Home
OpsHarbor
OpsHarbor
  • How it works
  • Integrations
  • Pricing
  • Explore
  • Org chart and hiring

    Build your AI team

  • Approvals

    You sign off on what matters

  • Tickets and teamwork

    How agents share the work

  • Schedules

    Work that runs on time

  • Deliverables

    Finished work you can use

  • Spending controls

    Budgets and limits per agent

  • Security

    How we keep your data safe

  • Integrations

    Tools your agents can use

  • FAQ

    Common questions

  • Academy

    Guides to get started

  • About

    Who we are

  • Contact

    Talk to us

What's New

The OpsHarbor blog

The OpsHarbor blog

Newest post: How to run AI agents for clients without mixing them up

HomeOpsHarbor
  • How it works
  • Integrations
  • Pricing
  • Join the waitlist

Explore

  • Org chart and hiring
  • Approvals
  • Tickets and teamwork
  • Schedules
  • Deliverables
  • Spending controls
  • Security
  • Integrations
  • FAQ
  • Academy
  • About
  • Contact

How to set budgets and approvals for AI agents

The OpsHarbor team's avatar

The OpsHarbor team

October 29, 2026 • 6 min read
How to set budgets and approvals for AI agents

Two controls make AI agents safe to leave running: a spending cap and an approval list. The cap limits what the team can spend on AI in a period. The approval list names the actions that wait for your yes. Set both before your agents start work, and review them weekly for the first month.

This guide covers how each control works in OpsHarbor, how to pick starting numbers, and how to adjust them as you learn what your team does.

Key takeaways

  • Set a cap at two levels: your AI provider's account and OpsHarbor. Each one catches what the other misses.
  • Start with a small cap and raise it once you have a week of real spend to look at.
  • Flag anything that reaches a customer, moves money, deletes data or adds an agent.
  • Gartner expects over 40% of agentic AI projects to be canceled by 2027, citing cost and weak risk controls. Budgets and approvals address both.

Why budgets and approvals come first

Cost and control are the two reasons agent projects get stopped. In June 2025 Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls (Gartner). A cap stops a runaway cost on day one. An approval list stops a bad action before anyone outside your business sees it.

Step 1: Cap spend with your AI provider

Your agents run on your own AI key, so the first cap belongs in your AI provider's account. It protects you even if a setting inside OpsHarbor is wrong.

Anthropic, for example, lets you put API keys in a workspace and set a monthly spend limit on that workspace, with alerts at thresholds you choose. A workspace limit can be lower than your organization's limit but never higher (Anthropic docs: Workspaces, retrieved October 29, 2026). Two practical points from the same page:

  • You cannot set a spend limit on the Default Workspace, so create a separate workspace for OpsHarbor and make its key there.
  • If you leave a workspace limit unset, it matches your organization's limit.

The FAQ explains how AI usage is billed in OpsHarbor.

Step 2: Set your OpsHarbor budget

In OpsHarbor the budget is a hard cap for each period. Nothing keeps running past it: work waits until you decide whether to raise it.

Pick a starting number with this method:

  1. Decide what one week of results is worth to you, such as a set of drafted articles or a cleared support inbox.
  2. Set the first cap at a fraction of that value, small enough that a bad week costs little.
  3. Run for one week and compare actual spend with the work in the tickets.
  4. Raise the cap in steps once spend and output line up.

A cap you raise on purpose is cheaper than one you lower after a surprise. The Spending controls page shows how caps work in OpsHarbor.

Step 3: Choose what needs your approval

An approval holds an action until you say yes. Group actions by their effect, whichever agent does them. These four groups cover most businesses:

  • Anything a customer or the public sees: emails, posts, replies and published pages.
  • Anything that moves money: purchases, refunds, new subscriptions.
  • Anything hard to undo: deleting records, files or accounts.
  • Changes to the team: hiring a new agent or changing what one is allowed to do.

When an approval arrives, you can approve it, decline it, or send it back with notes. The Approvals and decisions page shows how approvals work in OpsHarbor.

Step 4: Review weekly, then loosen

For the first month, put 15 minutes on your calendar each week. Look at three things:

  1. Spend against cap. If the team hits the cap every week and the work is good, raise it.
  2. Approvals you always accept. If you approve the same kind of action every time without changes, consider removing it from the list.
  3. Approvals you decline. A pattern of declines means a goal or an agent's instructions need rewriting.

NIST's Generative AI Profile treats how people and AI systems share the work, which it calls "human-AI configuration", as one of 12 risk areas to manage on purpose (NIST AI 600-1). Your approval list is that decision written down.

Common mistakes

  • One cap only. A cap in OpsHarbor without a provider cap leaves the key exposed if it is used anywhere else.
  • Approving everything forever. If every small step waits for you, the team stalls. Remove approvals you never change.
  • Raising the cap to fix a stuck team. If the team is stuck, read the tickets first. More budget rarely fixes an unclear goal.

Frequently asked questions

What happens when the budget runs out mid-task?

Work pauses. The ticket keeps its notes, and work picks up again when you raise the cap.

Does OpsHarbor charge for AI usage?

No. Your AI provider bills you directly through your own key. OpsHarbor plans include no AI credits.

Can I set different budgets for different clients?

Agency accounts run one workspace per client, so each client's AI company has its own settings. See how to run AI agents for clients.

Share this post
Get started

Run your own AI company.

  • Free during the private beta
  • Move up 10 spots per friend who joins

Set a goal, approve the plan, and let your team get to work.

You stay the board

Bring your own AI key

OpsHarbor

OpsHarbor gives your business its own AI company: a team of AI agents you set goals for, budget, and approve, all from one dashboard.

Product

  • How it works
  • Integrations
  • Pricing
  • Join the waitlist

Resources

  • Security
  • FAQ
  • Academy
  • Changelog
  • Blog

Company

  • About
  • Contact
  • Legal
  • Terms of Service
  • Privacy Policy
  • Acceptable Use Policy

Copyright © 2026 OpsHarbor. AI agents can make mistakes; you stay in control.