How to Price AI: Lessons From 12 Years at Salesforce
5 mins
November 10, 2026
The most reliable way to price an AI product is to charge on a metric customers already use to describe value, one they can predict and explain back to you. Then check the margins. Not the other way around.
Glenn Dearing ran pricing at Salesforce for 12 years and learned this from models that worked and models that had to be unwound. His advice focuses on seven moves:
- Let customers name the metric.
- Ship before you're certain.
- Favor metrics customers can predict.
- Switch to fungible credits once line items overwhelm the order form.
- Legibility over transparency.
- Be generous to early adopters.
- Package for how your sales team actually sells.
This is Part 2 of a series drawn from my webinar with Glenn.
- Part 1 covered why agents break per-user pricing, and why to price for value before checking the economics. Part 3 covers the systems that have to keep up.
1. Let customers name the metric
An attendee asked Glenn how to find a pricing metric that shows clear value to the customer and also scales with COGS. His answer started by focusing on the value you deliver to your customer rather than focusing on your cost to serve.
Before any new pricing project, Glenn interviews four or five friendly customers. He asks how the capability would make them money or save them money. Customers answer in their own terms: I wouldn't have to hire someone for this, or I could drop that other tool. Those words point to the metric.
Agents make this step more important. When an agent does the work, the value shows up in what got done, not in who logged in. Ask customers what the agent's work replaces, and the answer is usually an outcome metric they already count, such as tickets resolved, contracts reviewed or hours saved.
A follow-up question asked whether customers will really share that with a pricing person. Glenn said friendly ones will. They want you to stay in business and build the next thing that helps them.
Then check the margin
On COGS, Glenn's advice was to worry less about whether the metric tracks your costs. Pick a metric customers find intuitive. Then confirm the price sits high enough that the cost to serve can't push you underwater on margins.
If a few outlier users' usage could meaningfully hurt your margin, add protections. Glenn recommends limits on the capabilities that drive inference cost, plus a clear way to buy more capacity. Alternatively, consider fair usage clauses.
Timing matters too. Glenn doesn't worry about COGS for roughly the first six months. The figure he cares about comes once fixed costs are covered and usage is stable.
2. Ship before you're certain
When Salesforce launched its first generative AI products, Sales GPT and Service GPT, the team knew from day one the packaging wouldn't last. Customers bought credits at $3 per 100. A separate $0 SKU existed only so they could sign the legal terms.
Glenn remembers the first slide he showed the executive team: "We are gonna launch this fast. We are gonna learn fast. We are gonna iterate fast."
The point was to give customers something to buy. Boards were telling their companies to spend $100,000 to $200,000 on generative AI just to learn. Salesforce had something to sell, and many customers trusted it. Glenn said the team learned a lot working alongside those early buyers. The SKUs were gone within about six months.
Swap rights made it work. Salesforce wrote into contracts that early buyers could move to whatever replaced the original SKUs about six months later. Early buyers got a path to the newer products instead of being locked into the first version.
Treat early pricing as a learning budget. An attendee asked how Glenn overcame the fear of committing to hard-to-forecast products. He treated early pricing almost like an R&D expense. He'd hold a number in his head, for example a million dollars to learn whether a product works, and watch for cost overruns. As a finance person, I like that framing. The early months become a budgeted experiment, and the margin check runs on real usage data instead of guesses.
Glenn also noted that buyers are more willing to experiment now than they were a decade ago. They know it's early in AI deployments, and they're learning alongside their vendors.
3. Favor metrics customers can predict
A value metric still has to be one customers can forecast. In 2012, Glenn changed Force.com from per-user pricing to per-application pricing, charged by custom objects. The metric correlated with value. But it failed because customers couldn't predict how many objects they'd need.
Calculators helped CIOs. Procurement teams still didn't understand it, and the sales team couldn't explain it. Salesforce went back to mostly per user within about 18 months.
Glenn's lesson: "You can come up with the perfect metric that tracks your cost and is correlated to value. But if a customer can't predict how much they're gonna need, it's not very useful."
If you're already on a usage model that customers can't forecast, Glenn's fix is pattern recognition. When Salesforce launched Community Cloud, customers had to guess how many people would log in and how often. Within six to nine months, Salesforce had enough data from early customers to tell new ones roughly what to expect.
