HR hears that AI is "included." Finance sees a usage bill that grows every month. IT hits rate limits mid-rollout. All three are looking at the same thing from different angles: the token cost of running an employee assistant, and the fact that employee support is exactly the kind of high-volume, repetitive workload where that cost adds up.
This guide explains tokens in plain English, shows where employee support burns them, walks through an illustrative worked example, and gives you a procurement-ready list of questions for Finance and vendors. No engineering background required. The goal is simple: so that when someone says "AI is included," you know exactly what to ask next.
A token is a chunk of text, roughly a few characters or part of a word. Language models bill by the token, and they charge for two separate things:
Input tokens are everything you send the model: the system instructions, the retrieved policy content, the conversation history, and the employee's question. Output tokens are what the model writes back. Output almost always costs more per token than input, typically three to six times more, because generating text takes more compute than reading it (TECHSY, 2026). Two consequences follow that matter for employee support. First, long policy documents and long chat histories inflate input cost, because they are sent to the model on each turn. Second, "unlimited AI" is a pricing wrapper, not physics. Someone is still paying per token underneath, and if that someone is you on a consumption plan, the meter runs with every question.
The good news: prices have fallen hard. Industry trackers estimate LLM API prices dropped roughly 80% between early 2025 and early 2026 (iternal.ai). The bad news: employee support volume is high enough that even low per-token rates add up, and the expensive part is often not the model at all.
Four patterns drive most of the spend:
Wide-drive retrieval. If the assistant searches your entire SharePoint or file drive on every question, it pulls large amounts of text into the prompt as input tokens. A curated knowledge base sends far less text to reach the same answer.
Multi-turn conversations. Each follow-up resends the conversation so far. Long threads compound input cost turn by turn.
Multilingual interactions. Translation and non-English understanding can add processing steps, raising cost per interaction for global workforces.
Agent steps. When an assistant chains multiple reasoning or tool-use steps to complete a request, each step consumes tokens. More autonomy can mean more spend per resolution.
The theme connecting all four: how the system is built matters more than which model it uses. A well-scoped assistant on curated content can cost a fraction of a do-it-yourself bot pointed at an uncurated corpus, even on the same underlying model.
This is a simplified illustration to show the shape of the math, not a price quote. Token rates move constantly, so treat the numbers as order-of-magnitude.
Current LLM API rates span a very wide range: from about $0.10 per million input tokens for budget models to $30 per million for frontier reasoning models, with a common production model around $2.50 to $3 input and $15 output per million tokens (CloudZero, 2026).
Imagine 2,000 employees each asking a handful of questions a month. Two designs answer the same questions very differently:
On identical volume and the same model, the curated design can cost a fraction of the wide-corpus design, because it sends less text per answer and resolves faster. Two levers cut the bill further and are worth asking any vendor about: batch processing (a flat 50% discount on many providers) and prompt caching (up to roughly 90% off repeated input on some providers) (TECHSY, 2026).
The point is not a precise dollar figure. It is that architecture, not model choice, usually decides whether your token bill is small or alarming.
The token meter is the cost everyone watches. These are the ones that quietly cost more:
Fully loaded cost is license plus model usage plus retrieval plus all of the above. A cheap token rate on an expensive-to-run system is not a bargain.
For many HR and Finance teams, the appeal of per-employee-per-month (PEPM) pricing is not that it is always cheapest; it is that it is forecastable. You know your bill from your headcount, without watching a usage meter or absorbing a spike during enrollment season. Pure consumption pricing tracks usage volatility, which can be efficient at low volume and unpredictable at high volume.
Which model wins depends on your volume and your tolerance for variability. We compare the options in detail in our companion piece on employee AI pricing models, and you can see MeBeBot's transparent bands on the pricing page.
No. Token cost is what the underlying model charges per unit of text. PEPM is a pricing wrapper that charges a flat amount per employee per month, so the vendor absorbs the token variability. They are different layers of the same bill.
On a do-it-yourself build, usually IT or engineering, with Finance watching the total. On a PEPM platform, the cost is predictable and typically sits with the HR or IT budget line that owns the tool. Decide ownership before launch, not after the first invoice.
Often, yes. An assistant that retrieves a small, curated answer sends less text and resolves in fewer turns than one that drags in large uncurated documents and loops through multiple steps. Accuracy and efficiency tend to move together.
Per-token prices are falling fast, but employee support volume is high and the hidden costs (maintenance, rework, administration) do not fall just because tokens do. Architecture and accuracy will keep mattering even as raw token rates decline.
"AI is included" is the start of a conversation, not the end of one. Tokens are simply the text going in and out of the model, and employee support is high-volume enough that how the system is built (curated versus wide-corpus, few turns versus many) usually matters more than which model sits underneath. Bring the fully loaded cost into the room, ask the nine questions above, and decide whether predictable PEPM or variable consumption fits your team.
Want to model your real question volume? Book a demo and we will walk the numbers with you. For deeper cost context, see how to measure AI ROI in employee support, cutting cost per ticket, and the build versus buy TCO breakdown.
Token prices and provider discounts change frequently. Figures here reflect public rates at the time of writing and are illustrative, not a quote.
.svg)
.svg){kind=icon}
.svg){kind=icon}