The Turqoise Killfish reaches maturity in two weeks, faster than any other animal. The Tokenomics Foundation is growing up almost as fast. After launching just under 5 months ago with the mission of creating a shared discipline for managing AI and related token costs, it’s quickly evolved into something much more ambitious: a methodology to understand the total cost of AI and connect that cost to business value. In fact, this past September, the Tokenomics Foundation argued that token and API costs make up only 10% to 25% of total AI spend.  This evolution creates a challenge. You can measure AI costs without a taxonomy, but you can’t connect them to business value without one.

ROI Needs A Defensible Cost

The formula for AI ROI is simple: the numerator is the value created and the denominator the cost required to create that value. The value side of the equation is already difficult for most. Organizations must determine whether an AI investment saves employee time, increases revenue, improves customer outcomes, or reduces risk. Introducing the Forrester AI Value Matrix is a good starting point to help make this determination.

The denominator is tricker. Consider an example using two agents: one agent improves the evaluation and routing of sales leads for Product A and Product B. The other expedites prep for sales meetings, creating more customer-facing time for salespeople for both products. The combined capabilities of these agents result in a $2 million lift in sales of Product A and a $1M lift in sales for Product B.  Since each agent is splitting its capabilities between both products, how should the organization attribute the impact of each agent on Product A and Product B for its ROI calculations?

The first layer of the model categorizes AI costs, i.e. LLM costs, cloud services costs, on prem capitalization costs, labor, etc. The model then needs to attribute these costs to defined AI capabilities to understand the total cost of each agent, i.e. what combination of LLM costs, cloud, on prem, etc. map to each individual agent. Lastly, once we know the cost of each agent, the model needs to allocate each agents’ costs to defined business outcomes, in this case splitting them between Product A and Product B. This includes costs that aren’t easily taggable like on-prem storage or non-time sheet labor. Only then will you have a defensible cost basis where cost and benefit sides of your ROI calculation align.

Tokenomics Can Learn From TBM

Technology leaders have encountered this problem before. Technology Business Management (TBM) doesn’t attempt to connect a server or cloud instance directly to business value. Just like in the example above, it uses intermediate layers to translate technical resources and costs into applications, services, products, and ultimately the parts of the business consuming them.

AI needs a similar connective tissue. Will it come via an extension of TBM? Will the Tokenomics Foundation create its own AI cost taxonomy? Could some other organization create a best practice? Time will tell. However, what we’ve learned from TBM is that, while these models are powerful, they are also difficult to implement and maintain. The reason is that they’re not plug and play. As you progress through layers of the model, you have to make assumptions as to how to break up blocks of costs across multiple stakeholders, and maintaining these model assumptions requires resources.

For example, in the scenario presented above, it’s not immediately obvious how to split the cost of each agent equitably between Product A and Product B. Rather, it requires a formula to split this cost based on some indicative metric (such as # of salespeople, # units sold, etc.) and that formula needs to be agreed-upon across all stakeholders, owned and governed to create universal confidence in the model and the outcome.

What You Can Do Now

You can’t wait for a standard taxonomy to emerge – you need to make the right decisions on AI for your organization now. So, create your own taxonomy until a broader standard evolves. Here’s a simplified way of thinking through this to get you started, from which you can adapt to your own specific needs:

  1. Start with the outcome: define the nature of the benefit and the unit in which it will be measured.
  2. Define and map AI capabilities to outcomes: Identify the agents contributing to each outcome, define the capabilities they provide and establish allocation drivers where capabilities are shared.
  3. Build the total cost of each capability: Define your categories of AI spend and determine the total costs attributable to building/running your agents. In some cases, this mapping will be clean, such as if your observability tools provide cost attribution by agent. In others, such as labor, the mapping may be less clear, so align with all stakeholders on the assumptions that will drive attributions of these cost categories.
  4. Govern your assumptions: Document allocation methodologies, assign owners to maintain them and secure agreement from stakeholders.

Do this for one use case first to demonstrate the methodology and build a base for broader buy-in. Then expand the model as you gain support across functions as well as up the org chart. Lastly, don’t aim for perfect attribution on day one. The goal is to build enough structure between technical consumption and business outcomes that leaders can understand, challenge, and ultimately trust the economics.

Want to talk more about building your model for AI spend attribution? Set up a guidance session with me or email me at gzorella@forrester.com.

 

 

 

Share