Whose Budget Does the AI Go In?
Who pays for an AI pilot, who picks up the running costs, and how should vendors charge? A budget-season framework for business leaders and software founders.

Imagine you run customer support. Your company has spent the past year encouraging people to use AI, and your team has found a pretty good use for it. When a customer calls back, a rep can pull together what happened on Friday without making the customer tell the whole story again.
Now it’s budget season. Finance asks, “How much should we put in Support’s budget for AI next year?”
Up until this conversation, the number in your budget was zero. Product/IT or SG&A paid for the experiment. Engineering connected the support systems. Your senior reps spent hours going through old cases, checking what the AI picked up and what it missed. You never thought about the 'cost' of leveraging AI in your function.
We’ve been working through a version of that question at OneDigital. We have AI coworkers that help people across benefits, insurance and retirement. I want a practice leader to be able to look at what it costs to serve a client and see the AI doing some of that work alongside the people doing the rest. The idea is that only if you see the cost, do you really start to think of the value and ROI. And that is when you become more thoughtful and deliberate in you and your team's use of AI (no more Fable 5.1 on Ultra mode for a simple chat request!)
Another nuance to think about - Say two offices in the same practice get access to the same coworker. One office uses it every day. The other barely touches it. Should the first office get most of the bill? Or do we split it whether you use it or not? Like how you do for say, the Microsoft Suite of apps - some users are power users of email/Teams/Sharepoint, while the other barely goes through a few hundred emails a week. But they pay the same (or atleast that's how it's budgeted and accounted for).
These and more questions led me to think about “Who pays for AI?” more deeply.
1. Who pays while we find out whether it works?
Go back to the support team, before anybody knew the AI would be useful. I’d pay centrally for a trial on the last fifty cases. Give it a spending limit and a target date to make a go/no-go decision based on the results. At that meeting I’d want one of the senior reps who actually worked those cases. They can tell you whether the AI found the email that explained the whole mess or just summarized six pages of people being confused i.e. the true value of AI in your workflow.
The practice is already contributing something expensive here. Somebody who knows the work has to pick the cases, check the answers and explain what the AI missed. We’ve seen this with our own coworkers. The token bill is easy to pull. The time a supervisor spends teaching the thing to do its job is considerably harder to find on a dashboard.
I’d be comfortable asking a department head for that commitment before a pilot. Only once the pilot has been completed in a robust manner, would I ask her to forecast a year of usage and then add that to her budget.
2. Who pays once it becomes part of the job?
Say the pilot works. By March, the reps use the AI whenever a case changes hands. I’d put the recurring cost of those handoffs in Support’s budget. Product and Engineering would still pay for the shared platform, the integrations and the work of keeping them running. The support leader now has something she can reasonably budget for. She knows roughly how many cases change hands in a month and what it costs the AI to prepare one.
What I’d hope happens next is that she looks at a few of those cases. Maybe on a customer’s fourth call, with two reps and a billing specialist already involved, the AI history is worth every penny. On a ticket with three lines of correspondence, the rep could have read the whole thing in the time it took to generate the summary. Before this rigor, both counted as successful uses of AI on my dashboard because no one actually ran that POC to showcase value.
She might ask Product to change when the summary runs. She might decide it’s cheap enough to keep it on for everybody. Either answer is fine because it is an informed one. She knows what a good handoff is worth to her customers, and now she should feel comfortable paying for AI, just like she would if she had hired a few CS reps - because she can meaningfully quantify value.
3. How much goes in each team’s budget?
Here’s where the Microsoft comparison gets interesting. With two people on the same fixed-price license, an enthusiastic Teams user can send a lot more messages than I do and the company pays the same for both seats. With a metered AI coworker, putting it to work on thousands of additional cases can change the company’s bill. Somebody needs to see that number. The question is what we do with it.
For a sales coworker we want every producer to use, I’d forecast the annual cost from actual usage and split it by the number of producers in each office. I’d also show the managers how their teams are using it. If one office is getting good results and the other barely touches it, I want the second manager calling the first to ask what’s working. Billing the enthusiastic office for every extra conversation gives the second manager a perfectly good financial reason to carry on ignoring the whole thing.
