Six AI agents cost a customer success team $3,511 a year in model tokens on a book of 500 accounts, which is about $7 per account and $88 per account manager. The figures come from a cost model we built in October 2026 for a client's 2027 plan, scaled here to an illustrative book of 500 accounts and 40 account managers. At conservative benchmarks the same six agents hand back about 11,100 hours a year, or about 5,600 hours after we assume half of that time gets absorbed by the day instead of redeployed. Most of the arguments I hear about what agents cost are about the wrong line item, and these numbers are why.
At that size the token bill is a rounding error, whether the agents run in HubSpot or in Salesforce. The money in an agent program goes into building the agents, writing the business rules they run on, getting the data they read into the CRM, and changing how the team works once the hours come back. We build these agents for clients on both platforms, so we have a stake in the answer, and that is why every number below shows where it came from.
This matters right now because a lot of private equity operating partners are trying to put an AI efficiency number into their 2027 portfolio plans. The number that will hold up in a board meeting is hours returned to the team, net of what gets absorbed, measured after go-live. Run cost will not move that number in either direction, so it should not be what holds a program up.
The short answer
Six AI agents cost $3,511 a year to run on a 500-account book
$3,511
a year in model tokens for the six core agents, across 39,359 runs.
11,139
hours a year handed back to 40 account managers, or about 5,600 after halving for absorbed time.
$7
per account per year in tokens, which is about $88 per account manager.
Cost per run for each agent
The most expensive agent in the set costs $1.44 a run, and the four agents on the standard model cost between 2 and 6 cents a run. These are the six core agents, the ones whose time savings rest mostly on published benchmarks, and the run counts are a full year on a 500-account book with 40 account managers.
| Agent | What it does | Model | Tokens in / out per run | Passes | Cost per run | Runs per year | Token cost per year |
|---|---|---|---|---|---|---|---|
| QBR brief | Builds the quarterly review deck from usage, support and renewal data for the account manager to edit | Top tier | 60,000 / 6,000 | 3 | $1.44 | 1,337 | $1,925 |
| Renewal brief | Writes a one-page brief at each renewal milestone with health, risks and the plays to run | Top tier | 40,000 / 4,000 | 2 | $0.64 | 300 | $192 |
| Call prep brief | Writes a short brief before each customer call | Standard | 20,000 / 2,000 | 1 | $0.06 | 4,308 | $276 |
| Daily risk scan | Reads health signals across one account manager's book each morning and triages the accounts that moved | Standard | 20,000 / 1,000 | 1 | $0.05 | 9,200 | $478 |
| Call summary | Drafts post-call notes and the follow-up email from the transcript | Standard | 15,000 / 1,500 | 1 | $0.05 | 5,814 | $279 |
| Email draft | Drafts a customer email on request or when a workflow fires | Standard | 5,000 / 800 | 1 | $0.02 | 18,400 | $361 |
| Total | 39,359 | $3,511 |
The QBR brief is 55 percent of the token bill on its own, because it reads the most account history and runs three passes on the top model. Email drafts are almost half of all runs and about 10 percent of the cost.
How a run is priced
A run costs the tokens the model reads plus the tokens it writes, multiplied by the number of passes it takes. We priced the top-tier runs at OpenAI's list rate for GPT-5.6 Sol, which is $5 per million input tokens and $30 per million output tokens, and the standard runs at GPT-5.6 Terra's $2 and $12 (current list prices). The token counts per run are our own estimates from agents we have built that do the same jobs. HubSpot and Salesforce resell this capacity as credits, which we cover further down, and the token math underneath is the same.
Worked example
One QBR brief run comes to $1.44
We put the QBR and renewal briefs on the top model because they make judgment calls over a long account history and a wrong call costs a renewal conversation. Summaries, call prep, the morning scan and email drafts run on the standard model because they condense or reword data the CRM already holds. Model prices have mostly moved down this year, and OpenAI cut Terra's price in July, so these per-run numbers are more likely to fall than rise.
