Before you touch a workflow builder, name the job. Is the agent replacing a manual process? Filling a gap nobody has time for? Boosting something that already works?
For this walkthrough, I'll use an idea I've been sitting on: a customer lifetime value agent that lives inside your CRM or order processing system. It watches purchase events — timestamps, contacts, what was bought — and finds the moments where customers naturally graduate from one service tier to the next.
Most businesses handle this with a blanket database reactivation every three months. That's a time-based guess. The agent instead analyzes when each individual customer is actually ready, per service, and triggers the outreach then. Same job, sharper timing.
That's the setup. Now the math.
Say the data shows patterns like this: 75% of customers who buy Service A find success within three months and upsell themselves to Service B with zero persuasion. 50% do it in three to six months. 10% take six to twelve.
Those people are already upgrading on their own. The agent's job is to lift each tier by nudging at the right moment. Assume a conservative bump — plus 15% at the top tier, plus 5% at the two lower ones. That's a 25% total lift.
To keep the math clean, treat each percentage point as one person out of a base of 100. So 25 additional customers move up. If Service B runs $250/month and Service A runs $100/month, the added MRR isn't $6,250 — it's $6,250 minus the $2,500 they were already paying. Net added MRR: $3,750.
Those numbers are aggressive and made up. But the shape of the calculation is the point — you're modeling the delta, not the gross.
Pin down what the build actually costs over the window you care about. For this example: $500 setup, $150/month, 12-month commitment. That's $2,300 all-in.
On the reward side, $3,750 in added MRR over 12 months — if it hits every month, which it won't — is $45,000. Again, aggressive. Real months will have zero-progression stretches. Model both a conservative and an optimistic case if you want to be honest with yourself.
Net gain: $45,000 minus $2,300 = $42,700.
ROI: net gain divided by cost, times 100. $42,700 / $2,300 × 100 = 1,857%.
That's the number that tells you whether to build. Anything above zero is technically a return, but I want to see the ROI clear a threshold that justifies the operational overhead of maintaining the agent. A four-figure ROI on paper still means you have to run the thing, watch it, and fix it when it drifts.
Here's where I started paying attention to something I hadn't tracked before: net revenue retention.
NRR tells you whether your existing customer base is growing in value, shrinking, or holding flat — without counting new customers. It's the signal for whether you could stop acquiring tomorrow and still be fine, or whether churn and downgrades are quietly eating you.
An AI build that boosts upsells doesn't just generate ROI in isolation. It moves NRR. And that's the number that tells you the business itself is compounding.
You need five inputs for the period you're measuring: starting revenue, upsell revenue, downgrade revenue, churn revenue, and ending revenue.
Using the same example on a monthly basis: starting revenue $20,000. Upsells $3,750 from the agent. Downgrades $500 — some customers dropped from Service B back to A. Churn $250 — a couple exited entirely. Ending revenue: $23,000.
Formula: (starting + upsells − downgrades − churn) / starting × 100. Or just ending / starting × 100. Result: 115%.
Above 100% means your existing base is growing on its own. Below 100% means you're leaking, and every new customer you acquire is just filling the hole. That's the situation where an upsell agent isn't a nice-to-have — it's a structural fix.
Run the ROI math to decide if a build is worth it. Run NRR after it ships to see if the build is actually shifting the business. Two numbers, two jobs. Do both before you commit.