AI Automation Cost Explained: Token Pricing and Build 2026

When businesses start trying to price AI automation, the conversation gravitates to certain areas: which model is cleverest, which brand the board already trusts, whose demo made someone in the meeting room sit up.

Those feel important because they are things you can see.

Then the invoice arrives, and it arrives somewhere else entirely. The space between expectation and reality is where automation budgets get many of their surprises, and many of those surprises are avoidable once you understand where the money is actually going.

The fastest way to make that visible is with a number you can see and hold in your head.

At published June 2026 rates, drafting 5,000 first-pass customer support replies in a month costs roughly $13 of model usage on a fast, low-cost Claude AI model. Push the same 5,000 replies through a top-tier model and the figure is about $65. Either way, the AI itself is doing thousands of dollars worth of work for the price of a take out meal. And the figure sits a long way below whatever number the average automation proposal will show you.

That gap is the thing worth understanding before you sign anything.

Token cost is rarely the factor that decides whether an AI automation task is worth doing. Once that becomes obvious, the conversation changes shape and you read a proposal differently. You prioritise different tasks first, and you start asking the people building the system a sharper set of questions.

What you are actually paying for in AI automation

To see why those numbers come out so low, it helps to understand how AI models charge in the first place. They read and write in chunks called “tokens”, which are the building blocks of language for LLMs.

A useful rule of thumb: 100 words is about 130 tokens, so a short email is 100 to 200 tokens, a page of text 500 to 750 tokens, and a 20-page contract around 15,000. 

Watching real text turn into tokens in OpenAI's tokeniser will be genuinely helpful and take less than a minute.

You pay for every token the AI model handles, and the price splits in two.

Input tokens are everything you send to the model: the instruction, the customer's message, any documents you have attached.

Output tokens is everything the model writes back.

Both are priced per million tokens, and across the major models, output costs four to six times more than input. Anthropic's Claude runs at a clean 1:5 ratio between the two. OpenAI's flagship sits nearer 1:6. Google's Gemini models land in the same band.

The reason output is the pricier side is mechanical.

A model reads your whole (input) prompt in a single, fast pass, the way your eye glances across a page.

It writes its (output) answer one token at a time, each new word shaped by every word before it, which is slow and computationally expensive. You are paying for the harder of the two jobs.

Why the shape of the AI task matters more than the brand

That mechanical detail, output being more expensive than input, has a strategic consequence that catches many people off guard. What you ask the AI to do moves the bill far more than which model you choose to do it.

The cleanest way to see this is to fix the model and vary the job.

Take a single model, Anthropic's Claude Sonnet, and give it two pieces of work of the same total size.

The first job, summarising a long document, reads about 10,000 tokens and writes a 1,000-token summary, which comes out at around 4.5 cents a time.

The second job, drafting a long piece of writing from a short brief, reads about 1,000 tokens and writes 10,000, costing around 15 cents a time.

Same model, same total number of tokens used, but the writing-heavy job costs roughly three times more, purely because of how much the model has to write back.

This means the question, "which model should we use?" matters a lot less than it tends to feel in the meeting room.

Reading-heavy work stays cheap even on a capable model: summarising long documents, sorting and triaging, pulling figures out of forms. Writing-heavy work costs more wherever it runs.

The strategic implication then is obvious.

The first tasks worth automating are usually the high-volume, reading-heavy ones, because that is where AI delivers the biggest efficiency gain for the smallest cost. 

Think about sorting and routing inbound email, where the model's output is a single category label and the cost effectively rounds to zero. Reading numbers off invoices and forms. Drafting first-pass replies for a human to review and approve before they go out. These are the workflows where the maths is at its kindest.

The hidden multiplier, and two discounts that swing the other way

Before moving on, one detail is worth knowing because it can quietly inflate a bill if nobody is paying attention.

Some AI models are designed to reason, before producing an answer. They generate a stream of internal working - a “Chain of Thought” or CoT - that you never see in an automation, but will probably have seen if you’ve used the “Thinking” models in a chat interface. You will still pay for this at the output token rate.

On a difficult task, the CoT can run many times longer than the visible answer itself. For routine drafting, sorting and everyday automation, the simpler workhorse models handle the job, so the figures above hold. Reasoning models earn their place on the harder problems.

Pulling the bill in the opposite direction, there are two discounts a well-built system will use, and both are substantial.

Batch processing applies to work that can tolerate a wait of up to 24 hours, like overnight summarising or bulk document tagging, and runs at roughly half the standard price for the same job on the same model.

Prompt caching applies whenever the same large block of instructions goes out on every call. For example the knowledge base a support agent reads from. The system stores that block and reuses it at about a tenth of the normal input rate. On a support agent that reads the same knowledge base thousands of times a day, prompt caching is the single biggest lever on the model budget.

So if the AI automation itself costs pennies per task, why does a real automation project carry a large price tag?

And the answer is everything that surrounds the model.

Think of the model as the engine, powerful and efficient. But on its own it goes nowhere. The value sits in the harness built around it, and that harness has several parts that all take real work to build.

Every connection to a CRM, a helpdesk, an inbox or an order system is something that has to be designed, tested and kept working as those tools evolve. The underlying data and knowledge the AI draws on has to be clean and consistent enough to give useful answers, because no amount of model intelligence rescues messy inputs.

The instructions the system runs on have to be written, tested against real cases, and tightened until the behaviour holds. And around all of that, guardrails and review steps have to catch the mistakes that could otherwise reach a customer.

The benchmark data tells the same story across the industry. Zendesk's CX Trends 2026 found that the median support team running AI deflects around 41% of tier-one tickets, while the top-quartile teams deflect closer to 59%.

Gartner's reading of the same picture is sharper still. AI deflects 45% or more of queries on average, but only 14% of issues reach full self-service resolution.

The gap between the teams getting average results and the ones getting genuinely strong results sits almost entirely in how deeply the AI is integrated into the systems behind it: order data, account history, billing, the things that let the model actually answer the question rather than guess at it. Same models. Different harness.

The instruction-writing also rarely ends on day one.

A 2026 study presented at the CHI conference describes building a reliable AI application as a continuous cycle, in which the instructions and a growing library of real test cases evolve together as edge cases surface in production. In practice it feels less like installing software and more like onboarding a new hire. You watch the output, correct what is wrong, and hand over more responsibility as the work earns trust.

A common industry rule of thumb puts ongoing maintenance at roughly 15% to 30% of the original build cost per year, which is the part of the budget likely to be missing from a first quote.

Which is why, when an automation proposal lands on the desk, the useful thing to do is read it for the build. The tokens are the cheap part, and any clear quote will show model usage as its own line with the volume and the model behind it stated explicitly.

The rest of the proposal should account separately for integration, data preparation, testing and ongoing maintenance, with maintenance shown as a standing annual line rather than a one-off afterthought.

The question worth asking out loud in the meeting is who owns the system in twelve months, because if that answer is fuzzy the long-term picture is going to be fuzzy too.

The model cost itself is the easy part to settle. Our AI Automation Cost Calculator shows the model usage for real workflows across the main model options, so you can put an example figure against your own tasks before any conversation about budget.

The work that decides whether the automation will hold up is the foundation underneath.

That foundation: integration, clean data, reliable instructions, ongoing support, is the real investment, and it’s the work we focus on at RRCEO.