval thresholds yourself and confirm the vendor's platform enforces them.
- Negotiate a separate liability cap for agent errors and a data-protection super-cap, with wrong payments and breach-response costs defined as direct loss.
- Obtain a third-party indemnity covering employee, vendor and regulator claims.
- Collect a certificate of insurance for tech E&O and cyber, confirm AI output is not excluded, and require annual renewal evidence and a minimum limit.
- Specify action logs, approval records, retention and export rights in the contract.
- Sign a DPDP-compliant data-processing agreement covering purpose limitation, access scoping, breach timelines, subprocessors and deletion.
- Check that nothing in the contract waives your insurers' subrogation rights without their consent.
- Disclose the deployment to your crime and cyber insurers and seek written confirmation that cover applies.
- Review the arrangement after the first quarter of live use, once you can see real approval patterns and error rates.
GCCs and shared-services centres that do this upfront will find contract negotiation faster on the second and third agent deployment, because the template already exists. Those that skip it will discover the gaps the first time an agent pays the wrong vendor.