Modern Gravity is generally availableBlog ↗
Seen, but not chosenNational moving and storage company
The complaint I hear most about AI in a company is not about the model. Most companies do not yet have one agreed meaning for their key numbers. Even where the data team has governed the warehouse, what each number means still lives in people's heads. Then a leader asks a plain question, and the answer is almost right, or right with no way to say why.
That gap is the last mile of AI. This post is for data and operations leaders who are stuck in it.
My view, after 20 years of this work: the last mile closes fastest from the top down. Start with what your executives decide on, let AI build the business context and draft the definitions, and have your owners certify them.
Bottom up is the right foundation, and slow to reach the executive question
Governance tools and semantic layers, such as Snowflake, dbt, Microsoft Fabric and Databricks, are built by IT from the data up. Tables become models, models become metrics. That is the right foundation, and Gravity works with it.
The problem is where it starts. A bottom-up layer describes the data you have. An executive asks about the business they run: why on-time delivery slipped, which stores will run short, what a late shipment will cost.
Closing that gap the traditional way means meeting after meeting, leader by leader and team by team. What counts as on time? Which returns count? Whose inventory number is right? Each answer is reasonable. Agreeing on all of them takes months, and the executive needs the answer on day one.
Where most companies really are
Most companies think they are further along than they are. Here is how I see the stages.
- Everyone has a number. Each site and function keeps its own version of the same KPI. Meetings start with whose number is right.
- Governed data. There is a warehouse, a catalog and some certified datasets. The data is clean, but what each number means still lives in people's heads.
- Owner-certified metrics. Every number has an owner who certifies its definition and current source, matched to the management reports the business already runs on.
- AI that understands the business. The AI works from those certified metrics, a map of how the business runs and its key drivers. Every answer is checked before a leader sees it.
Governance tools get a company to stage 2, and that is real progress. But stage 2 only feels like the finish line. Trust starts at stage 3, and the last mile of AI is stages 3 and 4.
Gravity builds from the top down
Modern Gravity™ is built for stages 3 and 4, and it starts at the other end: the reports your leaders already run on, such as executive dashboards, QBR decks and KPI sheets. Nobody starts from a blank page. It works in two parts that are built together, one business area at a time. AI does the drafting on both. Your people make the calls.
Context: a map of how your company runs
AI builds your business ontology from those reports and an industry map. The map connects every function to the physical flows, business processes and drivers behind its numbers, so you can trace how a change in one part of the company affects the rest.
In supply chain, that is how one late factory shipment flows through to stock, chargebacks and lost sales. This map is what lets AI answer why a number moved and what to do about it, not only what the number is. Gravity walks the drivers back to the root cause.
Certification: definitions your owners sign
AI drafts each definition from the same reports and flags where teams disagree. Your business owners review and sign in plain language, not SQL.
Every answer the AI gives is then checked against what they signed. Agreement that took months of meetings becomes a review. It is still real work for the owners, just a lot faster.
| Bottom up (governance tools) | Top down (Gravity) | |
|---|---|---|
| Starts from | The data and its tables | The executive reports and decisions |
| Built by | IT and the data team | AI drafts, business owners sign |
| Answers well | What the data says | What happened, why, and what to do |
| Checks | The input: approved definitions and SQL | The output: every answer the AI gives |
The pitfalls I see most
These come up again and again, at companies of every size:
- Rolling out AI agents before there is one agreed definition of the metrics they will quote.
- Treating a certified dataset like a certified answer. Clean data can still produce a wrong answer.
- Moving to a new platform and carrying the old definitions over without anyone certifying them.
- Building the semantic layer from a blank page instead of the reports, docs and decks the team already runs on.
- Letting every site keep its own version of the same KPI and calling it flexibility.
The first two are the costly ones: an AI quoting numbers that nobody has certified. We measured what that looks like.
What happens when the last mile is skipped
In October we tested this. We asked Claude to list the figures for 8 leadership dashboards (headcount, joiners, leavers, offers), 3 times each, and graded every requested figure against a known answer key.
This was an internal test on synthetic company data. Claude alone had direct database access, with no semantic layer or certified metrics. The other arm was Claude working through Gravity. The comparison with a governance layer in place is next.
| Claude alone | Claude working through Gravity | |
|---|---|---|
| Runs with at least one wrong figure | 13 of 24 | 0 of 24 |
| Correct figures | 212 of 261 | 255 of 261 |
| Wrong figures | 49 of 261 | 0 of 261 |
| Figures held back with a reason | 0 | 6 |
What struck me was not the error count. Most wrong figures were off by 1 or 2. A headcount off by one does not look wrong on a dashboard. It looks like a fact, and someone acts on it.
If you want to see what this would look like in your business, the Gravity page shows how a Truth Sprint maps and certifies one business area in six weeks.
Top down meets bottom up
Governance tools are partners for Gravity, not competitors. If you already govern a semantic layer, Gravity works with it. If you have none, Gravity builds one from your reports. The two meet in the middle.
Snowflake's offering works well for the data team. Gravity does the hard work of the AI last mile.
We are adding this to our tests: the same AI with your governance layer alone, then with Gravity added. We will publish where Gravity helps and where it does not.
What this means for you
Start with one decision your leaders need to make, the numbers behind it and the owners who can certify them. The figure test shows the cost of skipping certification: plausible numbers that are wrong. With context and certification in place, Claude returned no wrong figures in 24 runs and held back 6 with a reason rather than guess.
If you lead data or operations, right answers are the minimum. From day one, the checked numbers and the map of what drives them give your leaders a basis for deciding which shortages to address, which shipment delays to escalate, and where chargebacks and lost sales are at stake.
The Truth Sprint is six weeks, one business area, one use case. Start on the Gravity page.
Where does your AI stall today: the answer, or the why behind it?
Figure test: internal, on synthetic company data, October 2026, Claude Code on a Mac, one model. 8 dashboards, 3 runs each. Claude alone had direct database access with no semantic layer or certified metrics. Only the figures asked for were graded, against a known answer key; no dashboard was published. No outside audit.
Results referenced are based on observed patterns in available data and are not guarantees of performance. Individual outcomes vary based on data quality, implementation, and market conditions. Modern AI recommends independent validation of any metric before business decisions are made.