How to Build the Business Case for AI Agents: The 3-Number Framework Your Finance Team Will Approve

Your CFO asks the question every CFO asks when faced with an AI proposal: "What exactly is the return and when?" You probably can't answer that clearly. Most teams can't. They talk about improved efficiency, better customer experience, operational excellence, all true, all unmeasurable. The CFO hears vagueness. Vagueness doesn't get budget approval.

The difference between AI projects that get funded and AI projects that sit in limbo is not the technology. It's the business case. And business cases are built wrong far more often than they should be. They lack a baseline. They lack specificity. They lack accountability. They lack timeline clarity. This article gives you the framework to build one that gets sign-off.

It's called the 3-Number Framework, and every business case for an AI agent should have exactly three numbers that matter. Everything else is context.

Why Most AI Business Cases Get Rejected

The rejection usually happens quietly. You present the business case to the CFO or the investment committee. They nod. They say "interesting" and "we'll think about it." Then nothing. Two weeks later, you follow up. They're reviewing it, they say. Two more weeks. Still reviewing. By week six, it's dead, and the conversation has moved on.

This doesn't happen because the technology doesn't work. It happens because the business case is built in a way that makes the finance team uncomfortable. The primary discomfort is that the benefits are promised but not proven. Someone claims the AI will save time, but there's no baseline showing how much time is currently being spent. Someone claims revenue uplift will happen, but there's no context for what's currently happening. Someone claims the implementation will take three months, but there's no detail on what that actually means.

The secondary discomfort is that there's no owner of the outcome. If the AI saves time, who is accountable for ensuring it actually saves that time? If it doesn't materialize, whose number missed? If no one owns it, then no one is responsible for making it happen, and the finance team has to assume it won't.

The third discomfort is the setup of the decision. Most business cases present a binary choice: do this AI project or don't. The finance team is being asked to choose between a known cost today and a promised but unquantified benefit tomorrow. That's not a decision that gets made. Decisions get made when you frame it as: what is the cost to do nothing versus the cost to do this.

Number 1: Hours Saved Per Week Per Person

This is the number the finance team understands immediately and can model easily. Start with a specific workflow that the AI will handle. Not "improve efficiency", that's too vague. A specific workflow. For example, NBFC KYC follow-up. Currently, when an applicant submits a KYC form, it goes to an operations team member who reviews it, requests missing documents via WhatsApp, follows up on outstanding items, and marks it complete once all documents are received.

Now measure the time this takes. Observe three operations team members doing this work for one week. Count the average number of KYC requests per person, per day. Measure the time per request, from receiving the form to marking complete. Let's say it's 45 minutes per KYC on average. Three people, forty KYCs per week per person, 45 minutes each equals about 30 hours per person per week spent on KYC follow-up.

An AI agent does the same workflow. It reviews the KYC form, sends a WhatsApp message requesting missing documents, re-sends reminders on day 2 and day 5, marks the KYC complete once documents are received. Time to complete the same KYC: 90 seconds of system time. The operations team member reviews the completed KYCs at the end of the day, takes two minutes per KYC to spot-check, and escalates issues. New time per person: 2 hours per week for QC and escalation management.

The math: 30 hours saved per person per week. Three people. ₹750/hour fully loaded cost (salary plus benefits plus overhead). That's ₹67,500 in labor cost saved per week. ₹27L per year in labour cost freed up. That's Number 1.

The finance team looks at this number and thinks: "Okay, so we're saving ₹27L per year in labour cost. If the AI system costs ₹5L to implement and ₹3L per year to operate, then the payback is less than two months, and we have ₹19L of net annual benefit." That's a decision they can make.

Number 2: Revenue Uplift from Speed and Coverage

Hours saved is one dimension. Revenue uplift is another. Some AI agents don't save labor, they increase revenue by serving more customers or converting faster. This number captures that.

Example: a real estate brokerage gets 500 leads per month from various sources. Current conversion rate is 8% (40 deals per month). Average deal value is ₹60L. Current sales process: when a lead arrives, a sales person manually reviews it, assesses fit, and decides whether to call. Response time: 8 hours. Some leads go cold by then.

