How to track AI spend at a startup before it gets out of hand
AI costs start as a few $20 seats and end as token bills nobody owns. How startups can see every AI cost, tie it to real work, and keep it in check.
5 min read
Key takeaways
- AI spend at a startup grows in small, reasonable steps: a few chat seats, a bigger plan for the heaviest user, then API and agent usage billed by the token.
- Start with an inventory of every AI subscription, seat, and API account, with the card it's billed to, its owner, and the job it does.
- Check usage-based AI costs every week and seats every month. Cancel anything nobody can tie to real work.
- Set a spend limit and an alert in every AI provider console you use, so a runaway script can't quietly run up the bill.
- In Opser, every agent, model, and paid tool draws from one credit balance, with usage by agent, an alert at 80% of the monthly credits, and a pause at zero.
There's a stage in every startup where things are going well enough that nobody reads the bills closely. AI spend loves that stage.
It starts with a $20 chat subscription. Soon everyone on the team has one. The person who uses it most moves to the $200 plan, an engineer adds an API key that bills by the token, and an agent starts running a job every morning. Every one of those steps was reasonable, and nobody ever decided to spend a few thousand dollars a month on AI. The bill got there anyway.
Growth hides this kind of drift. When revenue is coming in, you can afford to be wrong for longer, and small costs compound while everyone is busy. Getting it back under control takes a list, an owner for each line, and half an hour a month. Here's the system we recommend to small teams, and how Opser handles the AI spend that runs inside it.
Why does AI spend get out of hand at startups?
Most software costs are flat. You pay per seat, the price changes once a year, and the invoice looks the same every month. AI breaks that pattern in two ways.
Seats multiply quietly. Each person picks the assistant they like, often on their own card, and a seat rarely gets cancelled when someone stops using it. Plans creep upward too, because the heaviest users hit limits and upgrade, and every upgrade looks small next to the time it saves.
Usage billing is the bigger risk. API calls, coding agents, and automations are priced by the token, so the cost depends on how much work they do, which model they use, and how long they think. A script stuck in a loop or an agent switched to a bigger model can multiply the bill overnight, and nobody sees it until the invoice arrives. On top of that, the charges land on different cards, expense reports, and invoices, so nobody sees the total.
Where does AI spend hide?
| Where it hides | What it looks like | How to catch it |
|---|---|---|
| Seats | A chat subscription per person, including people who stopped using it | Compare seats with who actually used the tool last month |
| Plan upgrades | The heaviest user moves from the $20 plan to the $200 one | Write down who upgraded and why when you approve it |
| Usage billing | API keys and model calls billed per token | A weekly check, plus a spend limit in the provider console |
| Agents and automations | Scheduled jobs that call models while nobody watches | Usage per agent and an alert before the balance runs out |
| AI add-ons | AI features bolted onto tools you already pay for | Read the line items on every renewal invoice |
| Personal cards | Tools someone expensed, or never expensed | Ask the team for a list once a quarter |
How do you track AI spend as a small team?
A spreadsheet and a monthly habit are enough.
1. Make an inventory. One row per AI tool or account, with the plan, the number of seats, the monthly cost, whether it's a fixed price or billed by usage, the card it's on, its owner, and the job it does. Pull the rows from card statements, your expense tool, and your inbox, since receipts and invoices land in email. If your mailbox is connected to Opser, your agent can do the digging: "Find receipts and invoices from AI tools in the last 90 days and put them in a spreadsheet with the vendor, amount, date, and card."
2. Give every line an owner. Somebody has to be able to say why it exists. If you can, move AI tools onto one company card, so the next statement already is your inventory.
3. Tie each line to output. Next to every tool, write the job it does: drafting proposals, writing code, answering support email. If nobody can name the job, cancel it. If two tools do the same job, keep the better one.
4. Put a ceiling on usage billing. Set a monthly spend limit and an email alert in every AI provider console that bills your API keys. It takes a few minutes per provider, and it turns a runaway script from a surprise into a notification.
5. Review on a schedule. Usage-based costs every week, since they can jump overnight. Seats and subscriptions every month, when you can compare them with who used them. It's the same idea as an operations audit, applied to one fast-moving cost.
How do you keep AI agents from running up the bill?
Agents are the newest line on the bill and the hardest to see, because they work when nobody is watching. Give each agent a clear job and run it on the cheapest model that does that job well. Then look at what it costs next to what it produced.
This is where Opser does the bookkeeping for you. Every agent, model call, and paid tool in a workspace draws from one shared credit balance, priced in plain credits where 100 credits is $1, and each plan includes a monthly allowance. Admins can open Settings, then Usage, and see where the credits went by model, by tool, and by agent, for the last day, week, month, or all time. When the workspace has used 80% of its monthly credits, admins get an alert. When the balance reaches zero, agents pause until someone tops up, so a runaway job can't keep spending.
Each agent also has its own model and reasoning setting. The agent that sorts the inbox can run on a smaller, cheaper model than the one that drafts proposals, and image and video generation ask before they run by default. If your team pays for ChatGPT, you can link that subscription and run Codex models on your ChatGPT plan instead of credits.
When you want to know what the agents did with those credits, ask: "What did our agents do this week?" Your agent reads the workspace's activity log and answers with the actions they took, app by app.
What should a monthly AI spend review cover?
Block thirty minutes on the first Monday of the month and go through the same list every time:
- seats against the people who actually used each tool last month
- plan upgrades approved in the last month, and whether they paid off
- usage-based costs compared with the month before, starting with the biggest jump
- the agents that cost the most, next to what they produced
- one thing to cancel, downgrade, or move to a cheaper model
The last item matters most. A review that never cuts anything turns into a ritual, and the drift comes back.
The best time to do this is while things are going well enough. A $200 line is easy to question. Once the same habit has grown into a $5,000 one, it's tangled up with how half the team works.
Frequently asked questions
- What counts as AI spend at a startup?
- Chat subscriptions and seats, plan upgrades for heavy users, API usage billed by the token, AI add-ons inside tools you already pay for, and agents or automations that call models on their own.
- How often should a startup review AI costs?
- Look at usage-based costs like API tokens every week, because they can jump overnight. Review seats and subscriptions once a month, when you can compare them with who actually used them.
- How do I stop an AI agent from running up a bill?
- Give it a ceiling it can't pass. Set spend limits in the provider console for anything that runs on an API key. In Opser, agents draw from the workspace's credit balance, admins get an alert at 80% of the monthly credits, and agents pause when the balance reaches zero.
- Can I see AI usage per agent in Opser?
- Yes. Admins can open Settings, then Usage, to see where credits went by model, by tool, and by agent, for the last day, week, month, or all time.
- Can my team use its ChatGPT subscription inside Opser?
- For Codex models, yes. Link a ChatGPT subscription under Profile, in Linked accounts, and agents that use those models run on your ChatGPT plan instead of Opser credits.
Founder · Senior Software Engineer
Simone Zamparini is a founder and senior software engineer.
