Before shows like “Game of Thrones” and “Succession” and affordable wide-screen television sets, there used to be a huge quality gap between watching TV in your living room and seeing a movie in the theater. And, likewise, there used to be a big difference between the video games you could play at home on your Nintendo and the games available at your local arcade. If you wanted the full experience of Dance Dance Revolution or Mortal Kombat, you needed to drive to the local mall, go to the arcade, exchange your money for tokens, then keep feeding those tokens into the game machine to keep playing.

But while the idea of paying to play arcade games might seem like an outdated nostalgia trip, it’s actually a good metaphor for AI in the workplace, given how AI has a “variable” cost that grows the more people use it.

My own company uses AI to help clients train their human workforce, leveraging models like ChatGPT, Gemini, and Claude to power interactive role play simulations, machine-graded free responses assessments, and even “instructor-led” training sessions delivered by an AI agent. But while clients are usually intrigued and impressed by what AI can do, there’s always a moment of vertigo when the conversation turns to price:

“Wait, so… we have to pay each time an employee talks to the AI?”

For the past 10-15 years, organizations have used e-learning and online video as a cost-saving measure for large-scale training programs, and e-learning modules and videos have effectively zero cost of delivery beyond the cost of development (or, in some cases, the cost of hosting on a learning management system). So the notion of having to pay $1.00 for a worker to take a machine-graded, free-response workplace safety quiz or $2.00 for a physiotherapist to get an AI copilot’s advice on a patient’s exercise regimen or $3.00 for a salesperson to spend an hour rehearsing their pitch with simulated AI customer might seem exorbitant.

However, as we’ll discuss in this article, the variable cost of AI makes a lot more sense once you understand how these apps work and why it’s more appropriate to compare AI-based applications to human instructors / experts than a “static” piece of media like a video, web page, or e-learning module.

What’s a “Token”?

In the world of AI, a “token” is a unit for measuring the computational power used by an AI model, based on the complexity of the text you put in and the complexity of the output it generates. When you interact with an AI, each message you send is broken down into these tokens, which the AI processes to understand and generate responses, and most AI providers charge a few cents or dollars for every million tokens entered into or generated by an AI model.

To give an extremely rough rule of thumb, 1 word usually equals 1.35 tokens. This is because a “token” represents a unit of meaning – not necessarily a word. So while the word “questioning” might be one token, “unquestioning” might be two because the prefix modifies the meaning, and then the model also has to account for things like punctuation.

Calculating the cost of tokens is complicated by the fact that an AI model doesn’t just read each piece of text once. Rather, every time you send a message, the AI re-evaluates the entire conversation, processing all the tokens anew, so it can ensure its responses reflect the entire context of the interaction. This means that longer interactions, with more back-and-forth, naturally incur higher costs. However, it’s worth noting that reprocessing existing tokens is much less expensive than generating new text, so you aren’t paying the full cost all over again.

What Makes One AI Interaction Cost More (or Less) Than Another?

So, if tokens are the currency of the AI world, how does that translate into the costs for using an AI application to train workers or automate a task?

Just like work done by humans, the price of AI interactions reflects the cost of materials, the cost of “labor”, and the overall value of the task being performed or the problem being solved.

Tokens can be likened to the cost of materials. When it comes to “labor” costs, different AI models come with varying capabilities and price points. For instance, a high-capacity model like the latest version of ChatGPT might cost around $2.50 per million tokens, while a more streamlined “mini” version could be as low as 15 cents to process the same volume of tokens / text, with the difference lying in the quality of the output.

In our experience helping organizations use AI for workforce training, we’ve found that a less expensive “mini” model is often sufficient for straightforward tasks like grading free-response questions about food safety or anti-money laundering regulations. However, if you want an AI model to simulate an upset customer in a role play activity for customer service representatives, the more advanced models will usually do a more convincing job of generating human-sounding dialogue than the low-cost models. While the cheap model can do a fine job of writing a textbook, it’s not great at capturing the emotions and nuance of an unhappy customer complaining about a broken refrigerator.

In this sense, AI models are similar to human experts and instructors: just as you wouldn’t hire a world-famous surgeon to teach a course on basic first aid, you should match the AI model to the complexity of the task. However, a major difference is that while human wages tend to increase over time and based on someone’s education level, the cost of AI models tends to go down as newer, more efficient versions are developed which deliver superior performance at lower per-token prices.

How Does the Cost of AI Interactions Compare to Human Interactions?

Just as an interior designer’s work is worth more than the cost of upholstery and paint, it’s important to consider not just the technical costs but also the value provided by AI interactions.

As we’ve discussed in other blogs, designing AI interactions (“prompt engineering”) is an art and science, and ensuring an AI model reliably performs complex tasks to a consistent level of quality takes more than typing “design a training program for doctors” into ChatGPT’s consumer interface.

To gauge the value of an AI interaction, our company will ask “How much would it cost to have a human perform the same task?” then set the subscription fee for the AI interaction at a fraction of the human rate. For example, an AI that provides basic information retrieval might be priced lower than one that offers expert-level coaching or consultation.

Of course, saying “an AI is worth 50% or 35% of a human” doesn’t capture the full value. For instance, AI agents are available on demand – 24/7/365 – whereas a human consultant or trainer might take weeks or months to fit you into their schedule. If comparing an AI-facilitated role play activity against a role play in a workshop, we’d need to account for the cost not just of the coach and the person doing the role play, but also the other 15 people sitting around in the classroom or on the Zoom call, waiting for their turn. And there’s also the value of speed: once properly designed, an AI agent can accomplish any task in seconds (or milliseconds), even ones that might take the most qualified human several hours or days.

Budgeting for AI in the Workplace

As organizations integrate AI into their operations, understanding and managing the variable costs of AI interactions will become crucial for planning projects.

In the context of workforce training, it might seem strange to compare the costs of AI to human instructors rather than “traditional” digital learning content, however – once you get past the strangeness – it’s a better paradigm for measuring the value AI brings.

For the sake of simplicity, our company has found ways to make AI licensing and usage costs resemble the familiar “seat license” model for e-learning platforms and live workshops (i.e. pay $XXX per user per course per month), however, as organizations become more comfortable with AI, we expect a shift towards more flexible approaches.

One very large organization we advise is already doing this, giving employees a monthly “token allowance” and letting them spend it on whatever company-approved AI applications they need. It’s not unlike parents in the 1980s or 1990s dropping their kids off at the video arcade with a fistful of quarters- proving that, even at the cutting edge of AI technology, everything old is eventually new again.

Hopefully this article has offered some useful perspective on the costs of operationalizing AI in the workplace.  If you’re interested in using AI to train your workforce, or just need guidance on your organization’s overall AI strategy, please consider reaching out to Sonata Learning for a consultation.