Chapter 6: Build your operating plan

Make the controls someone’s responsibility

You now have the pieces. The last job is to make them someone’s responsibility, because an unowned dashboard and an unwritten exception process will not survive the next expensive month.

Put shared controls in one place

An AI gateway is a service that sits between your coding tools and the model providers. For tools that support it, the gateway applies shared identity, model access rules, and spend controls in one place. It is useful infrastructure, and it is also where Databricks and Coinbase put their routing and budgets. It is not a prerequisite for learning where your money goes.

Start by combining each tool’s usage export under consistent user and team identities. Move compatible traffic behind shared enforcement as you need it. Some tools bill separately or expose limited telemetry, so record those gaps instead of claiming one gateway sees everything. And do not expect to negotiate your way out: there is almost no volume discount on consumption pricing.

A gateway sees requests. Your issue tracker and your repository see delivery. Join the two before you evaluate outcomes. And restrict access to session traces with a defined retention window, because traces contain proprietary source code and sensitive task context.

Spot the policies working against you

Tick the traps you recognize. They feed your scorecard, and they are diagnostic prompts, not a ranking of your team.

• Usage leaderboards. Rewarding tokens consumed rewards the wrong thing.
• One monthly cap doing two jobs. Too small to permit normal work or too large to catch a runaway afternoon. See the budget policy.
• Permanent limit increases. One expensive project becomes a lifetime entitlement. Time-box every raise.
• Suspension as the first response. Cutting a developer off on the 24th interrupts work and produces exceptions. Downshift first.
• Public benchmarks as the adoption decision. They don’t predict your codebase. Build the eval.
• Per-call optimization. Per-call savings that raise per-task cost are losses.
• Seat-based budgeting. Seat count misses how much work each agent performs. Separate fixed fees from variable consumption.
• Token totals as the board metric. They say what the tool charged, not what it bought.
• Annual commitment before measurement. Collect a representative usage baseline before locking in a large commitment.
• Per-tool budgets with no total. Controls each bill, not the sum, and pushes engineers toward whichever tool has headroom.

Check what is actually in place

The exercises tested your understanding. These sixteen checks test your operating setup. Tick only what you can point to today. A strong quiz score does not mean the infrastructure exists yet.

Make the next 90 days concrete

Run the first model evaluation during the baseline period. Do not spend two months installing controls before you ask whether the default model is good value. And sign the annual contract after the 60 to 90 day baseline, not before. It is the cheapest insurance in this guide.

The decision you should be able to make

Pick one expensive workflow. Explain what it costs, what it delivers, and what you would change next. Then show how a team can get extra capacity for valuable work without inheriting a permanent blank check.

That is the whole job. Keep useful work moving. Make waste visible. Give every improvement a test and an owner.

Key terms

Further reading