Two Sets of Data,Two Ways to Buy AI

The question

How should a company buy AI for analytical and forecasting work? Spending is climbing even as the cost of comparable model capabilities falls: enterprise large language model (LLM) spending tripled over the 12 months to the end of 2025, according to data McKinsey cites. Contracts have not kept pace. In a survey of 202 senior enterprise leaders, 65% of buyers still pay for services on FTE (full-time equivalent) terms: the price follows the headcount assigned, even as AI takes on more of the work. A CEO approving AI spend meets two kinds of contract, one priced by headcount and one by subscription or usage, and needs a way to tell which fits which workload.

How to sort the work

Start by sorting analytical AI workloads by how fast their inputs, definitions and predictive relationships change. This paper covers analytical and forecasting work, not every use of AI, such as drafting, coding assistants or image generation.

Most organizations have both kinds of analytical work. One looks back at a defined record: historical data and the analytics built on it. The other predicts outcomes whose inputs or relationships keep changing, driven by market inputs: sales, demand forecasting, and similar. These are workload categories, not permanent labels for datasets.

What to do

1. List your analytical AI workloads. Mark each one by whether its inputs, definitions or predictive relationships change. Historical claims work whose coding rules or payer terms change needs revalidation. Put it in the changing column when those changes alter the analysis.
2. For the stable ones, start with a reproducible analytical pipeline your team controls. Add a model only where the question requires one. Revisit the work when records, definitions, questions or performance change.
3. For the market-driven ones, decide who is responsible for keeping performance current, a vendor or your own team, and name an owner for watching its accuracy.
4. Match the pricing to what you are buying: people (FTE) when you need continuing human capacity, access (subscription), usage (consumption), or a managed result. Decide separately where it runs: on-premises or cloud. Finance sets the accounting from the arrangement.

Why the split holds

A demand forecasting model learns patterns from past data, and those patterns can stop holding when consumer behavior shifts. A model that is off on demand throws off inventory planning, and the error carries through the supply chain. A January 2026 preprint describes scheduled retraining every three to six months as common practice in supply chain forecasting.

For historical and semi-static data, established methods often do a great job, and some questions need no model at all, only calculations and attribution. Where a model is needed, a 2022 benchmark across 45 tabular datasets found tree-based models such as XGBoost state of the art on medium-sized data, about 10,000 samples in the authors' benchmark. That supports testing established methods first. It says nothing about how often to refresh. Old model weights do not by themselves establish a security vulnerability, and accuracy drift is a different problem from security. But attacks can target the model itself, through poisoning, evasion or model extraction, so the model and the software around it need an owner and controls.

Infrastructure and staffing follow the same logic. Steady, high-utilization workloads can favor owned capacity, while variable or lower-volume workloads can favor consumption-based cloud pricing. Ongoing, predictable demand goes to core headcount, and contingent labor meets temporary needs. Internal ownership usually means people on payroll and capacity you control, whether on-premises or in your own cloud account. Those are implementation choices, not consequences of the rule. Token spending can build capability, run internal work or serve customers. Finance decides what qualifies for capitalization and, when costs are expensed, whether they belong in cost of goods sold. The accounting follows the use, the arrangement and the applicable standards, not the tier.

Some say FTE pricing makes no sense in an AI-led world. In my view that is not quite the case. Buyers expect FTE-based pricing to shrink sharply, and shrinking is not gone. Headcount is a poor unit when AI takes on more of the work. It can still fit when you need continuing human capacity.

Words worth recognizing

Analytical pipeline

Reproducible calculations and attribution, using SQL, statistics or BI (business intelligence) tools, that answer a retrospective question without a learned model.

scikit-learn

A general-purpose machine-learning library for building and evaluating models.

XGBoost and LightGBM

Gradient boosting for structured, tabular data.

statsmodels and Prophet

Classical statistical forecasting. Prophet, from Facebook, handles seasonality and holiday effects and is now maintained mainly for fixes.

skforecast

Turns any scikit-learn-compatible model into a forecaster, with backtesting built in.

Chronos (Amazon) and TimesFM (Google)

Pretrained forecasting models that can forecast a supplied series without task-specific training on it.

Language models as forecasters

The TimesFM authors found time-series-only models beat them at lower cost. A 2025 study found text models competitive in some cases, so test on your own data.

Questions that start the conversation

• CEO to CIOWhich of our AI workloads run on market-driven data, and which on historical data? For the historical ones, did we build the analysis once, or do we pay a recurring fee for something that does not change?
• CIO to the teamDoes each task need machine learning, or would a reproducible analytical pipeline do it? Where we use models, which tools do we use: scikit-learn, XGBoost, LightGBM, statsmodels or Prophet? What is the retraining interval on each market-driven model, who decides, and who owns watching its accuracy? Who owns the model and the software around it, and its controls?
• CFO to CIOWhat exactly are we buying: people, software access, usage or a managed result? What does the lifecycle cost include?
• COO with the CEO and CIOWhich operations depend on a model that can go stale, and what happens to them when it does?