What to Ask a BPO Before They Touch a Regulated Process
AI agents are cutting processing times by half or more. The question is no longer whether your BPO uses them, but whether it can govern them inside your regulated systems.
by Rohit Gupta | 11th November 2026 | 11 mins read
AI is saving real time, but governance is the gap
The time savings from AI in financial services are no longer theoretical. Some published examples:
- Bajaj General Insurance: an AI agent that checks motor claims against policy terms cut claims admissibility time by 60%, GenAI extraction halved claims registration time, and AI document analysis cut surety bond underwriting time by 80%, according to Bajaj Finserv’s Q2 FY27 disclosures. AI bots served 68.9% of its customers.
- A large US health insurer: automating call summaries cut after-call work by 80% and average handle time by 20% across more than 1,000 agents.
- Tower Insurance, New Zealand: AI in the contact centre improved call handling time by 26% and freed 8,200 agent hours in six months, as published by AWS.
- Independent research: a study of 5,179 support agents published by the National Bureau of Economic Research found AI assistance raised productivity by 14% on average, and by 34% for newer agents.
So the business case is real. Many BPOs will now offer AI agents as part of their delivery, and they will promise faster turnaround and lower cost.
The hard part is what happens when those agents enter a regulated system. On 15 October 2026, CERT-In’s Critical Advisory CIAD-2026-0049 described AI agents that reached live systems of Indian organisations through exposed credentials, insecure APIs and excessive cloud permissions, and read personal and confidential data. The same month, major US banks began restricting third-party AI agents that tried to act beyond what customers had asked. RBI’s April 2026 advisory on AI-accelerated cyber threats asks regulated entities to control AI agents’ privileged access and third-party AI risk.
That is the real question for a bank, NBFC or insurer: can this BPO prove that its AI agents will stay within their limits, that no critical decision is made without a person, and that the agents cannot be used, or misused, to commit fraud?
Start with your own objective
Before asking a BPO anything, be clear about what you are trying to achieve. The right partner for cutting handle time is not always the right partner for handling a volume spike, and each objective changes which questions matter most.
| Your objective | What to measure | Questions that matter most | Watch out for |
|---|---|---|---|
| Reduce handling time (AHT) or turnaround | Average handle time, turnaround, first-time-right rate | How AI is used, where humans check, how accuracy is measured | Speed gained by skipping checks |
| Cut cost per transaction | Cost per file or call, including rework | Pricing model, staffing mix, how AI savings are shared | Low headline rates with high rework or hidden fees |
| Handle larger or seasonal volumes | Backlog, ramp-up time, SLA during peaks | Ramp-up capacity, training time, multi-site cover, AI scalability | Quality dropping when volume spikes |
| Improve quality and compliance | Error rate, audit findings, complaints | QA sampling, maker-checker, audit trails, regulatory experience | A vendor that treats compliance as your job |
| Extend hours or languages | Coverage hours, language coverage, customer satisfaction | Shift design, language capability, voice and chat bots | Overnight shifts with thin supervision |
Most institutions have more than one objective, but one usually matters most. Tell the BPO which one it is. A good partner will shape its proposal around it. A weak one will send the same deck it sends everyone.
What banks, NBFCs and insurers can and cannot outsource
Next, be clear on what your regulator allows. The rules differ by type of institution, but they share one idea.
Banks and NBFCs. RBI’s outsourcing directions, now updated for commercial banks in the 2025 Managing Risks in Outsourcing Directions, prohibit outsourcing core management functions: policy formulation, internal audit, compliance, determining compliance with KYC norms, sanctioning loans and managing the investment portfolio.
Insurers. The IRDAI (Outsourcing of Activities by Indian Insurers) Regulations 2017 prohibit outsourcing decision making in underwriting and claims, grievance redressal, product design, actuarial functions and investment.
US institutions. The 2023 Interagency Guidance on Third-Party Relationships, New York DFS Part 500 and the NAIC’s AI model bulletin all hold the institution responsible for its vendors and their AI.
| A BPO can do this | The regulated entity must keep this |
|---|---|
| Collect, check and index KYC documents | Decide whether KYC is compliant |
| Process loan applications and prepare credit files | Sanction the loan |
| Triage AML alerts and draft investigation notes | Decide whether to file a Suspicious Transaction Report |
| Prepare chargeback and dispute files | Decide the dispute outcome |
| Collect claim documents and prepare claim files | Decide whether an insurance claim is paid |
| Log complaints and draft replies | Own grievance redressal and the final response |
The work can move. The accountability cannot. An AI agent that “approves” KYC files or small loans under a threshold is making a decision that belongs to you.
An example scenario to test the answers
Questions are only useful if you can tell a good answer from a weak one. To make that easier, the rest of this piece uses one example scenario.
