ConvoMargin What a turn costs, what it earns
A turn is one user message and one model reply. ConvoMargin works out what a turn costs you, compares it to what that conversation earns, and shows what is left. That leftover is your net margin per turn. Book a paid Audit Sprint now, or join the waitlist for the tool that watches the number for you.
Three steps: add up what a turn costs, work out what the conversation earns, then subtract one from the other. You end up with a dollar figure per turn instead of one large monthly bill. The steps are the same whether you buy an audit or wait for the monthly tool.
Every turn burns tokens. Retries and tool calls burn more. We read that from your provider logs, an API proxy, or the calculator below, so the cost sits on the turn instead of disappearing into a monthly invoice.
Use the revenue one conversation brings in, or take the plan price and divide it by the number of conversations that user has. Either way you get dollars per conversation, which we then split across the turns inside it.
What is left after cost is your margin. Positive means the turn paid for itself. Negative means you are buying that conversation for the customer. We report the number; your team decides what to change.
TLDR: Cost of a turn, minus what the conversation earns, equals your margin. See pricing.
Three posts on the three places AI spend usually hides: your subscription plans, your company seats, and your support queue. They are worked examples built from published prices, not from customer data of ours. Full articles live at /blog/.
SaaS Founders
Ordinary software costs about the same whether one customer uses it or a thousand do. AI does not. Every extra prompt costs real money, so a flat monthly price with no usage limit is a bet you have not priced.
Read the postFinance and ops
Engineering runs long automated loops while sales barely logs in, and one combined invoice hides both. Work out the cost per prompt by team before you sign the renewal.
Read the postCustomer Support
Every extra turn re-sends the whole conversation, so a long support chat costs far more than a short one. Left unwatched, free-tier users can cost more than they will ever pay.
Read the postTLDR: Read a post, then run your own numbers in the calculator. All posts.
Two ways to get the same number. The Audit Sprint is a paid one-off review you can book now. The monthly tool watches the number for you and is still on the waitlist.
Live now
$1,500 – $3,500
Waitlist
$49 – $199/mo
Being straight with you: the Audit Sprint is the only thing you can buy today. The monthly price is what the tracking tool will cost when it ships, so the waitlist is a place in the queue, not a purchase. It does not yet read your billing account or cap your spending.
The calculator below runs the same math on numbers you type in. You get the cost of one turn, what one turn earns, and what is left over. It is the quickest way to sanity-check your own figures before you pay anyone, and it shows what the monthly tool will track for you.
Set your volume, turns, tokens, model, and revenue. Everything runs in your browser. We can save the estimate without your name attached, and if you book an audit or join the waitlist we attach it to your company so we start from your numbers, not ours.
Pick which one you want. We keep the numbers you set above with your request, so when we reply we are talking about your traffic instead of a generic example.
Hard to say, and that is the problem. The table below is our estimate from published model prices, not real customer data. Treat it as a rough yardstick for the kind of numbers an audit looks at.
Most teams have nothing to compare against, because almost nobody publishes what a single AI conversation costs them. Per CloudZero/Benchmarkit research, only 43% of organizations track AI spend by customer, and just 22% track it to the transaction.
| Use case | Typical tier | Est. cost / turn | Notes |
|---|---|---|---|
| Customer support bot | Mid-tier | $0.003 – $0.013 | Short context, high volume |
| Sales / onboarding assistant | Mid to premium | $0.013 – $0.05 | Longer context, some reasoning |
| Research / analysis agent | Premium | $0.05 – $0.38+ | Multi-step, high token, agentic |
These figures are estimated, not measured. They come from published rates and ordinary usage, not from ConvoMargin customers. Your real cost depends on prompt length, which model you pick, whether you cache, and how often you retry. Put your own numbers into the calculator above.
Want your real numbers instead of our estimates? Book an Audit Sprint or join the waitlist and measured data can replace this table over time.
TLDR: This table is an estimate. An audit gives you the measured version. Try the calculator.
Because it falls between two teams. Engineering dashboards stop at tokens, traces, and retries. Finance stops at revenue per customer. Nobody puts the two on the same line, so a chat feature can quietly lose money for months. ConvoMargin puts them on one line and reports what is left after cost. It measures; it does not block your model calls or cap your spending.
TLDR: We measure the margin. We do not cut your spend for you. See how it works.
It tells you whether each AI reply pays for itself. We take what a turn costs you in model spend, compare it to what that conversation earns, and report what is left. You can buy that as a one-off audit today, or join the waitlist for the tool that tracks it every month.
One user message and the model's reply to it. A conversation is usually several turns. If the model has to retry, that retry costs money too, so it counts as extra spend on that conversation.
The money a turn brings in, minus what that turn cost you to run. It is a working number for the AI part of your product, not an accounting figure for your tax return. The calculator uses published model prices and the numbers you type; an audit uses your own logs.
Observability and LLMOps tools show you tokens, traces, latency, and retries. They are good at that, and none of them know what your customer pays. We take those usage signals, put your revenue next to them, and give you one number per turn. We sit beside those tools rather than replacing them.
The audit is a paid review ($1,500 – $3,500) that takes about 48 hours and gives you margin per turn by customer group, the prompts that cost most, and where tokens can be cut. The monthly tool ($49 – $199) sits in front of your model calls, counts tokens and turns per user, and warns you when margin turns negative. The audit is live; the tool is on the waitlist.
The calculator runs in your browser. We can keep the estimate it produces, the inputs, spend, and margin, under an anonymous session with no name attached. If you book an audit or join the waitlist, that estimate gets attached to your company so we do not start from scratch.
No. It is our estimate from published model prices and ordinary usage patterns. It shows the shape of the numbers an audit looks at. Real measured data will replace it once enough customers are sending us their actual usage.
Three places, each covered in the blog: flat-price plans where heavy users cost more than they pay, company-wide seat deals where a few teams burn the whole budget, and support chats where long conversations cost more than the ticket was worth. The calculator is free; measured work is the audit or the waitlist.
No. It works anywhere a model call is tied to something you charge for: assistants, agents, copilots, and features billed by usage.
An open page at /build-log.html where we publish what it cost to build this site: the prompts, the failed deploys, the retries, and the estimated model spend. Same math as the calculator, and labelled as an estimate.