The question was governance. The answer was cost.
The table had been set for governance: where AI pilots actually break, who owns a data product when it fails, and what has to be true before an agent goes near a regulated or customer-facing process. The first substantive answer of the night was about money.
One leader described the AI now arriving inside enterprise software her company already licenses. The capability was not the problem. The pricing model was. She could not yet build a cost-against-return case that her internal customers in finance, who know their numbers precisely, would accept, and she was asking her software partners for help with it.
From there the table kept coming back to economics from different directions. A model a step behind the frontier already answers most of what an enterprise asks. If so, paying frontier prices for routine questions is a choice, not a necessity, and several guests predicted that the architecture that wins inside companies will route each task to the cheapest model that does it well. The counter-argument was put just as firmly: as frontier prices fall, the savings shrink, and open models arrive without the surrounding support, so reaching the same quality takes real work.
Nobody argued that cost was the only question. But it was the one everyone could answer from their own budget, and it pulled every other topic toward it.
Rot is now a line item
The observation that stayed with the table joined the governance argument to the cost argument. Every enterprise holds data nobody uses: duplicates, records past their useful life, copies of copies, fields nobody ever defined. That used to be a housekeeping problem. Storage is cheap, so keeping it cost little, and deleting it never reached the top of anyone’s list.
AI changes the arithmetic. A model does not skip what nobody needs. Retrieval indexes it, pipelines transform it, prompts carry it, and the meter runs on all of it. Data that was merely untidy in a warehouse becomes expensive the moment an AI system has to read it, and more expensive again when the answer it produces is wrong and someone has to catch it.
Ungoverned data is no longer merely untidy. It is measurably expensive to process.
That is why we think governance belongs inside the AI business case, not beside it. Retiring, reconciling and curating data used to be justified on risk and tidiness, arguments that rarely win a budget. It can now be justified on run cost, which finance already measures. The same discipline sits behind an AI business case a CFO will fund: count the costs that the first estimate leaves out.
What enterprises are actually buying
A second consensus surprised some at the table. Once a cheaper model answers well enough, extra intelligence is rarely what decides a purchase. What buyers are paying for is security: clarity about how their data will be used, the confidence to defend a decision, and, in some contracts, a vendor that stands behind the output legally. Vendors that are easier to negotiate with about data use were said to be winning enterprise business on that alone.
The worry is wider than personal data. It covers confidential and proprietary information of every kind, formally secret or not: what goes into a prompt, where it is kept, and whether it ends up improving someone else’s model.
Ownership is a board question now
The most animated stretch of the evening was about who owns what. The table discussed a recent copyright settlement involving an AI developer in which, as the room read it, the outcome turned on how the training material was acquired rather than how it was used. It discussed the asymmetry by which a large publisher can afford to litigate and a small company cannot, and the way voice and likeness are protected only where a contract says so.
Data Products has been on the receiving end of this. A company approached us to buy our data to train a large language model, and we said no. Smaller companies are being approached all the time, and many accept for very little. That is how ownership quietly changes hands, and it is why we now treat a data inventory as a list of assets as well as a list of liabilities.
Illinois came up as the counterexample. Its biometric privacy law requires a company to disclose why it collects biometric data and how long it keeps it before collecting any, and it was passed after a company that let shoppers pay with a fingerprint went bankrupt, leaving open the question of who would get the fingerprints. One view at the table was that a law that strict would struggle to pass today.
The best return so far is unglamorous
Asked where AI had actually paid off, one of the most sophisticated analytics teams at the table gave the least glamorous answer of the night: cleaning up technical debt. Systems that grew in awkward ways over years can be rebuilt the way they should have been built in the first place, and AI makes that work faster.
The same is true of data. The cheapest token is the one you never spend, and in many organizations the first job for AI should be finding what can be retired before anything is fed to a model.
The table was just as firm that automation without transformation fails. Laying AI over an existing process produces cost without value, and the resistance is as human as it is technical: people comfortable with a legacy process are attached to it, and everyone wants someone else to change first. Many of the problems brought to internal AI teams turn out to be process problems with an AI label, which is why adoption is a business decision, not a launch event.
Self-service needs a curator
The contribution most often repeated afterwards came from a team that is deliberately holding back from putting a conversational interface on top of its data products. The reason is not the technology. Without the context around a data product, meaning its definitions, its caveats and the question it was built to answer, it is easy to misread, and a fluent interface makes misreading faster. Whether that caution is temporary or permanent, they could not yet say.
The wider diagnosis was that most organizations lack a curation layer: someone who decides what is fit to go into an AI system in the first place. Bad inputs do not produce proportionately bad outputs. They produce more of them, faster, delivered with confidence, and models will infer connections that were never in the data. The safeguard the table preferred was not a person somewhere in the loop but a person in the lead, accountable for what the system is allowed to decide.
What to do before the next model upgrade
- Find the rot before you feed it. List what your AI systems actually read, and retire or archive what nobody uses before it is indexed, embedded or put into a prompt.
- Put data quality in the business case. Price the processing, review and rework that poor data causes, so that cleaning it competes on run cost, not on tidiness.
- Name a curator. Decide who says what is fit to go into an AI system, and give that person the authority to say no.
- Route by task. Send each request to the cheapest model that does it well, and keep the most capable models for the work that needs them.
- Treat your data as an asset. Know what you own, what you have licensed and what you would never sell, before someone offers to buy it.
The September table took this argument one step further, into how AI cost should be measured at all: Measuring AI cost in the wrong place.
Where to start
- A Data Readiness reviewWhat your AI reads, what it should not, and what fixing the data underneath would change.
- Value & EconomicsWhat an AI system is worth, what it costs to build and what it costs to run.
- Private Executive BitesThe tables this article comes from, and how to request a seat at a future one.
