There is a specific way software costs money now, and it is unlike anything most small businesses budgeted for.
The old model was a seat. You paid a number per person per month, you knew it in advance, and the only way it changed was if you hired someone. It was boring and you could forecast it a year out.
Notion Custom Agents work differently. They bill by the run, they run on their own schedule without anyone asking, and one team reportedly consumed 150,000 credits in a single month. At the current rate that is $1,500, on top of their subscription, for software doing exactly what they configured it to do.
What Custom Agents actually do
Notion launched Custom Agents on 24 February 2026, and the capability is genuinely a step beyond the AI assistant most people picture. These are not chat helpers waiting for a prompt. You give an agent a job and a trigger, and it works without further instruction.
They operate across Slack, Mail, Calendar, Figma, and Linear as well as inside Notion itself, which means they reach into the tools where work actually happens rather than being confined to a single app. The typical uses are task triage, answering internal questions, running daily standups, and producing status reports, all of which are real recurring work that nobody enjoys doing.
Notion built them with reasonable governance, which deserves credit. You control what each agent can access, you can disable any of them at any time, and every run is logged so changes are visible and reversible. Compared to the over-permissioned agent problem we covered in the Anthropic breach article, that is a considerably more careful default than the industry norm.
There is also now a public beta of Notion Agent APIs, which lets you talk to a Custom Agent from Slack bots, internal tools, or mobile clients. That extends the reach further and, relevantly for this article, extends the number of ways a run can be triggered.
How the credits actually work
Custom Agents were free to try through 3 May 2026. From 4 May they moved to usage-based billing through Notion Credits, at $10 per 1,000 credits, and they are available only on Business and Enterprise plans.
Credits pool across your entire workspace rather than being allocated per person, they are billed monthly, and they reset at the start of each billing cycle. That last point matters more than it sounds: there is no rollover. Unused credits vanish, so there is no advantage to buying ahead and no reservoir to cushion a heavy month.
Consumption scales with complexity rather than being a flat rate per run. Longer instructions, more tool calls, and more database reads all push the number up, which means the same agent can cost meaningfully different amounts depending on what it encounters. Conservative estimates put a typical run at roughly $0.11 to $0.33.
Those per-run numbers are the reason this catches people. Eleven cents feels like nothing. Eleven cents is nothing. The problem is that autonomous agents are defined by running without anyone asking, which means the multiplier is not something you decide, it is something the trigger decides.
How a team spends $1,500 without noticing
The reported case of a team consuming 150,000 credits in one month is worth reasoning through, because the mechanism is entirely ordinary and nobody in that story did anything foolish.
At the midpoint of the estimated range, roughly $0.22 a run, 150,000 credits is somewhere near seven thousand agent runs. That sounds enormous until you distribute it. Six agents, running hourly across a working month, is already around three thousand runs before anyone triggers anything manually. Add agents that fire on events rather than schedules, such as every new Linear issue or every inbound email, and the count becomes a function of how busy your business was rather than how you configured anything.
That is the structural trap. With a per-seat tool, a busy month costs the same as a quiet one. With event-triggered agents, your best month is also your most expensive month, and the bill arrives after the fact with no natural checkpoint in between.
The absence of rollover compounds it. A business with variable workload cannot bank credits during quiet periods to absorb busy ones, which means the pricing model is least forgiving exactly when your business is doing well. None of this is hidden or unfair, and all of it is easy to miss when you are configuring something helpful on a Tuesday afternoon.
When they are genuinely worth it
The cost framing should not obscure that this is a good product solving real problems, and the arithmetic often works comfortably in its favour.
A daily standup agent producing a written summary that would otherwise cost twenty minutes of someone's morning is trivially worth thirty cents. Run across a working month that is perhaps seven dollars against roughly seven hours of recovered attention. There is no version of that comparison where the agent loses, and it applies to most scheduled, predictable, genuinely repetitive work.
Internal question answering is similarly strong. An agent that handles where-do-I-find-this questions inside Slack saves interruptions, and interruptions are expensive in a way that does not show up in any budget. The value there is real even though you will never be able to put a precise figure on it.
Where it stops working is high-frequency event triggers on low-value work. An agent firing on every inbound message to perform a trivial categorisation is the pattern that produces surprising invoices, because the volume is set by the world rather than by you and the per-event value is close to zero. The question worth asking before configuring any trigger is not whether the agent is useful, but what happens to the bill on a month when volume triples.
How to control the spend
The practical measures here are unglamorous and take about twenty minutes.
Start by finding your current consumption rather than estimating it. Notion exposes credit usage, and the majority of teams running agents have never looked. The number is either reassuring, in which case you have stopped worrying permanently, or it is not, in which case you have found something worth acting on. Either outcome beats a guess.
Then audit your triggers specifically, not your agents. Which fire on a schedule and which fire on events? Scheduled agents have a cost ceiling you can calculate in advance: runs per day multiplied by days multiplied by per-run cost. Event-triggered agents have a ceiling set by your business volume, which is precisely the thing you cannot control, so those are the ones that deserve scrutiny.
Finally, question hourly schedules honestly. Most agents configured to run hourly were configured that way because hourly felt responsive, not because anyone needed hourly. Moving a status-report agent from hourly to twice daily cuts its cost by roughly 85% and, in most businesses, nobody notices any difference in usefulness. That single change frequently resolves the entire problem.
The pattern this belongs to
Notion is not doing anything unusual here, which is the reason this article matters more than one product's pricing page.
The same sequence has now played out repeatedly this year. A capability launches free during a build-and-test window, businesses wire it into their operations, and the meter starts once switching away has become inconvenient. Meta Business Agent did exactly this on 1 August, which we covered in the Meta pricing change. Notion did it on 4 May. The pattern is now the norm rather than the exception.
What makes agent pricing genuinely different from earlier software billing is that the thing consuming your budget operates without you. A per-seat tool costs what it costs whether anyone opens it. A metered agent costs more precisely when it is being most useful, which inverts the intuition every business owner built over a decade of SaaS subscriptions.
The habit that handles all of it is small. When you enable anything that runs on its own, work out what it costs at ten times the volume you expect, and write down where you would look to check. That takes two minutes at setup and it is the entire difference between a tool you chose to keep paying for and an invoice that explained itself to you afterwards.