TBM for AI Value: Applied Guidance
AI is new. Its economics aren't.
AI Solutions are built from the costs and resources TBM practices have managed for years: labor, cloud, infrastructure, software, data, vendors and contracts, and shared platforms. AI adds new signals, including tokens, training and inference lifecycles, rapidly changing models, and constrained capacity. TBM for AI Value: Applied Guidance shows how to bring them into the model you already run.
Available to TBM Council Members.
Membership is free.
An AI Solution costs far more than its token bill, and TBM models all of it
Token-driven model usage is one cost among many. Labor, outside services, hardware, cloud, and software make up the rest, and TBM already models each of them. TBM for AI Value: Applied Guidance walks teams through every step, from modeling these costs to managing consumption and the economics of each AI Solution. In the two example AI Solutions on this page, tokens are 20 percent of the total, in line with the independent studies below.
The eight costs followed in the two walkthroughs below: an insurance claims assistant built on a model API, and claims document processing on the insurer’s own GPUs. Illustrative FY2026 figures, based on the allocation patterns in Modeling the Total Costs of AI, the second article in this series.
What the research says: token costs are a minority of AI cost
20–25%
Tokens as a share of an AI agent’s variable run costs for a banking customer-service task. Human oversight is 70–75%.
McKinsey · August 2026
~33%
Tokens as a share of what an AI agent costs to run, by EY’s estimate. The rest is subscriptions, platform infrastructure, governance, change, failure, and regulation.
EY · September 2026
RSM (August 2026): “…visible costs of AI typically account for less than half of its true cost over a three-year horizon.” Each study measures a different scope, so the figures are shown side by side, not averaged.
Only 1 in 5 organizations can accurately model the full cost of AI
2026 State of TBM · TBM Council and ISG Research
Most organizations are scaling AI without a reliable cost basis. Just 5 percent produce fully burdened costs across their AI Solutions, which makes it hard to forecast demand or compare alternatives as AI grows.
What changes as compensation rises?
12x
Practitioners with a mature Solutions layer are nearly 12 times more likely to accurately cost AI. The full cost of AI comes together in the Solutions layer and reaches the business through the Consumers layer, and that is where mature practices pull away.
2.8x
Mature TBM practices are 2.8 times more likely to deliver Data/AI total cost of ownership.
1.8x
Mature TBM practices are 1.8 times more likely to have clear visibility into AI consumption.
93% vs. 50%
Of organizations that can cost AI, 93 percent already deliver Product & Service TCO, compared with 50 percent of all TBM practitioners.
AI makes mature TBM more valuable, and more urgent.
The next gap runs from AI cost to AI value
Accurate cost is the foundation, but cost alone can’t decide whether an AI investment should be funded, scaled, optimized, or retired. Few organizations connect consumption, unit economics, and value to it yet.
Only 16%
can attribute AI consumption to the Solutions and Consumers driving it.
Addressed in Managing AI Consumption and Variable Costs with TBM
Only 13%
use unit economics regularly to manage AI costs.
Addressed in Managing AI Consumption and Variable Costs with TBM
Only 25%
formally track value against the business case for most investments.
Addressed in Managing the Economics of AI with TBM
Source: 2026 State of TBM, TBM Council and ISG Research. Base: all respondents, n=221; TBM practitioners, n=116. The 12x and 4x figures compare moderate or high with low or no layer maturity. The study measures the capabilities needed to manage AI investments, not AI investment returns.
Taxonomy 5.0 gives AI cost its own place in the model
AI cost is classified in the same four layers as every other technology cost. Taxonomy 5.0 added AI elements at the two layers where AI needs more detail: the resources it runs on and the Solutions it delivers. With them, a TBM Practice can separate GPU capacity and model platforms from general compute and applications, and build TCO for each AI Solution.
The Artificial Intelligence Solution Type
Five Categories: Agentic, Generative, Interpretive, Predictive, and Prescriptive. Each has standard Solution Names, such as Artificial Intelligence:Generative:Text Generation, so each AI Solution carries its own TCO. An application that adds AI features keeps its Solution Type and is tagged AI Enabled.
Four AI Resource Sub-Towers
- Infrastructure:Compute:AI Compute
GPU-based servers and other compute built for AI workloads - Infrastructure:Storage:AI Storage
High-performance storage that feeds AI training and inference - Infrastructure:Network:AI Network
Low-latency fabric that connects GPU clusters - Applications:Application:AI Models
Model APIs billed by the token, and platforms for building and deploying AI
No new Cost Pools
Token charges, GPU capacity, and AI labor are tagged by type of expense, like any other purchase. Most of an AI Solution’s cost still sits in standard Resource Sub-Towers such as Development and Support & Operations, as the examples below show.
How it works
Follow one AI Solution through the model you already run
The TBM Model follows every cost by the type of expense (Cost Pools), the function of the resources it funded (Resource Towers), what was delivered (Solutions), and who drove demand (Consumers). The examples below follow AI costs through those layers, using the Taxonomy 5.0 AI elements where they apply.