The scenario: A mid-sized NBFC wants a BPO to handle loan application processing and KYC document checks. Its main objective is to cut turnaround from 48 hours to the same day. It also wants to handle three times the normal volume during the festive season, without adding permanent headcount. The BPO proposes a mix of trained staff and AI agents that read documents, check them against the application and prepare credit files.
For each group of questions below, you will see what a weak answer and a strong answer sound like in this situation.
Questions about scope and accountability
What to ask:
- Which steps will your people handle, which will your AI handle, and which stay with us?
- How do you make sure no decision our regulator reserves for us is made by your team or your tools?
- Who on your side is accountable for this process, by name?
How the answers differ in the scenario:
| Question | Weak answer | Strong answer |
|---|---|---|
| Who does what? | “We’ll follow your SOP and handle the whole loan journey.” | “AI extracts and matches documents, our team checks exceptions and prepares the credit file, and your credit team sanctions. Here is the step-by-step map.” |
| What about reserved decisions? | “Our AI can auto-approve simple cases to save time.” | “Our agents never approve or reject. Every file goes to your credit team with a recommendation and the evidence.” |
| Who is accountable? | “Our account manager will coordinate.” | “Our delivery head owns this process, and our compliance lead reviews it monthly with your team.” |
The weak answers sound helpful, and the auto-approval offer may even look attractive against a same-day target. But it would put a sanction decision with the vendor, which RBI does not allow.
Questions about people and access
What to ask:
- How do you screen staff, and how often is screening refreshed?
- How is access to our systems approved, reviewed and removed?
- How will you add people for the festive peak without losing quality or control?
How the answers differ in the scenario:
| Question | Weak answer | Strong answer |
|---|---|---|
| Screening | “All our staff are background-checked.” | “Background and reference checks before access, refreshed every year, with records available for your audit.” |
| Access | “Team leads manage access.” | “Access is requested per person, approved by your team, reviewed every quarter and removed the same day someone leaves.” |
| Festive peak | “We’ll hire more people when volume rises.” | “We cross-train a reserve pool in advance, add AI capacity for document checks, and keep the same maker-checker rules at three times the volume.” |
Questions about AI agents in your systems
AI tools that only suggest are one thing. AI agents that can log in, read records and take actions are another. These questions test whether the BPO knows the difference.
What to ask:
- Which AI tools or agents will touch our process, what does each do, and which model or platform sits underneath?
- Does every agent have its own identity, separate from human users, so every action is traceable?
- What is each agent allowed to read and change? Is it read-only by default?
- Can an agent change sensitive fields such as a customer’s bank account, mobile number or payout details?
- How do you stop an agent immediately, and when was that last tested?
- Will our data be used to train any model?
How the answers differ in the scenario:
| Question | Weak answer | Strong answer |
|---|---|---|
| What AI do you use? | “We use advanced AI to speed things up.” | “Two agents: one extracts data from KYC documents, one compares it to the application. Both run on a named model hosted in India.” |
| Agent identity | “The agents run under our team’s login.” | “Each agent has its own service identity, so every read and every change is logged against that agent.” |
| Access | “They need full access to work properly.” | “Read-only access to the documents and application fields they need. They can draft a credit file but cannot submit or change customer data.” |
| Sensitive fields | “Agents can update records to save time.” | “Agents cannot change bank accounts, mobile numbers or payout details. Those changes need a person and a second approval.” |
| Stopping an agent | “We can switch it off if needed.” | “One switch stops all agents on your process. We test it every month and can share the test log.” |
| Training on your data | “Our AI keeps learning from your data.” | “Your data is not used to train any model. That is in the contract and switched off in the platform settings.” |
The sensitive-fields question matters most for fraud. In financial services, the classic fraud is not reading data, it is quietly changing where the money goes. An agent that can change a bank account or a mobile number is a fraud route, whether through a bug, a manipulated prompt or a compromised credential.
How AI governance should work in a regulated process
Good answers to individual questions are not enough. You need to see how the BPO governs AI as a whole, before, during and after an agent goes live. This is what I would expect to see.
1. An AI register. A single list of every AI tool and agent used on your process: what it does, what data it touches, which model sits underneath, who owns it and when it was last reviewed. If an agent is not on the register, it should not be running.
2. Risk tiers. Not every agent carries the same risk. An agent that summarises calls is lower risk than one that prepares credit files. Higher-risk agents get more testing, more human checks and more frequent review.
3. Joint approval before go-live. No agent goes live, and no change is made to an agent, without written approval from both the BPO and the institution. That approval should follow testing for accuracy, invented outputs, and behaviour under unusual or manipulated inputs.
4. Identity and least access. Each agent has its own identity, only the access it needs, read-only by default, and limits on how many records it can touch in a given time. This is the “Know Your Agent” principle the IBA chairman has called for: verify the agent, limit it, log it and make someone accountable for it.