Deploy a Lead Qualification Agent. It reviews every lead in real time, assesses fit based on property type, price range, location, and buyer profile. Sales-ready leads get routed to a sales person with context. Response time: 5 minutes. Conversion rate doesn't change dramatically, but does response time matter? Let's say it does. Conversion rate goes from 8% to 9% (500 leads × 1% improvement × ₹60L = ₹30L additional monthly revenue).

But coverage matters too. Sales people only have time to call 60% of leads because they're busy with existing conversations. The AI calls everyone. Maybe another 5% of people who would have been missed anyway now get called and convert. That's another 25 deals × ₹60L = ₹15L per month.

Number 2: ₹45L per month = ₹5.4 crore per year in additional revenue from speed and coverage.

Now the finance team's mental math shifts. They're not just thinking about labor savings. They're thinking about revenue upside. The AI system that costs ₹5L to implement and ₹3L per year to operate now delivers ₹27L in labor savings plus ₹5.4 crore in revenue uplift. The payback is days, not months.

Number 3: Risk Cost Avoided

This is where most business cases get weak. They don't account for the costs that don't happen. The compliance risks that don't materialize. The SLA breaches that don't occur. The customer churn that doesn't happen. These are real costs that are just harder to quantify.

Example: a fintech company handles loan applications. Compliance risk is high, if they miss regulatory deadlines or documentation requirements, the regulator fines them. Currently, deadline tracking is manual: a person maintains a spreadsheet of key dates, sends reminders, follows up. Miss rate: 5% of deadlines slip. Average fine when a deadline slips: ₹25L. Average missed deadline per quarter: 3. Cost of missed compliance: ₹75L per quarter = ₹3 crore per year.

A Deadline Tracker Agent monitors all deadlines automatically, sends escalating reminders, and flags at-risk deadlines to a compliance officer before the deadline. Miss rate drops to 0.5% (99.5% of deadlines are met). You save 2.5 × ₹25L = ₹62.5L per quarter = ₹2.5 crore per year in fines avoided. That's Number 3.

Now the business case is: ₹27L labor savings + ₹5.4 crore revenue uplift + ₹2.5 crore compliance risk reduction = ₹8.17 crore in annual benefit for ₹5L implementation cost and ₹3L annual operating cost. Payback: 5 days. ROI: 155x annually.

That's not a maybe. That's a yes.

How to Frame the Decision

The framing matters as much as the numbers. You're not asking the finance team to choose between "do AI" and "do nothing." You're asking them to choose between "grow without AI" and "grow with AI."

Here's the framing: "The NBFC is growing at 40% year-on-year. To handle next year's volume, we either hire three more operations people at ₹10L per person fully loaded (₹30L cost), plus some overhead, for a total ₹35L in additional fixed cost. Or we deploy a KYC AI agent for ₹5L initial and ₹3L annual. The KYC agent handles 3x the current volume, so we don't need to hire additional headcount, and we actually free up 30 hours per week per person for higher-value work. The cost to scale is either ₹35L annually in headcount or ₹3L annually in AI. We're asking you which."

That's not a risky bet on an unproven technology. That's a choice between two known costs. One costs more and scales slowly. One costs less and scales faster. The finance team chooses the cheaper option.

Timeline and Accountability

Finally, specify the timeline clearly and assign ownership. Not "implement in Q2," which is vague. "Week 1-2: system configuration and training data prep. Week 3-4: pilot with 10% of KYCs. Week 5-6: measure pilot results and rollout decision. Week 7-12: full deployment and monitoring." Someone owns each phase. Someone is accountable for hitting the projected hours saved and revenue uplift.

When the AI doesn't save the projected hours, you know the specific reason because someone is measuring it. When it does save the hours, you have proof and can allocate the freed-up time to higher-value work. The business case becomes something the organization actually tracks against reality instead of something that gets filed and forgotten.

That's how AI projects that start with strong business cases actually deliver on their promises.