Hours each agent returns
The six core agents return about 11,100 hours a year on a team of 40 account managers, which is about six full-time people at 1,880 paid hours each. We set each time credit at or below the bottom of the published range wherever a published range exists, and we say where one does not.
| Agent | Manual time today (published) | Time with automation (published) | Hours we credit | Hours returned per year | Value at base salary |
|---|---|---|---|---|---|
| Daily risk scan | 45 to 60 minutes a day to review 40 accounts by hand (Gainsight) | No published figure | 15 to 30 minutes per account manager per day, our assumption | 3,958 | $210,532 |
| QBR brief | 4 to 6 hours to build a 30-slide deck (ScaleUp CS) | 30 to 45 minutes to review and adjust (ScaleUp CS) | 2.25 hours per review, 1.5 on the lightest tier | 2,510 | $133,511 |
| Email draft | 15 to 20 minutes saved per message (Gainsight); about 2 fewer hours a week on email for Copilot users across 7,137 workers at 66 firms (Microsoft and Harvard field experiment) | Included in the saving | 1 hour per account manager per week | 1,840 | $97,872 |
| Call summary | 20 to 30 minutes of notes and follow-up per call (Gainsight) | Under 5 minutes (Gainsight) | 15 minutes per call | 1,454 | $77,340 |
| Call prep brief | 30 to 45 minutes per account (Gainsight) | 5 to 10 minutes to review (Gainsight) | 15 minutes per call | 1,077 | $57,287 |
| Renewal brief | No published figure we could verify | No published figure | 1 hour per renewal, from our own builds | 300 | $15,957 |
| Total | 11,139 | $592,500 |
The value column prices an hour at a $100,000 base salary over 1,880 paid hours, which is about $53 an hour before benefits. Loading benefits at 30 percent would raise it to about $69 an hour, and we left that out on purpose so the number reads as a floor. The daily risk scan returns the most time because it runs every working day for every account manager, while the QBR brief returns the most per run.
The total for a 500-account book
Across the whole book, the six core agents cost $3,511 a year in tokens and return hours worth about $592,500 at base salary, or about $296,000 after we halve the hours for time that gets absorbed. That is roughly 84 dollars of time for every dollar of tokens on the conservative reading.
Return per dollar of tokens
Every core agent returns at least $35 of time for each $1 of tokens, after the haircut
We halve the hours because saved time does not turn into new work on its own. In the Microsoft and Harvard field experiment, people who used Copilot spent about two fewer hours a week on email and their meeting time did not change, which tells you the time came back but says nothing about where it went. A plan that counts every saved hour as capacity will overstate the result, and the haircut is there so the number survives a board review.
Three more agents in the model are retention plays: a renewal-at-risk brief, a weekly check for renewing accounts whose calls never mention the renewal, and an expansion agent that refreshes the whitespace view. They add about $410 a year in tokens, and their value shows up as renewals caught earlier and expansion qualified, so we do not count them as hours. We also left seven more agents out of the total because their time credits rest on the client's own estimates, and nobody has measured them yet.
Where the real cost sits
The costs that decide whether an agent program pays back are the build, the business rules, the data the agents read, and the change in how the team works, and none of them show up in the token number. Our own model says so in its last line, because build labor sits outside it and it is the cost a CFO should ask about.
Most of a well-built program is ordinary workflow, in HubSpot or in Salesforce. In a renewal risk program, a calculated field or a workflow detects that an account crossed a threshold, another workflow tells the account manager, and a report shows whether the save play worked. The model only runs when someone needs to know why the account moved, which means reading the support tickets, the usage trend and the last few calls. Those deterministic steps cost nothing per run and give the same answer every time, so a good design puts the model in as few steps as possible.
Renewal risk program
One step in four calls a model
Detect
A calculated field or workflow flags an account that crossed the risk threshold.
WorkflowNotify
A workflow creates a task for the account manager who owns the account.
WorkflowExplain
The agent reads the support tickets, the usage trend and the last few calls, and writes why the account moved.
ModelVerify
A report shows whether the save play worked, and the result goes back on the record.
ReportThe rules take the real effort. Someone has to decide which accounts get which play, at what threshold, in what order and at which service tier, and then tune those rules against a real book. Building the agent itself has become the easy part: an agent that would have taken a quarter to build eighteen months ago took us an afternoon in HubSpot Agent Builder this summer. On Salesforce, Agentforce needs Data Cloud configured first, and in our Agentforce work the technical setup has taken about a week while the knowledge content behind the agent has taken about a quarter.
The hours also arrive on a schedule. Every hour in the model shows up on the go-live date of the workflow that produces it, so the call summary and call prep agents return nothing until a call recorder is running on the customer success side, and the risk scan returns nothing until the health signals it reads are in the CRM. A plan that books all 11,100 hours on day one will miss its first quarter.
The same agents on HubSpot or Salesforce
Both platforms sell agent capacity as credits, and a CRM-reading agent run lands in the same range as the raw token math on either one. On HubSpot, a renewal prep agent we built in our own portal billed 92 credits for a live run, which is 92 cents at the capacity pack rate of $10 per 1,000 credits. On Salesforce, Agentforce bills 20 Flex Credits per action at $500 per 100,000 credits, which is 10 cents an action, so the cost of a run depends on how many actions the agent takes.