5. Human approval points. For every step that affects a customer’s money, account or record, a named person approves, and the system records who.
6. Fraud controls on the agents themselves. Agents cannot change payout details, bank accounts or contact numbers. Unusual agent behaviour, such as reading far more records than normal or working at odd hours, triggers an alert. Maker-checker rules apply to agents the same way they apply to people.
7. Monitoring after go-live. A simple dashboard tracking accuracy, how often humans override the agent, error rates and any drift over time. If the override rate climbs, something has changed.
8. Full audit trail. Every agent action logged: what it read, what it produced, what a person did with it. Logs kept for at least 180 days in India, in line with CERT-In’s directions.
9. A tested kill switch and an AI incident plan. A way to stop agents immediately, tested regularly, and a plan for what happens next, including how you will be told and how the process continues manually.
10. Regular review. A monthly joint review of the dashboard and incidents, and a fuller review every quarter. Standards such as ISO/IEC 42001 for AI management systems and RBI’s April 2026 advisory give a useful structure for this.
In practice, the day-to-day split between AI and people should look like this:
| AI does | Human does |
|---|---|
| Reads documents and extracts key fields | Checks low-confidence fields and anything flagged |
| Matches documents against the application | Resolves mismatches and decides what to ask the customer |
| Drafts the credit file or case summary | Reviews and approves before it goes to the institution |
| Scores and prioritises AML alerts | Investigates and recommends whether to escalate |
| Monitors files and calls for compliance gaps | Investigates flagged cases and coaches the team |
Ask the BPO to walk you through one real case, showing where the AI acted and where a person did. If they can’t, the governance exists on a slide, not on the floor.
Questions about where your data is stored
“Your data is on the cloud” is not an answer. There are very different ways a BPO can store and process your data.
| Storage model | What it means in simple terms | What to ask |
|---|---|---|
| Your systems, accessed remotely | The BPO’s people work inside your applications through secure virtual desktops; data stays with you | Can anything be downloaded or copied out? |
| The BPO’s own cloud account | Data is copied into infrastructure the BPO rents from a provider such as AWS, Azure or Google Cloud | Which region? Who holds the encryption keys? |
| Third-party software tools | CRM, telephony, OCR or AI platforms that store your data in their own systems | Which tools, which vendors, which countries? |
| AI model providers | Documents and prompts are sent to an AI model, sometimes hosted outside India | Is data sent out of India, stored or used for training? |
| Private servers | Data sits on servers in the BPO’s own data centre | Physical security and disaster recovery |
| Backups | Extra copies kept for recovery, often in another region | Where are they, and are they deleted at exit? |
How the answers differ in the scenario:
| Question | Weak answer | Strong answer |
|---|---|---|
| Where is the data? | “Everything is on a secure cloud.” | “Your team works inside your loan system through virtual desktops. The only copies we hold are in an India cloud region, encrypted with keys you control.” |
| AI processing | “Our AI provider is enterprise-grade.” | “Document AI runs in an India-hosted environment. No documents leave India, and none are stored by the model provider.” |
| Exit | “We’ll delete everything when the contract ends.” | “We delete data from our systems, backups and every third-party tool within 30 days and give you a signed deletion certificate.” |
For payment data, RBI’s localisation rules require storage only in India. And under the DPDP Act, the institution remains the Data Fiduciary even when the BPO does the work, as I covered in our piece on data privacy in outsourcing.
Questions about security and incidents
What to ask:
- Which certifications do you hold, and can we see the latest audit reports? Look for ISO 27001, SOC 2 Type II, ISO 27701, PCI DSS where card data is involved, and increasingly ISO 42001 for AI.
- How quickly will you tell us about an incident?
- What happens if one of your sites goes down during the festive peak?
How the answers differ in the scenario:
| Question | Weak answer | Strong answer |
|---|---|---|
| Certifications | “We’re ISO certified.” | “Here are our current ISO 27001 and SOC 2 Type II audit reports, with the findings and how we closed them.” |
| Incident notice | “We’ll inform you within 48 hours.” | “Within two hours of detection, so you can meet CERT-In’s six-hour reporting rule. Here is the name of our incident lead.” |
| Site failure | “We have a backup plan.” | “A second site in another city is trained on your process and can take over within four hours. We tested it last quarter.” |
Contract clauses that protect you when things go wrong
Most of the pain in a failed outsourcing relationship comes from what the contract doesn’t say. Our guide to BPO contract clauses covers several in more depth. The ones I would check first:
- A separate, higher liability cap for data breaches, since standard caps are often only a few months of fees
- Breach notification in hours, starting from detection
- Approval rights over sub-contractors, sub-processors and AI providers
- AI change control: no new agent, tool or model change without written approval
- A ban on using your data to train models
- Audit rights and regulator access, with the BPO’s cooperation committed
- Exit and transition support, with data returned in a usable format and deletion certified
- Termination for regulatory breach
- Clear limits on auto-renewal and price increases
Business risks that rarely make the checklist
Financial stability. A BPO in cash trouble cuts corners on staff and security long before it fails. Ask for financial statements.