I would draw the line somewhere, though. Say one office decides to use the same coworker to research every prospect in its territory overnight. That manager has made a spending decision the other offices had no part in. If it adds materially to the bill, I’d show her the estimate and let her decide whether the research is worth funding out of her budget. I’d be quite happy for her to say yes.
That’s roughly where I land. Give people a predictable share of the capabilities you expect everybody to use. Keep measuring consumption so you know when that forecast needs revisiting. And when a team makes a choice that meaningfully changes the cost, give them the number while they can still do something with it.
What would I charge if I were selling this?
I sold software in a previous life, so I keep thinking about the person on the other side of this deal. At SmartCue, I could ask a buyer how many people needed to create demos and give them a price. If one of those demos got forwarded to the entire sales team, great! That was pretty much what I wanted to happen.
Now imagine I’m selling the AI that prepares those case handoffs. Two support teams could each have twenty reps. One handles twice as many cases, and its customers have years of account history to go through. I can quote both teams the same price per seat. It’s wonderfully easy for both buyers to budget. I’m the one who has to work out whether I can afford the second customer.
I could charge for usage instead. Salesforce, for instance, offers Agentforce pricing where actions such as retrieving a record or summarizing a case draw down credits. That gives me, as the seller, a way to earn more when the software does more work. It also gives my buyer a question she didn’t have when she bought twenty seats: “I know how many cases we handle. How many actions does your product take to handle one?” If my pilot can’t answer that, I’ve given her a rate card and left her to figure out the budget.
Or I could charge for an outcome. Intercom charges $0.99 when Fin resolves a support conversation over chat or email. An answer can count as resolved if the customer stops replying, so counting only confirmed fixes could underestimate the bill. If the customer comes back to that conversation for more help, Intercom deducts the resolution, even across billing periods. That detail matters to both sides. The buyer knows what counts, and the seller has a reason to get the answer right the first time. But what would I call an outcome for our handoff tool? The AI finds a detail the rep would have missed. The rep gets on the phone and saves the account. I’d have a hard time telling that rep how much of the renewal belonged to my summary.
For that particular product, I’d probably try a subscription with a pool of handoffs included and an agreed price for additional volume. The buyer gets a number for ordinary months. I get paid if she rolls it out to five more offices. The fifty-case pilot gives us somewhere to start on the size of the pool. I’d want her to see when she’s getting close to the limit, though. Finding out from an invoice six weeks later would make for a fairly short renewal conversation. We see that in our Anthropic and OpenAI bills every month. Not pleasant and impossible to budget for. No wonder most large orgs are trying to figure out how bring some of these capabilities in house - either through cheaper open source models, custom trained models running on their own hardware, or a combination. Not choices leaders like to make, but their hands have been forced.
I can see a case for each model. A seat makes sense when I’m selling a capability I want every rep to have. A completed handoff is easier to buy when the customer can forecast case volume. A resolution makes sense when the vendor can deliver it and both parties can recognize when it’s happened. What’s interesting to me is that the customer can buy one way and budget another. Finance can pay a vendor by the handoff and still give each office a predictable share of the annual cost. There’s no particular reason the rep answering the phone needs to care how many credits the vendor just used.
So what could the head of Support say when Finance asks what to put in next year’s budget?
“Put the handoffs in mine. We tested them. On the longer cases, customers used to have to explain the problem again every time they got a new rep. The reps who picked up those calls showed us where the summaries helped. We’ve switched them off on the short tickets. Here’s how many of the longer ones we expect, what the vendor will charge us, and what it would cost to add the weekend team.”
That’s the conversation I’d want on both sides of the deal. As the seller, I’ve given someone enough information to buy more of my product if it works. As the person running an AI program, I’ve got a practice leader who can explain where it works, what she’s willing to spend on it, and why.
The days of AI being free for everyone in the org, but still costing the org an arm and a leg are numbered. All organizations need to solve for this equation sooner rather than later!
Robin's Notebook
A new entry every couple of weeks. No promotion, no funnel, no manifesto. Just the unfiltered version.
Subscribing opens beehiiv.com in a new tab to confirm. Unsubscribe anytime; I won't share your email.