Platform pricing, October 2026
Both platforms price agent work in credits that cost a cent or less each
HubSpot
- Unit price
- $10 per 1,000 HubSpot credits in capacity packs (HubSpot)
- What a run costs
- 92 credits billed for one live run of our renewal prep agent, about $0.92
- Included
- 500, 5,000 and 10,000 credits on Starter, Professional and Enterprise (HubSpot pricing)
- Before you spend
- Test runs in Agent Builder consume no credits, and a monthly run limit caps the bill
Salesforce
- Unit price
- $500 per 100,000 Flex Credits (Agentforce pricing)
- What a run costs
- 20 Flex Credits per Agentforce action, which is $0.10 an action, so a four-action run is $0.40
- Included
- 500,000, 1 million and 2.75 million Flex Credits with the Core, Advanced and Max editions (Salesforce)
- Before you spend
- Agentforce needs Data Cloud configured first, and the knowledge content takes longer than the setup
We wrote up the HubSpot side in detail in What a HubSpot CS Agent Actually Costs to Run, including how to forecast credits before spending any. On HubSpot, a monthly run limit set before the agent goes live turns a variable cost into a fixed ceiling that finance can sign.
Running the math on your own book
Any customer success team can rebuild this model in a spreadsheet in an afternoon, and the steps below are the ones we follow.
- List each agent with the event that triggers it, and count how often that event happens in a year on the book: reviews, customer calls, renewals, and working days for each account manager.
- Price one run by estimating the tokens the agent reads and writes, choosing the model tier, and multiplying by the number of passes. On HubSpot, run the agent in test mode against real accounts and read the estimated credits per run in the run limits panel. On Salesforce, count the actions the agent takes in a run and multiply by 20 Flex Credits.
- Credit the time saved per occurrence at or below the low end of a published range, and write the source next to every number.
- Value the hours at base salary and halve them for absorbed time before anyone outside the team sees the total.
- Give each workflow a go-live date and start counting its hours from that date.
- Put the build labor, the rules design and the data work next to the token cost, because that is the comparison a board will ask for.
We run this model on a client's own book as part of our two-week CS AI Diagnostic, alongside the process review that decides which of these agents are worth building at all. If you want to see it on your numbers, book a call with us.
Frequently asked questions
How much does an AI agent cost to run for a customer success team?
In our October 2026 model, a customer success agent cost between about 2 cents and $1.44 per run in model tokens at OpenAI's list prices. Six core agents cost $3,511 a year on a book of 500 accounts with 40 account managers, which is about $7 per account and $88 per account manager.
Which customer success AI agent costs the most to run?
The QBR brief costs the most at $1.44 per run, because it reads about 60,000 tokens of account history and runs three passes on a top-tier model. It made up 55 percent of the token cost for the six core agents in our model.
How many hours can AI agents save a customer success team?
In our model, scaled to a 500-account book, six core agents returned about 11,100 hours a year to a team of 40 account managers, which is about six full-time people. After halving the hours to allow for time that gets absorbed, the conservative figure is about 5,600 hours.
What does a customer success agent cost in HubSpot or Salesforce credits?
A renewal prep agent built in HubSpot Agent Builder billed 92 credits for a live run in our portal, which is 92 cents at HubSpot's capacity pack rate of $10 per 1,000 credits. Salesforce Agentforce bills 20 Flex Credits per action at $500 per 100,000 credits, which is 10 cents an action.
What is the biggest cost in an AI agent program for customer success?
The biggest costs are the build, the business rules the agents run on, getting the data they read into the CRM, and changing how the team works once hours come back. Model tokens came to about $3,500 a year for six agents on a 500-account book in our model.
Does every step of a customer success program need an AI model?
Most steps in a customer success program run as ordinary workflows that detect a condition and notify an owner. The model is needed only where someone has to read unstructured material such as tickets, call transcripts or email to explain what changed and what to do next.
About the author
Jeff Kushmerek is the founder and CEO of Infinite Renewals, a Boston customer success consulting firm and HubSpot Solutions Partner that builds agents and workflows in HubSpot and Salesforce. He has spent 25 years running customer success and professional services teams, including a professional services organization that went public in 2012, and he wrote Retention Starts in Implementation and the HubSpot Customer Success Workspace Operators Manual. Prices and figures in this post are as of October 22, 2026.