Concentration and lock-in. One vendor running several critical processes, or depending on one AI platform, is a single point of failure.
Conduct risk. For collections, sales and complaints, the BPO’s staff act in your name. RBI’s new directions on loan recovery, effective 1 January 2027, set stricter requirements on recovery agents, and the lender stays answerable.
Location risk. A single delivery site is exposed to floods, power cuts and local outages. Multiple sites in different cities reduce this.
Conflicts of interest. Ask whether the BPO runs the same process for your competitors, and how it keeps teams and data apart.
Three questions for BFSI leadership
Do we know every AI agent that can touch our systems through our vendors today? If not, the CERT-In scenario is already possible inside your organisation.
Can any vendor’s AI agent change where our customers’ money goes? If the answer is anything other than a clear no, that is the first fix.
If we had to exit this vendor in 90 days, could we? If the contract and the data would not let you, you are more dependent than you think.
Good BFSI BPO selection is no longer about finding the cheapest capable vendor. It is about finding one that can deliver your objective with AI, and prove it can govern that AI inside your regulated systems.
At Venturesathi, we expect banks, NBFCs, fintechs and insurers to ask us every one of these questions. We run regulated BFSI processes with human accountability on every decision, governed use of AI, and data handling we are happy to show in detail. If you are evaluating partners, we’re glad to walk you through our answers.
Frequently Asked Questions
What should I ask a BFSI BPO before signing?
Start by telling the BPO your main objective, such as lower handling time, lower cost or higher volume. Then ask what it will and won’t do under your regulator’s outsourcing rules, how its people and AI agents get access to your systems, how AI is governed and approved, where your customers’ data is stored, how fast it reports incidents, and what the contract says about liability, sub-processors, audit rights and exit. Ask for evidence, not just answers.
What can a bank or NBFC not outsource under RBI rules?
RBI’s outsourcing directions prohibit banks, NBFCs and other regulated entities from outsourcing core management functions. These include policy formulation, internal audit, compliance, determining compliance with KYC norms, sanctioning loans and managing the investment portfolio. A BPO can support these processes, but the decision must stay with the regulated entity.
How should a BPO govern AI agents in a regulated process?
A BPO should keep a register of every AI agent, rate each by risk, get joint approval before any agent goes live or changes, give each agent its own identity and the least access it needs, require human approval for decisions affecting customers, block agents from changing payout or contact details, monitor accuracy and overrides, log every action, keep a tested kill switch and review everything regularly with the client.
What is “Know Your Agent” in financial services?
“Know Your Agent” is the idea that financial institutions should verify and control AI agents the way they verify customers. It means giving each AI agent its own identity, limiting what it can access and do, setting transaction limits, keeping audit trails of its actions and making a named person accountable for it.
How much time can AI agents save in BFSI operations?
Published results vary by process. Bajaj General Insurance reported that AI cut claims admissibility time by 60% and halved claims registration time. A large US health insurer cut after-call work by 80% by automating call summaries. Independent research found AI assistance raised support agent productivity by 14% on average.
How should a BPO store BFSI data on the cloud?
Data should be stored in approved regions, encrypted at rest and in transit, with clear control over who holds the encryption keys. Payment data must stay in India under RBI’s localisation rules. The BPO should list every place a copy exists, including backups, logs, third-party tools and AI platforms, keep logs for at least 180 days within India, and certify deletion when the contract ends.
About the Author
Rohit Gupta is a Chartered Accountant and the Founder of Venturesathi, on a mission to prove that world-class global operations aren’t defined by geography, but by discipline, systems, and intent.
In 2016, Rohit launched his first BPO in Rourkela, Odisha, mastering the complexities of global delivery from the ground up. Today, he leads Venturesathi, a team of 300+ professionals delivering high-tier CX, software development, and back-office operations that bridge the gap between tier-3 economics and tier-1 execution standards.
With over a decade of experience, Rohit specializes in building “audit-ready” scalable models. His background in finance (ISA) and deep technical expertise in data tools (Power Query, DAX, Automation) allow him to design operations that are as measurable as they are efficient. At Venturesathi, the philosophy is simple: don’t just provide a service, act as a Sathi (partner), helping global clients scale without the chaos.
Connect with Rohit on LinkedIn.
Related Reading on Venturesathi
- From Logging Complaints to Resolving Them: Grievance Redressal in Indian Insurance
- Why Cashless Approvals Still Take Hours: Inside the Pre-Authorization Bottleneck
- Outbound Calling in Insurance: Staying Compliant Without Losing Reach
- Data Privacy in Insurance Operations: What Changes When You Outsource
- Banking & Financial Services (Service Page)