How Much Do AI Sales Agents Cost in LLM Tokens?
The deal-review agent cost next to nothing in the demo, and then RevOps pointed it at every open opportunity.
The pilot bill and the production bill
Picture a RevOps engineer who builds a deal-review agent in a week. It pulls the Gong transcripts for an opportunity, hands them to a general model such as Claude, and asks for MEDDPICC status, the top risks and a next step. On the one deal in the demo, the answer reads well and the bill is a rounding error.
Then the agent goes nightly across 400 open opportunities. Same prompt, same model. The bill now grows with every transcript on every deal, every night, and with every step the agent takes to reason through them.
Nobody did the arithmetic wrong in the pilot. The pilot measured one deal, once, and production cost is driven by how often the model re-reads the same material.
📊 Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. — Gartner |
What you pay for: tokens in and tokens out
Model providers bill by the token, which is roughly a fragment of a word. They price input tokens, what the model reads, separately from output tokens, what it writes, and output usually costs more per token on the major APIs.
A sales agent reads a lot. One hour-long discovery call produces a transcript of many thousands of words, and a late-stage enterprise deal can carry a dozen calls plus the email thread. Every time the agent answers a question about that deal, it reads whatever context you hand it.
It writes, too. Summaries, reasoning steps, drafted field values and the recap email all bill as output.
Cost driver | What makes it grow | Where it shows up |
|---|---|---|
Transcript length | Longer calls, more calls per deal | Input tokens |
Deals in scope | Running on the whole pipeline | Input and output |
Run frequency | Nightly runs, or a run on every new meeting | Everything, multiplied |
Agent loop steps | Tool calls, retries, self-checks | Input re-read on each step |
Generated text | Summaries, rationale, drafted fields | Output tokens |
The rows multiply each other. Deals in scope times runs per week times steps per run turns a modest per-call figure into a line on the budget.
Reasoning tokens bill as output
Models that think before they answer generally bill that thinking as output tokens, including on APIs that never show you the text. An agent asked to reason carefully through every risk on a deal can spend more on the reasoning than on the answer it hands back.
Why agent loops multiply the read
A single prompt reads the transcript once. An agent loop sends the conversation so far back to the model at each step, including every tool result it fetched along the way. A five-step run can read the same transcript five times.
Retries add to it. When the agent checks its own answer, calls another tool or re-plans after an error, each turn carries the full context again. If you log token usage per run, look at how much of the input had already been read on an earlier step.
Prompt caching helps, within limits
Prompt caching on the major model APIs lowers the price of re-reading an identical prefix. It helps less when the context changes from step to step, and a deal's context changes every time a new meeting lands.
The cost that never appears on the invoice
Ask a general model the same question about the same deal on two different nights and it can return two different answers. Teams respond by adding checks: a second pass, a judge model, or a manager reading the output before it reaches the CRM. Each check costs more tokens or more human time.
The review step decides whether the agent survives past the pilot. A manager who validates every output has been handed a new queue to work through.
Reps feel it too. When the agent's MEDDPICC read changes from Tuesday to Wednesday with nothing new on the deal, the rep has to defend a status they never set, and the next deal review spends its time on the tool.
📊 Only 1 in 5 AI use cases delivers a positive return. — Gartner, IT Symposium/Xpo keynote, October 2026 |
Estimating the bill before you scale
Log tokens per run, split by input and output. Provider dashboards and API responses report both, and the split shows which lever to pull.
Count how often each transcript is read. Add up steps, retries and reruns per week. If one call transcript is read dozens of times, storing answers will save more than switching models.
Separate repeat questions from one-off questions. MEDDPICC status asked every week on every deal is a repeat question. A research question about one account is not.
Price the review step. The minutes a manager spends checking agent output belong in the cost model.
Test consistency before you scale. Run the same deal through the agent twice with no new activity. If the answers differ, budget for the checks you will add.
Set a token cap per deal. A weekly budget per opportunity stops one long enterprise deal from eating the run.
A decision layer that writes nothing
Spotlight.ai runs on Spotlight's Calibrated-Decision AI-Model (CDAM), and technical buyers usually ask first which model we trained. We didn't train one. CDAM is a calibrated decision architecture: 40M+ atomic signals, a closed answer space for every question, and composition in code, with a general model used at two of the six stages and only inside a space the software defined before the call.
That shape changes the bill. The model answers narrow questions one at a time, and each answer is a typed value: true or false, one option from a fixed list, or a score on a fixed rubric. Every answer carries a confidence score and only counts once it clears the bar. Qualification, the Spotlight Score, gaps, risk and the forecast are then calculated in code from the stored values.
So the decision layer uses zero output tokens, because it writes nothing. We still pay for what the model reads during extraction, once per question as new evidence arrives. Ask the same question of the same deal a thousand times after that and you pay for the model zero times, since the answer is stored and the same evidence gives the same answer.
Those economics hold across the revenue team. Sales gets every deal qualified on MEDDPICC and forecast on buyer evidence. RevOps gets complete deal data and the win, loss and slippage patterns behind the number. Value teams get a business case on every deal from the buyer's own words, and Customer Success qualifies and forecasts renewals with the same rigor.
Where a general agent still earns its tokens
A general agent answers questions nobody anticipated, and that flexibility is real. For a one-off account question, a first draft of an email or an ad hoc analysis, an agent loop is a sensible tool and the token cost of an occasional question stays small.
The cost problem appears when the same questions run on every deal, every week. Spotlight's MCP server lets your own agents read and act on deals CDAM has already qualified and scored, so they spend tokens on the new question and skip re-reading transcripts to rebuild MEDDPICC from scratch.
Price one question across your pipeline
Pick one question your agent answers on every deal, such as whether the economic buyer is engaged. Multiply transcript tokens per deal by open deals, runs per week and steps per run, then price it on your provider's current rate card. That number is your floor before retries and review time.
Then compare it with the same question answered once and stored. A Spotlight.ai demo can walk through both on deals shaped like yours.
FAQs About AI Sales Agent Token Costs
How much does an AI sales agent cost to run?
It depends on how many tokens it reads and writes per deal, how many deals it covers and how often it runs. Transcript length, agent loop steps and nightly reruns usually matter more than the price per token.
What are input and output tokens?
Input tokens are what the model reads, such as transcripts, emails and instructions. Output tokens are what it writes, such as summaries and field values, and they usually cost more per token.
Why do agent loops cost more than a single prompt?
An agent sends the growing conversation back to the model at every step, so the same transcript can be read several times in one run. Retries and self-checks add more passes.
Does prompt caching solve the token cost problem?
It lowers the cost of re-reading an identical prefix. It helps less when the context changes, which happens on a live deal every time a new meeting or email arrives.
What does zero output tokens mean for Spotlight.ai?
Spotlight's decision layer writes nothing: qualification, scoring, risk and the forecast are calculated in code from stored typed values. Spotlight still pays for what the model reads during extraction.
Should we build our own sales agent?
General agents are a good fit for one-off questions. For questions asked on every deal every week, point your agent at stored, qualified answers through Spotlight's MCP server so it does not re-read every transcript.
Comments