Reading NVIDIA’s Auto Ontology project got me thinking about this. Many organisations I talk to are still experimenting. Individuals use AI to get parts of their work done faster, and teams are starting to build their first agents.

I think this is the right time to start two things in parallel.

1. Help people and teams work better with AI

Start with tasks people already do. Where does AI save time? Where does checking its work take just as long? Share those lessons so the next person can build on them.

Then pick a workflow the whole team repeats. For example, a reporting team might use AI to prepare a weekly summary and flag anomalies, with an analyst checking the figures before it goes out. Measure how long the whole workflow takes, including review and rework.

Uber Engineering offers a useful way to think about measurement. It breaks AI spend down into users, sessions per user, turns per session, requests per turn, tokens per request and price per token. It also tracks cost per completed outcome and quality. I’d bring that same distinction to business workflows. How many people use the tool, what does it cost to complete the work, and is the result good enough to use?

2. Build the technical and data foundations alongside that work

In BI, an agent can write a query in seconds. To answer a question about active customers, it needs to know how the business defines “active”, which data to use, and how to check the result. NVIDIA’s Auto Ontology shows one way to draft those definitions and relationships from database metadata, have people verify them, and make them available to agents.

Shopify’s Under the River shows the same idea in software engineering. A shared codebase, reproducible environments and written-down engineering practices give River enough context to work across teams, run tests and open pull requests.

The two tracks can inform each other. Start with one recurring workflow and use it to find out which definitions, permissions and checks the agent actually needs. Put those in place, measure the next run, and share what the team learns.

Which workflow would you start with, and what would an agent need to know to do it reliably?

Reference reading