Pay-Per-Resolution Only Makes Sense If Your Problems Are Text-Shaped
Salesforce just changed how they charge for Agentforce. Starting this month, their new Help Agent uses pay-per-resolution pricing: a flat $2 per autonomously resolved issue, and nothing if the customer asks for a human or walks away unhappy. “Your cost is tied to outcomes, not activity,” they said.
That’s progress. Finally, a vendor willing to put money on what actually matters.
But here’s what nobody is saying out loud: that pricing model works brilliantly for some companies and will be financially irrelevant for others. Not because the AI is better or worse. Because the problems are different shapes.
Why Per-Resolution Is a Real Step Forward
The old metric was deflection. How many contacts did you avoid? How many people didn’t call?
Deflection is a garbage metric dressed up in a suit. It counts every conversation the bot touched and declared done — including the ones where the customer gave up and called back ten minutes later, or never got their problem solved but also never complained. In other words, it’s a vanity number that tells you how little you spent without telling you whether anything actually got fixed. (We’ve written before about how rising AI deflection rates often correlate with rising truck rolls — the two numbers move together when the root cause never gets addressed.)
By contrast, pay-per-resolution forces a different question. Did the customer’s problem go away? Was it resolved — fully, without a human in the loop? You pay when it works. You don’t pay when it doesn’t. As a result, there’s a natural incentive to ensure the AI actually closes the loop rather than just appearing to.
This is why Salesforce’s shift is worth paying attention to. In fact, it’s an implicit admission that deflection was the wrong scorecard. When you tie payment to resolution, you’re betting that you can define — and verify — what “resolved” means. That’s a harder bet than it sounds. But it’s an honest one.
The 90% vs. 25% Gap Isn’t a Quality Problem
Look at two Salesforce customers and you’ll immediately see what I mean.
Heathrow Airport deployed “Hallie,” an AI agent built on Agentforce, and achieved a 90% chat resolution rate — meaning 9 out of 10 conversations closed without a human handoff.
GE Appliances deployed Agentforce for their Bodewell appliance care service and landed at a 25% autonomous resolution rate.
If you read those numbers without context, you’d assume GE did something wrong. They didn’t. Instead, they have a fundamentally different support problem.
Heathrow’s support surface is almost entirely informational. Flight status. Terminal maps. Security wait times. Lounge access rules. Parking prices. Visa requirements. Connection times. Every single one of these questions has a definitive text-based answer. The AI knows the answer. The customer reads it. Done. Because there’s no physical artifact involved, you never need to see anything.
GE Appliances is a different universe. Their customers are standing in their kitchen with a dishwasher that’s leaking onto the floor. Or an oven that’s making a noise. Or a refrigerator that stopped cooling. These aren’t informational queries — they’re physical-world situations that require diagnosis. And diagnosis requires context. What does the error code look like? Where exactly is the water coming from? Is that a door seal issue or a pump issue? Is the condenser coil frosted over?
Text can’t answer those questions. You need to see the hardware.
The 25% rate isn’t a failure of the AI. Instead, it’s the AI hitting the ceiling of what language can do without visual input. The other 75% of GE’s issues are physical-world problems that require eyes on the hardware before anyone can say what’s actually wrong.
Problem Shape Is the Variable You’re Not Measuring
Here’s the taxonomy that actually matters: not which AI you’re using, not how well it’s trained, but whether your support problems are text-shaped.
Text-shaped problems have these properties:
- The answer exists in a document, database, or policy
- Resolution requires no physical inspection
- The customer can verify completion in-channel (a password reset worked, a booking was confirmed, a refund appeared)
- The problem is the same regardless of the physical environment
Password resets. Account lockouts. Billing disputes. Booking changes. Shipping status. Warranty registrations. Return authorizations. These are all text-shaped. AI can own these end to end, and as a result, the economics of pay-per-resolution will look great.
Physical-world problems have different properties:
- Resolution requires understanding the physical state of something the customer has
- Diagnosis depends on context that doesn’t exist in any document
- The customer can’t verify completion without a technician or visual confirmation
- The same symptom maps to multiple possible root causes
Appliance failures. Hardware install issues. Network cable routing. HVAC diagnostics. Vehicle repair. Building infrastructure. These problems are physically instantiated. The AI doesn’t know what it doesn’t know, because what it doesn’t know is sitting in someone’s basement.
If your product category lives in the second column, you’re not going to see Heathrow-style resolution numbers — ever — without a different approach.
What Teams With Physical Products Need to Do
The answer isn’t to lower your expectations for AI. Instead, it’s to instrument the last mile.
The fundamental problem for physical-world support is an information gap. The AI doesn’t have the data it needs to resolve the issue. It doesn’t know what the technician would immediately see if they walked into the room. Ultimately, that gap is what drives the 25% ceiling.
There are two ways to close it:
1. Give the AI eyes. Remote video changes the diagnosis conversation fundamentally. When a support interaction starts with the customer showing you the problem — real-time, on camera — the agent (AI or human) has context that text-only conversation can never provide. You can see where the leak is. You can see whether the installation was done correctly. The actual error code is right there on the display — not whatever the customer typed. That visual context, captured and structured, starts to feed the data that AI needs to resolve rather than just triage.
2. When you can’t resolve, hand off fast and informed. For the subset of issues that genuinely need a human technician, speed and context are everything. The cost isn’t in the 25% you couldn’t resolve autonomously — it’s in how long the customer waits and whether the technician shows up with the right information. A handoff that includes a recorded video of the problem, customer-annotated symptoms, and a structured issue summary isn’t just better for the customer. It also reduces truck roll time, parts guessing, and repeat visits.
This is why Viewabo exists: remote visual support as a layer that goes in between the AI’s ceiling and the truck roll’s floor. Not replacing either one. Closing the gap where text ends and the physical world begins.
The Honest Version of the Pricing Story
Pay-per-resolution is an honest pricing model. It aligns incentives correctly, and the industry is better off moving toward it.
But the honest follow-up question every CX leader at a company with physical products needs to ask is: what percentage of my support volume is actually resolvable from a text conversation? What’s my version of 90%, and what’s my version of 25%?
If you have a Heathrow-shaped problem — customers asking questions that have text answers — you’re going to love the economics of this model. Consequently, your resolution rate will be high, your cost per resolution will look great compared to a human agent, and you’ll be rightfully bullish on AI owning more and more of your support queue.
If you have a GE-shaped problem — customers with physical products in variable states of working or broken — the metric that matters isn’t your autonomous resolution rate. Rather, it’s how quickly you can get visual context into the conversation, and how informed your human handoff is when that context reveals something the AI can’t close.
Per-resolution pricing doesn’t change what AI can and can’t do. It just makes the gap visible. That’s actually useful.
Now go figure out what shape your problems are.