From the guidance:
Token costs are a quarter of the cost
A cost basis built from the model provider’s bill alone would understate this AI Solution by $1,280K.
Labor is the largest Cost Pool
Development labor ($450K) costs more than token usage. Run support adds $300K.
Every dollar follows rules TBM already uses
Provider bills are broken down by meter, account, and tag before they are allocated, the same way any cloud bill is.
See how TBM accounts for the two most common AI solutions
Most organizations either build an app on a commercial model they pay for by the token, or run a model themselves on GPU capacity. Each walkthrough follows the costs that matter most for that kind of Solution, one step at a time, from the general ledger to the business units that use it.
Please visit on desktop for the interactive walkthrough
Insurance claims assistant on a commercial model API
An app the insurer built helps claims and customer service staff draft replies and summarize claim history. It calls a commercial model, for example Anthropic’s Claude API, and pays per token.
- Solution
- Artificial Intelligence:
Generative:Text Generation
- Runs on
- Model API, plus app servers and a document search index on AWS
- People
- Internal engineers build it; a shared support team runs it
- Showback
- Claims Operations and Customer Service, by interactions handled
Costs you’ll follow
- Token usageWhat the model provider charges per request, priced by the token
- Development laborEngineers who build and extend the app, part of a larger team
- Support laborA shared support team; the assistant’s share comes from ticket data
- Document search (AWS)The policy documents the assistant checks before it answers, known as a retrieval index
Based on the guidanceModeling the Total Costs of AI, the second article in this series, Sections 3 and 5 · Illustrative figures, FY2026, $ thousands
Open walkthroughClose walkthrough
Claims document processing on the insurer’s own GPUs
The insurer tunes a free, open-weight model, for example Meta Llama, to read claim documents and pull out the details staff need. It runs on GPU servers the insurer owns and shares with other AI Solutions.
- Solution
- Artificial Intelligence:
Interpretive:Document Processing
- Runs on
- Owned GPU servers for daily work; rented AWS GPUs for tuning runs; a SaaS data platform
- People
- An outside data science firm tunes and tests the model
- Showback
- Claims Operations and Underwriting, by documents processed
Costs you’ll follow
- Owned GPU serversBought and shared by several AI Solutions; they reach the model as depreciation
- Rented cloud GPUsRented by the hour on AWS for occasional tuning runs
- Data science consultantsSpecialists who prepare the data and tune and test the model
- Data platform (SaaS)The model is free; the software to tune and manage it is a subscription
Based on the guidanceModeling the Total Costs of AI, the second article in this series, Sections 7 and 8 · Illustrative figures, FY2026, $ thousands
Open walkthroughClose walkthrough
Use Next and Back, the step track, or the arrow keys. Next on the last step moves to the next cost.
TBM for AI Value: Applied Guidance
Four articles, each matched to a gap in the research
The series applies established TBM standards and practices to AI. Use it before investing in an AI Solution, or on AI Solutions already in development or production. Each article builds on the one before it and follows the same five example AI Solutions, from a licensed AI assistant to a self-hosted agent.
Calibrating TBM Practices for AI Value ManagementPlanning AI Value & Preparing the TBM Practice to Prove It
- Plan each AI Solution’s business value proposition: Value Drivers, KPIs and baselines, and value thresholds.
- Find which TBM use cases and capabilities to mature first for the AI Solutions in scope.
- Prepare data, methods, and roles for tokens, lifecycles, and constrained capacity.
Modeling the Total Costs of AIBuilding Solution TCO Across AI Architectures
- Scope the cost model by architecture, from Commercial AI SaaS to a Proprietary Model.
- Apply standard allocation paths for every AI cost, including shared GPU pools.
- Build fully burdened annual cost and TCO, and allocate it to business units and value streams.
Managing AI Consumption and Variable Costs with TBMConnecting Tokens & Usage to Forecasts and Defensible Showback & Chargeback
- Reconcile token and usage records to billed cost and to the ledger.
- Calculate cost per unit of work and forecast cost from demand.
- Set guardrails and allowances, and run showback or chargeback.
Managing the Economics of AI with TBMSteering Funding, Capacity, Skills, & Compliance Across AI Portfolios
- Rank AI Solutions for budget, constrained capacity, and specialist staff.
- Plan GPU capacity margins, order points, and supply options.
- Use value thresholds to fund, scale, shape, optimize, or stop each AI Solution.
Concise ReferenceKey tables, allocation paths, and checklists for AI cost, consumption, economics, and data
- The main tables from all four articles and the AI chapters of Data Strategy & Architecture for TBM Practices.
- For teams who have read the guidance and need a working reference while they scope, build, and run AI cost and value models.
Help shape the next version:
This guidance will be updated as AI and TBM practice mature. Tell us what worked, what was missing, and what your teams found when they applied it. Send feedback to the TBM Council Standards team.