So watch your first customers closely. Their usage becomes the forecast you hand the next buyer.
4. Switch to fungible credits once line items overwhelm the order form
What is a fungible credit?
A fungible credit is a unit of prepaid capacity that customers can spend across several products or actions. The vendor sets an exchange rate for each action. Customers buy one pool and spend it where they get the most value, instead of forecasting each product separately.
After Salesforce bought ExactTarget, the product was tied to a CPM metric that was commoditizing fast. Salesforce replaced it with "super messages." Customers could spend them on SMS, email, or WhatsApp messages. Nobody had to forecast how many texts they'd send in each country. The pool flexed with whatever the marketing team wanted to run that quarter.
Glenn sees another advantage. When you ship a new capability, you can add it to the credit pool. Customers can try it with credits they already own, with no sales motion required.
Credits also handle the margin check without changing what customers bought. SMS costs varied widely by country and over time. Glenn's example: an SMS in Brazil might cost one super message, then 1.8 or 2 when costs got scary. The customer still bought super messages. Only the exchange rate moved.
When to switch. I asked Glenn how to choose between dollar-denominated commitments and credits. I'd seen credit models get hard when a long product list leaves customers mapping credits to products themselves. Glenn answered with a rule for when to move away from itemized line items altogether.
Glenn's rule of thumb: switch when the account team can't explain each item on the order form. His example was a platform with 18 capabilities. A rep can't walk in and say you need a thousand of these and twelve thousand of those. At that point, a flexible pool serves the customer better.
5. Legibility over transparency
For years, Glenn told teams their pricing should be transparent. In the last two years, he changed the word to legible. Tokens forced the change.
His analogy: imagine a gas station posting its price in dollars per kilogram. It's transparent. You know the price and how much you get. But unless you race Formula One cars, you can't picture a kilogram of gas, so you can't tell what filling your tank costs.
Glenn's test for legibility is simple. Can the customer explain your pricing back to you? If they can read the order form and understand what they get and what it costs, it's legible. If they can't, "it may be transparent, but it's not terribly helpful."
Tokens fail that test on every count. Customers don't know how many they need, what each one does, or which token applies. By the time they've run experiments long enough to find out, the price has changed.
Cost-plus pricing makes this worse. A markup on token costs puts tokens on the order form, and tokens are the unit you pay for, not the unit your customer gets value from. Price on the metric customers named in your interviews, and they can explain your pricing back to you.
6. Be generous to early adopters
In early 2016, Salesforce repackaged nearly its entire portfolio, about 80% of revenue. The plan raised prices and reset functionality. Most existing customers would have paid roughly 8% to 15% more to move to the new products.
Marc Benioff saw the plan and got nervous. Some of those customers had been with Salesforce for fifteen years. He asked Glenn what it would cost to let existing customers move to the new products at their current price. The answer was more than $100 million.
Benioff approved it anyway. His reasoning was that loyal customers would buy new products at higher prices over time. Glenn didn't get the impact he wanted in the first quarter. Over the long run, he said, average selling prices kept rising.
Glenn applies the same logic to AI. Your earliest customers take a risk on new functionality. Reward them by helping them migrate generously as the product evolves.
I added a practical point on the call. A few power customers get outsized value from your product or engage more than everyone else. They're the ones to test new pricing with, and the best source for the value language behind your metric. Glenn wants "forward deployed pricing strategists" working with those customers to learn how they get value.
7. Package for how your sales team actually sells
One more lesson ties pricing to the field. At Salesforce, many revenue conversations were add-ons to existing customers, not new logos. So Glenn's team built products that fit that motion. Sandboxes (development environments) and services were priced as natural add-ons, often 10% to 20% on top of current spend.
Packaging has to work for the buyer. It also has to work for the motion your account teams already know how to run.
Part 3 covers what Finance and RevOps need so these lessons survive contact with quoting, metering and billing.
Next in the series: Part 3: What Finance Needs to Run AI Pricing That Changes
Previous: Part 1: Why AI Pricing Breaks the Per-User Playbook
Watch the webinar
Glenn covers all eight lessons, with the stories behind each, in "Pricing AI Is Hard: 8 Lessons (and a Few Scars) from 12 Years at Salesforce."
