Every Number Traces to the System of Record
For a mission-driven organization, statistics are not decoration. They are the case for support. 571 Scholars. More than $30 million committed. Over 350 partner schools. A donor reads those numbers and decides whether to trust you with their money. A family reads them and decides whether to trust you with their future.
When dealing with such important numbers there are two distinct failure modes, and both are unacceptable
- Stale Numbers: the site says 135 partner schools when the real figure passed 350 long ago, which is exactly the gap No Greater Sacrifice was living with on their old platform. The organization always knew its real numbers. The site just couldn't keep up.
- Hallucination: This occurs because modern AI rarely fails by drawing a blank. This means it fails by producing something plausible, but confidently wrong.
The case for support
The near-miss that proves the point
During the NGS build, we were creating an interactive map showing where Scholars live across the country. The AI proposed a beautifully rendered, entirely reasonable-looking distribution, including a specific count of Scholars in a state where the true number was zero.
Nothing about it looked wrong. It rendered perfectly. It was proportional and professional. It was also fiction: a "reasonable placeholder" that would have gone in front of donors as fact.
It never shipped, because it hit a wall we had already built. No figure appears on the platform unless it traces to the system of record. The real map data came from a direct query of NGS's donor database, actual records and actual states, and automated tests now pin those values so they cannot silently drift.
That is the honest lesson about AI and data. The model is not malicious. The model is helpful, but helpfulness without provenance is how fabricated numbers reach donors. The defense is not better prompting. The defense is a rule with no exceptions.
How it works: one source of truth, wired everywhere
The architecture is simple to describe.
Salesforce is the single source of truth. NGS's donor database, where every Scholar, gift, and school relationship already lives, is the authority. The website is a view of it, never a second copy of it.
Verified facts live in exactly one place. Every canonical statistic was verified against the database, ratified with the client, and stored once. Site pages do not contain the number "571." They contain a reference to the Scholar count, resolved from the canonical source when the page renders.
Update once, correct everywhere. When a new Scholar cohort arrives, the figure changes in one place and every page citing it get simultaneously updated. The category of error where page A says 571 and page B still says 493 is structurally impossible.
Automated review watches for strays. A content scanner checks every published page for hardcoded figures that should be references, and for retired numbers that ratified language decisions removed. Drift gets flagged before a donor ever sees it.
The result for the organization: the website stopped being a thing that goes stale. It is now downstream of the same system NGS already keeps meticulously current. The human discipline that makes rules like this stick is its own story, told in The Most Important Layer of Our AI Stack Is Human (opens in a new tab).
The most sensitive data gets layers, not vibes
One more real moment, told carefully because the data deserves it.
Part of the build involved connecting Scholar profiles on the website to their corresponding records in the database. Automated matching handles the volume, however names are treacherous. Families can include multiple people with similar names, and a "close enough" match on this data means one person's page carrying another person's story. For an organization serving Gold Star families, that is not a bug class. That is a breach of trust.
So the pipeline is layered. An automated uniqueness safeguard hard-rejects any two profiles claiming the same database record to block a plausible, but fabricated, match between two Scholars with similar names. A second, independent audit re-verifies that every linked record's name actually agrees with the profile it is attached to, and it caught a separate near-miss the first layer could not see. Above both layers sits the final rule: nothing publishes automatically. Every change lands as a draft. A human approves anything that touches a Scholars's story.
Safeguard, audit, human. Three layers, because the data warranted three layers.
What this means if you're evaluating AI for your organization
Three questions worth asking any partner, including us.
"Where does this number come from?"
If the answer is not a named system of record, the number is a liability with a publish date.
"What happens when the AI doesn't know?"
The dangerous answer is "it does its best." The right answer is "it is structurally prevented from inventing, and here is the mechanism."
"When a figure changes, how many places need editing?"
The right answer is one.
The AI era does not lower the bar for factual discipline. It raises it, because the failure mode got quieter. Plausible fabrication does not look like an error. It looks like content. The organizations that thrive will be the ones whose platforms make provenance structural.
NGS's platform does. Every number on nogreatersacrifice.org (opens in a new tab) traces home.
No Greater Sacrifice
Next read: Built to Be Cited: Winning AI Discovery
Steve Hamilton
SVP, DXP and Custom Solutions Practice
Stay Informed
Get industry-leading insights delivered to your inbox.
Industry Leading Insights
Our latest thinking on personalization, digital transformation and experience design
