Founders running real AI agent workloads now show boards a burn number with no structure behind it, and that single gap is costing them credibility in the room.
- Track AI agent spend as a per-task unit cost, not a lump monthly line item
- Separate token spend from inference compute and human review time in your board deck
- Cursor's engineering team has published that it costs roughly $20 to $40 in model spend per resolved PR-sized task
- Build a three-scenario forecast tied to usage growth, not a single flat burn line
- Flag the agent cost curve against revenue per customer so the board sees the ratio, not just the total
Here's how it usually goes. A founder walks into a board meeting, flips to the slide with "AI costs," and shows one number: $340,000 last quarter, up from $190,000 the quarter before. A board member asks why. The founder says usage grew. Nobody in the room can tell if that's a healthy trajectory or a warning sign, because there's no unit behind the total.
That's the gap. Most founders are now running meaningful AI agent spend, whether that's a coding agent, a customer support agent, or an internal ops agent chaining multiple model calls together, but almost nobody has standardized how to present that spend to a board. Sales has pipeline coverage ratios. Engineering has velocity metrics. AI agent spend just gets a dollar figure and a shrug.
A board doesn't actually want to know what you spent. They want to know whether what you spent is getting cheaper or more expensive per unit of value delivered, and whether that trend is one you control. A raw total answers neither question. $340,000 could mean your agent handled 50,000 tasks at $6.80 each, down from $9 the quarter before, which is a great story. Or it could mean the same 20,000 tasks at nearly double the per-task cost, which is a very different conversation.
Replit ran into a version of this problem publicly in 2025, when a runaway agent session during a product demo burned far more compute than the team expected, and the story became less about the dollar figure and more about the lack of per-session caps. The lesson wasn't that AI agents are too expensive to use. It was that nobody had built a unit-level forecast that would have caught the anomaly before it hit the bill.
Why Does My AI Coding Agent Bill Keep Increasing Overnight Why does my AI coding agent bill keep increasing when nobody touched the keyboard overnight? Usually it's a runaway tool-call loop compounding token costs with every retry. This guide walks through the mechanism, a real incident involving Replit's agent, and the specific guardrails that stop it. - why AI coding agent bills spike overnight - preventing runaway tool call loops in agents
So the fix isn't a bigger spreadsheet. It's a different clause structure entirely, one that breaks total spend into components a board can actually reason about.
The three-line structure that replaces a single burn number #
Instead of one AI cost line, present three, and let the board watch how they move relative to each other quarter over quarter.
The first line is raw inference spend: what you pay OpenAI, Anthropic, or whichever model provider you're metered against, broken out by model tier if you're mixing a cheap model for routing and an expensive one for final generation. Anthropic's own usage-based pricing for Claude, billed per million input and output tokens, makes this line straightforward to pull directly from your API dashboard rather than estimate.
The second line is orchestration and retry overhead: the extra calls your agent makes when it fails a task and retries, when it calls a tool that errors out, or when a multi-step chain has to backtrack. This number is almost always higher than founders expect on first measurement, often 20 to 40% on top of the "successful" task cost, and it's the line a board should watch most closely because it's the one engineering can actually reduce without cutting agent capability.
The third line is human review time, costed at loaded hourly rate, for any agent output that still needs a person to check it before it ships. This is the line most decks leave out entirely, and it's the one that quietly determines whether an agent is actually saving the company money or just moving the cost from a salary line to a cloud bill.
Once you have those three lines, the board deck metric that matters is cost per completed task, not cost per month. Cursor's engineering team has written publicly that resolving a PR-sized coding task with their agent runs in the range of $20 to $40 in model spend depending on task complexity and how many iterations the agent needs. That's a number a board can track quarter over quarter and immediately understand whether it's trending the right way.
Building the forecast, not just the historical line #
A backward-looking cost breakdown is useful, but a board deck clause needs a forecast, and the forecast is where most founders get lazy and just extrapolate last quarter's growth rate in a straight line. Don't do that. Build three scenarios tied to usage drivers you actually control or can measure.
How to Audit AI Agent Usage Before You Walk Into Renewal A startup sells an AI agent into an enterprise account. Eighteen months later, renewal lands on the calendar and the founder walks in with a slide about the roadmap. The customer's procurement lead walks in with a spreadsheet showing forty of the two hundred seats never logged in. - how to audit AI agent usage before renewal - what to do before enterprise software renewal negotiations
The base case assumes your current tasks-per-customer ratio holds and your customer count grows at the rate sales is already forecasting. The high case assumes a specific, named catalyst, a new feature that triggers more agent calls per session, or a planned price tier that unlocks heavier usage, and shows the cost curve under that catalyst. The low case assumes you ship the efficiency work already on the roadmap, like swapping a reasoning-heavy model for a smaller one on routine subtasks, and shows the cost curve bending down even as usage grows.
The point of three scenarios isn't to hedge. It's to show the board you know which levers move the number and that you're already pulling on them. A single flat-line extrapolation tells the board nothing about your judgment. Three scenarios tied to named product and engineering decisions tell them everything.
Tie the forecast to one more ratio before you close the slide: AI agent cost per customer against revenue per customer. If a customer pays you $200 a month and your agent spend against their account is $340 a quarter, or roughly $113 a month, you have a margin problem hiding behind a usage-growth story, and the board needs to see that ratio directly rather than infer it from two separate slides.
Frankly, most boards don't ask for this breakdown because most founders haven't shown them it's possible. The first founder in a portfolio to bring a three-line, three-scenario AI cost clause into a deck sets the bar every other portfolio company gets measured against in that same boardroom. That's not a hypothetical. Investors sitting on multiple boards talk to each other, and a clean framework travels fast once one portfolio company adopts it.
Build the clause once, in a format you can reuse every quarter: three cost lines, one unit metric, three forward scenarios, one ratio against revenue. Fill it in with real numbers pulled straight from your model provider's billing dashboard and your own task-completion logs. Don't wait for a board member to ask why the AI line jumped. Show them the unit economics before they have to ask.
Also read: How to Negotiate a Most Favored Nation Clause in Your First Enterprise Deal • How to Forecast AI Agent Usage for Pricing When Nobody Can Predict Volume • How to Build an AI Agent Usage Audit Trail Before Security Asks
This article is posted in Startup News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.