Enterprises are moving from experiments to outcomes. Leaders are no longer asking whether conversational AI works, they are asking where it moves the P and L first, and how to deploy it safely at scale. In this conversation, Nurix founder Mukesh Bansal and co founder Abhishek Asawa unpack what agentic AI really changes for the enterprise. The short version is simple. Agents remove headcount ceilings, unlock flexible capacity, and make it possible to operate at a level of speed and consistency that humans cannot sustain alone. The long version is what follows. It is a field guide for executives who want business impact, not just demos.
Why AI Agents have become important
Over the last year, the most visible traction has been in conversational workloads. Voice agents that qualify leads, book appointments, and resolve routine support tickets are already reducing wait times and raising conversion rates. That is only the first wave. As Mukesh notes, agentic systems are general purpose. Anything that can be described as a repeatable workflow with rules, context, and measurable outcomes is a candidate.
Early movers are going beyond calls and chats. They are using agents for internal communication tasks, drafting meeting notes, summarizing Slack threads, generating status reports, and pulling research from large document sets. They are also using agents behind the scenes for data hygiene, transaction checks, and back office automations. The pattern is clear. Start where conversations and follow ups are happening at scale. Expand into research, operations, and analytics that benefit from relentless consistency.
CEO's guide to what matters most for Agentic AI
Abhishek frames the practical question most leaders have. If opportunities are everywhere, where should a CEO start. Mukesh offers a mental model that changes the calculus. In a headcount world, growth depended on how many people you could hire and onboard in a given planning cycle. Five, ten, or twenty seats determined how many calls you could make and how many tickets you could handle. Agents change that constraint. You can run one hundred, one thousand, or five thousand concurrent conversations without linearly scaling headcount. You move from capacity scarcity to capacity on demand.
How do you turn that abstraction into a first project. Use three filters.
- Business impact. Focus on a workflow where revenue lifts or cost reductions are easy to measure. Outbound lead follow up, inbound lead triage, first line support, and recovery calls are prime. These already have baseline numbers for pick up rate, contact rate, conversion rate, average handle time, and cost per resolution.
- Variability. Prefer workflows where performance swings by time of day, by day of week, or by backlog size. For sales, same day calls often convert two to three times better than next day calls. If you cannot surge capacity to hit the same day window, you leave money on the table. Agents remove that bottleneck.
- Integration readiness. Pick a domain where you can connect to the systems of record. CRM for sales, ticketing for support, order and billing systems for operations. When an agent has context about a customer and can take action in the system, outcomes improve dramatically.
Start with one use case that is high impact, variable, and integration friendly. Prove value, then pattern match to similar workflows.
Measuring what matters for agent evaluation
A frequent question is how to price and evaluate an agent program. Minutes are tempting because they are easy to count, but minutes are not the goal. Outcomes are. Mukesh recommends translating agent efforts to unit economics that executives already understand.
Two simple metrics keep teams aligned.
Cost per Lead, CPL
CPL = Total Marketing Spend divided by Number of Leads Generated
Example: Spend of ₹10,000 that generates 100 leads yields a CPL of ₹100.
Customer Acquisition Cost, CAC
CAC = Total Sales plus Marketing Spend divided by Number of New Customers
Example: Spend of ₹2,000 that closes one customer for a ₹15,000 product yields a CAC of ₹2,000.
If an AI agent can match or beat your current CPL or CAC while improving speed to first touch and consistency of follow up, the program is working. When agents lift conversion rate while holding cost flat, CAC drops. When agents cut handling time while holding conversion constant, cost per resolution drops. In either case you are buying business impact, not minutes.
What separates a great agent from a demo
There is a misconception that agents are prompt engineering with a phone number. Real systems that ship revenue require three layers.
- Domain and company knowledge. The best human agents do not just read scripts. They recognize intent quickly, know which phrases build trust, and understand the product deeply. They also know what your brand stands for, how you speak, and what to do when a customer is confused or frustrated. Building a great AI agent means encoding the behavior of your top ten percent performers. That requires transcripts, playbooks, objection handling guides, and examples of great calls.
- Integrations and context. Agents convert when they can answer questions and take action. That demands connectors to CRM, ticketing, order management, calendars, and payment gateways. It also demands guardrails. An agent should check customer identity, pull the right record, and choose the correct next step without leaking information or creating duplicate entries.
- Production grade infrastructure. Real conversations are messy. People interrupt. Background noise spikes. Latency breaks trust. You need reliable speech to text, reasoning, and text to speech in a low latency loop, plus concurrency so one thousand calls do not degrade quality. You also need observability, alerts, and rollback paths so that if a model update changes behavior you can catch it fast.
Put simply, agents are a product, not a prompt. They combine knowledge, integrations, and a platform that keeps them fast and reliable.
A deployment approach that works for enterprises
At Nurix we use a two pillar model that aligns stakeholders and removes friction.
Pillar one, Workflow Strategy. Map the as is process, identify business value, and re design the workflow for agents. We interview operators and listen to calls to understand how top performers behave. We mark decision points, data needs, and failure modes. We design the agent’s conversation strategy, escalation rules, and success metrics. This is consulting level work that creates the blueprint for impact.
Pillar two, Platform Expertise. Execute with a platform that delivers stable infrastructure, scalable concurrency, and production reliability. This includes SIP trunking and telephony, low latency speech models, retrieval over your knowledge base, secure integrations, and evaluation tooling. It also includes the knobs that business owners need, such as brand voice dictionaries, eligibility rules, and outcome targeted experiments.
