// THE PRODUCTION RECORD
Our own record, misses included.
The industry number is that 95% of enterprise AI pilots return nothing. Rather than quote it and move on, here is our own record: every client system and product described on this site, and what actually happened to it. It is a small dataset by design, 14 systems, and it is our own, not an industry survey. It is not every project we have ever started: our failure taxonomy counts sixty-nine, and fourteen of them were built and never launched. We publish the ones we refused and corrected alongside the ones that shipped, because that is the whole point.
Record last updated
7 in production
1 pre-release
1 dormant
1 not deployed
1 refused
3 corrected
7
In production
1
Pre-release
1
Dormant
1
Not deployed
1
Refused
3
Corrected
The table scrolls sideways: swipe it for what happened to each system.
| System | Sector | Outcome | What happened |
|---|---|---|---|
| Water-quality consumer product | Energy | In production | Shipped after several earlier approaches failed. Publishes water-quality measurements against federal legal limits across roughly 300 Connecticut ZIP codes. |
| Operations unification across five platforms | Insurance | In production | Five disconnected systems reconciled into one operational picture; the engagement is ongoing. |
| Property enrichment pipeline | Distribution | In production | 101,862 properties in scope, roughly 40% enriched. We found and reported our own dedup bug to the client rather than absorbing it quietly. |
| Court intelligence pipeline | Legal | Dormant | We found the original discovery loop spending money and writing zero rows, and rebuilt it on the county bulk feed. It works. It is also dormant: the wider lead pipeline it was meant to feed was never built out, and no lead from it was ever converted. A correct diagnosis is not the same thing as a business. |
| Outage-intelligence pipeline | Energy | In production | A live outage-intelligence pipeline. Its public product is not showcased on this site yet. |
| Country-risk intelligence platform | Enterprise security | In production | Country risk tied to enterprise exposure, built on official government advisories and open data feeds; ongoing. |
| Daily customer-intelligence briefing suite | Operations | In production | Automated daily briefings generated from unified operational data. |
| Voice agent for live customer calls | Operations | In production | Handles live inbound customer calls. |
| Edge SEO and AI-discovery layer | Workplace safety | Not deployed | Built and delivered: a CDN-level worker that adds titles, schema and structured credentials to a site its host platform locks down. It has never run in production, because the client's DNS was never moved to the CDN it depends on. Delivered work that never ran produced no value, and we say so. |
| Field-workforce productivity analysis | Operations | Corrected | Our first pass put field-technician productivity at 31%. Our own follow-up audit found 57% of the supposedly unproductive hours within 100 meters of a real customer site; we corrected the figure to 43% and withdrew five of seven names from a suspected-moonlighting list. |
| Notification approval console | Operations | Corrected | It marked time-sensitive campaigns as sent without ever calling the mail vendor, and the audit trail agreed. We fixed the send path and added an expiry so a stale campaign dies instead of going out late. |
| Three predictive ML models | Multiple | Refused | Built and tested. Accuracy collapsed once we removed leaked labels and circular targets, so we kept them shadow-only and shipped none of them. |
| Edwin (AI memory product) | Own product | Pre-release | Built and used internally; not yet generally available. Its own pre-registered benchmark found it tied a full-context baseline on accuracy, and it publishes that result. A provisional patent application is filed; two of its 35 experiments disproved our own hypotheses, and the filing discloses them. |
| Consumer marketplace directory | Own product | Corrected | Added a confidence spine that grades every listing by its provenance and serves it by trust tier instead of hiding weaker rows. Established that 906 of 1,290 active listings were genuinely bookable, meaning a real future session and a real price. |
Methodology
- In production means a system is processing real data in production use.
- Dormant means we built it, it works, and it is not being used. We report it separately because counting it as production would be the flattering answer rather than the true one.
- Pre-release means built and in internal use, not yet released to anyone else.
- Not deployed means we built and delivered it, and it was never put into production.
- Refused means we built and tested it, then chose not to ship it.
- Corrected means we found and fixed a material error in our own output.
- Each system appears once, under the outcome that says the most about it, so a Corrected or Refused entry can also be a running system. The homepage receipts are counted per finding (corrections, refusals and one publication), not per system, so one system can carry more than one; the token-cost receipt corrects our own estimate and the advice not to build is a recommendation, so neither is a system. Sector is the client's industry; Operations where we have chosen not to tie a system to a client's industry; Multiple where one row spans several clients; Own product where there is no client.
- This is our own engagement record, not an industry survey, and the sample is small on purpose. Two security-assessment engagements are excluded pending review, so they do not appear here.