How to Scale AI Agents Without Losing Control of Costs

Scaling AI agents starts with a question your team should be able to answer before adding more users: what does it cost to complete a useful piece of work?

On October 5, Cohere introduced North 2, highlighting multi-step agent orchestration alongside spending controls. Its announcement describes request and token consumption tiers, user quotas, organization-wide caps, and alerts. It also lists reusable skills, shared libraries, memory, and human oversight for critical actions among the platform’s capabilities.

Those controls deserve attention. As agents take on more steps, the economics depend on everything required to finish the job: software, connected services, failed attempts, and the people checking the result.

A useful rollout plan makes those costs visible before expanding usage.

Watch Cohere’s official North 2 video for a look at the product. Consider which capabilities would improve your workflow and what they would cost to run.

Measure the cost of accepted work

Tokens are the units of text models process and generate. Tracking them helps explain consumption, but it gives you only part of the cost picture.

Consider an agent that prepares account briefs for a sales team. A finished brief needs accurate information, relevant context, and enough detail to support a customer conversation. If someone has to rebuild it, the workflow has consumed resources without delivering the intended result.

Define an accepted outcome before you start measuring. For this example, it could be a brief that passes an agreed factual check and is ready for the account owner to use.

Then calculate:

Cost per accepted outcome = total workflow cost divided by accepted outcomes.

Include:

• Platform charges and model usage allocated to that workflow.

• Connected tools, data access, and infrastructure costs.

• Human review, corrections, and exception handling.

• A reasonable share of implementation, training, and ongoing maintenance.

Keep the costs of failed and abandoned attempts in the total. Count each accepted outcome once, even if it took several attempts.

A simple illustrative example

Assume a pilot attempts 100 account briefs and produces 80 accepted briefs in a week. These figures are hypothetical, not vendor pricing or a HiVoltage result.

• Software, model, and connected-service costs total $100.

• Review and rework take four hours, valued at $50 an hour, or $200.

• Allocated implementation and maintenance add $50.

The total is $350, or about $4.38 per accepted brief.

Compare that figure with the existing process at the same quality standard. Also check turnaround time and whether the team uses the briefs. Saved employee time creates capacity; it becomes a cash saving only when spending changes.

Match the workflow’s scope to its value

An agent can keep searching, checking, revising, or calling another tool. Each additional step needs a reason.

Anthropic’s guidance on building effective agents recommends starting with the simplest approach and adding complexity when it improves the result. It also notes the cost and latency tradeoffs of agentic systems.

Apply that thinking to the account-brief example. Define which sources the agent can use, how much detail the brief needs, and what should happen when information is missing. A clear instruction to flag an unresolved fact may be more useful than repeated searches that still fail to verify it.

For custom AI agents, those boundaries belong in the design brief. Test simpler approaches alongside the agent. A fixed sequence may handle predictable work well. More flexible reasoning may earn its cost when the task varies substantially. Let measured results guide that choice.

This is also where reusable instructions and shared knowledge can help. Treat them as assets that need testing and upkeep. Reusing an outdated instruction across more workflows can multiply the correction work.

Use spending limits at several levels

A company-wide cap helps contain total exposure. Teams also need limits close enough to the work to explain where money is going.

For a pilot, set:

• A workflow budget tied to expected volume and an acceptable cost per outcome.

• A run limit for elapsed time, tool calls, or retry attempts, using controls the system supports.

• An early warning threshold that leaves time to investigate before reaching the cap.

• A fallback for unfinished work when a limit is reached.

Decide what happens at the threshold. Does the workflow pause, reduce optional research, or hand the task to a person? Test that behavior before relying on it.

Cohere’s announcement provides a timely example of platform-level cost controls. It does not publish a North 2 price list or establish the return your organization will achieve. Ask prospective vendors how usage is billed, which connected-service charges are separate, and how limits affect work already running.

Run a pilot that can support a scale decision

A small pilot should give you evidence about ordinary work and the awkward cases your team handles every week.

1. Record the current process. Measure time, direct costs, quality, and rework for a representative set of tasks. Include the people doing the work in defining what counts as acceptable.

2. Choose one bounded workflow. Start with a clear input and finish point. Keep optional additions out of the first test so you can see which parts create value.

3. Track the whole attempt. Record software cost, completion status, elapsed time, retries, and human review. Group similar tasks together so easy cases don’t hide expensive exceptions.

4. Compare under consistent conditions. Test changes to models, instructions, and tools against the same cases and quality requirements. Anthropic’s agent evaluation guidance describes using a stable set of tasks to track cost, speed, and error rates over time.

5. Expand when the evidence supports it. Increase volume in stages. Watch both cost per accepted outcome and total spend. A lower unit cost can still produce a larger bill when usage grows.

Set the decision criteria before the pilot. You might require acceptable quality, manageable review time, and a delivery cost below the current process. If the workflow improves speed at a higher cost, decide whether that speed is worth paying for.

Keep human review in the economics

Review is part of producing reliable work. Budget for it openly.

The important question is where human attention adds value. Checking a consequential recommendation may be worth the time. Repeatedly fixing the same formatting problem suggests the workflow needs improvement. Track those separately so the team can reduce avoidable rework while protecting necessary judgment.

Before the next expansion, ask the people using the output what still slows them down. Their answers may explain a cost problem that a usage dashboard misses.

HiVoltage helps teams design agentic workflows around business priorities and human review. Start with one workflow, its full delivery cost, and a clear test for whether it’s ready to scale.