These pillars meet in one document that every stakeholder can read. A single page shows the user journey, the systems involved, and the goals. Executives see the numbers. Operators see the flow. Engineers see the interfaces.
Iteration is the strategy
Shipping version one is the start. Agents learn like teams do. They need goals, feedback, coaching, and upgrades.
- Evaluation. Define success metrics that mirror the business. For sales this might be contact rate, qualified rate, and booked meetings. For support it might be first contact resolution and CSAT. Track these daily.
- Review loops. Sample calls by outcome. Audit failures and near misses. Identify patterns such as misunderstood intents, missing knowledge, or uncertain tone. Feed those insights back into training data, retrieval content, and prompt strategy.
- Versioning. Change one thing at a time, measure it, keep what works, and roll back what does not. Treat prompts, guardrails, and knowledge snippets like code with versions and changelogs.
- Governance. Maintain redaction, access controls, and audit trails. Use test suites that cover sensitive flows, refunds, and identity checks before any update goes live.
Iteration turns a good agent into a great one. The goal is not just accuracy. The goal is faster cycles of learning than competitors can match.
A practical first project
Let us ground this with a sales example, since Abhishek called it out as a sweet spot.
The challenge. A business receives hundreds of thousands of leads per month. Agents call in waves, but peak times get overloaded. Same day contact rates are much higher than next day. Conversions drop when callbacks slip.
The agent. A voice agent reaches out in under five minutes, confirms interest, asks two to three qualifying questions, and books a time with a human rep when needed. The agent updates the CRM, sets reminders, and sends a follow up text with meeting details.
Integrations. CRM for lead records, calendar for scheduling, telephony for calls and texts, analytics for reporting. Optional payment link if the product allows instant purchase.
Targets. Increase same day contact rate, increase qualified appointment rate, and reduce no show rate. Hold or reduce spend per lead. Show CAC improvement over a four week run.
Guardrails. No sensitive data over text without masking. Explicit consent language. Clear escalation to a human for pricing exceptions or specific compliance questions.
Pilot this in one region or one product line. Publish a weekly dashboard that shows the baseline and the agent’s numbers. Once you see the lift, roll the workflow to other segments.
Internal communication and research, the quiet wins
Not every impact story needs a phone call. Many teams spend hours per week collecting updates and writing reports. Agents that summarize Slack threads, pull key decisions from meeting notes, or generate a weekly status page save precious time and reduce errors. Research agents that pull facts from thousands of documents help analysts move from search to synthesis. The more consistent the structure of the output, the easier it is to review and trust.
The trick is the same as customer facing deployments. Define the business value, integrate with the tools where the information lives, and create a review loop so the agent learns the tone and depth your team expects.
From SOPs to re engineered workflows
Mukesh makes a useful point for any enterprise that plans to hand standard operating procedures to an AI system. SOPs encode the way you worked when humans were the only option. Agents allow you to change the work itself. Instead of moving data across three screens, the agent can call an API and make the update instantly. Instead of waiting for a manager to approve every exception, the agent can enforce rules and only escalate edge cases. If you copy old SOPs, you get old outcomes with a new cost structure. If you re design the process, you unlock new outcomes.
This mindset also lowers risk. The more you can simplify a workflow, the fewer branches an agent needs to handle. The fewer branches, the easier it is to test and certify behavior.
Human oversight is a feature, not a concession
Enterprises do not have to choose between machines and people. The strongest systems are hybrid. Humans provide oversight, judgment, and approvals. Agents provide speed, recall, and availability. The right ratio changes by workflow. A recovery call with a refund might need a human in the loop. A follow up to collect missing information might not. The important part is that the choice is explicit and adjustable.
A practical way to make oversight visible is a control panel metaphor. You tune sliders such as speed versus quality, task level versus strategy level, and friendly versus professional tone. Then you monitor outcomes and adjust. Business owners understand this framing, and it keeps decisions grounded in results.
What good looks like in production
Enterprises that succeed with agents share a few behaviors.
- They stack rank workflows by value, complexity, and integration readiness.
- They align on outcomes before they choose models or voices.
- They build a small knowledge system that is curated and versioned rather than dumping a wiki.
- They ship to production with a clear fallback to a human and a clear escalation path.
- They review calls and transcripts like they review sales pipelines and incident reports.
- They choose partners who can operate both pillars. Strategy to design the workflow, platform to keep it live at scale.
When those ingredients are present, the shift from pilot to production happens fast. The agent starts as a helper and becomes an always on teammate.
The executive checklist
If you are deciding whether to start or scale an agent program, use this list in your next review.
- Do we have a workflow with measurable value, clear variability, and available integrations.
- Do we agree on unit economics targets such as CPL, CAC, or cost per resolution.
- Do we have transcripts and examples of top performance to model.
- Do we have system access to read and write the data the agent needs.
- Do we have a platform that meets latency, concurrency, security, and observability needs.
- Do we have an evaluation plan that business owners will read weekly.
- Do we have a change management plan for operators and customers.
Answer yes to those questions and you will have a project that earns its keep.
Closing
Mukesh puts it plainly. No role is untouched by what agents can do, including the CEO. That is not a threat. It is an opportunity to design work that is faster, more reliable, and more humane. Humans do the judgment and the relationship work. Agents do the follow through and the heavy lifting. Together they move the business.
If you want to see how this looks in your environment, start with one workflow and a simple goal. Call more leads the same day. Reduce first response time. Clear a backlog every evening. Prove it in four weeks. Then scale what works.
