“Our employees are our greatest asset.” Every company says it. Bellwether is how you prove it — your people write down what they actually know, and sovereign AI makes it findable. Plugs into Oracle HCM or any HR portal. Runs on your hardware.
Ask any HR platform who can help build a quoting tool and it will return everyone whose profile says “Inside Sales”. That is a classification, not an answer. What actually determines who can build the thing is context that lives entirely in people's heads:
“Denver ships partial pallets on Fridays because the carrier cutoff is 14:00, and it isn't in the WMS.”
“The pricing exception logic looks insane because it preserves a 2004 AS/400 behaviour.”
“Customer IDs below 40000 are ambiguous — every trend older than 1998 is quietly wrong.”
None of that is in a system. All of it is load-bearing. And it walks out of the building the day the person who knows it retires.
Every company now has more data than it can read. Warehouses, lakes, ERP tables going back decades. And almost none of it interprets itself — a number is only worth what someone can tell you about why it looks like that.
That interpretation is business context, and it is the most valuable asset in your data estate. It is also the only part that was never written down. It sits in the heads of the people who were there when the decision was made.
A model with your data and none of your context produces confident, well-formatted, subtly wrong answers. A model with both produces something your people recognise as true.
This is the part of AI adoption everyone skips. Companies connect the model to the warehouse, get plausible nonsense, and conclude the technology isn't ready. The technology was fine. It simply had no idea that Denver's Friday cutoff is 14:00, or that customer IDs below 40000 are ambiguous — and nobody had ever given it a way to find out.
This is not general knowledge. It is the specific, hard-won understanding that makes your company different from your competitor — the closest thing to a moat you have. It is the last data you should be uploading to someone else's model. Bellwether reads it, ranks it and answers with it entirely inside your network.
Context is what turns a general model into your model. Ask which of your people can brief an AI build, which knowledge is thinnest, who should be in the room. The company stops guessing at its own expertise and starts routing work to it.
“Our employees are our greatest asset” is the most repeated sentence in corporate life and the least evidenced. It appears in annual reports, on lobby walls, in every all-hands. And if anyone asks which asset, or how much, or what happens when it leaves, there is no answer — because nothing in the building measures it.
Every other asset class is instrumented to death. You know the depreciation schedule on a forklift. You know the carrying cost of the inventory in aisle six. You know, to the dollar, what the building is worth. The asset you call your greatest is the only one with no ledger at all.
Bellwether is that ledger. Not a headcount — a record of what your people actually know, in their own words, and what it would cost you if they walked.
Stated in every annual report. Backed by nothing an auditor could test.
Twenty-two things this organisation knows, attributed to the six people who know them, in language they chose themselves.
Four of those things have exactly one owner. One of those owners retires in nineteen months. That is a number a board can act on.
It also does something quieter and more useful: it makes the claim true. People who write down what they know get found, get asked, get credited. The sentence stops being a poster and starts being how work gets routed.
A manager assigns a form. HR owns the taxonomy. Employees pick skills from a dropdown someone else wrote. The data is stale in a quarter and nobody trusts it, because filling it in honestly has no upside and describing yourself modestly has a cost.
The employee writes it, owns it, and edits it whenever they like. It is a brag sheet, not a compliance task — the more precisely you describe what you know, the more often you are the person the work comes to. The incentive finally points the right way.
Plain language in. Ranked people out — each with the passage they wrote that justified the match, so a staffing decision is defensible rather than a score from a black box.
$ bellwether find "we keep getting phantom inventory variances at quarter end" Searched 22 knowledge entries from 6 people — locally, nothing left this machine. 1. Marisol Herrera (Distribution Operations Supervisor) match 0.72 ↳ Cycle count variances that aren't real "Reels get counted by footage on the floor and by weight in the system. On anything under 250ft the weight math is wrong often enough that we get phantom variances every quarter-end. I know which SKUs do it and I re-count those by hand." 2. Dan Whitcombe (Systems Analyst · IT) match 0.30
Real output from the sample registry. The evidence is the product: a name alone is a guess, a name with the paragraph behind it is an answer.
Which knowledge does exactly one person have? No HR system on earth can answer that today, at any price. Bellwether answers it by finding the entries nothing else in the organisation comes close to.
$ bellwether risk SUCCESSION RISK — knowledge no one else has claimed 1. Marisol Herrera — The Alameda dock height problem nearest anyone else comes: 0.30 2. Priya Raman — Which contractors need drawings before pricing nearest anyone else comes: 0.35 3. Marisol Herrera — Cycle count variances that aren't real nearest anyone else comes: 0.41 4. Dan Whitcombe — The COBOL copybooks nobody can read nearest anyone else comes: 0.42 "I can map them to modern field names from memory. That map does not exist in writing." If one of these people leaves, that knowledge leaves with them.
In the sample registry, Dan's own overview says he retires in nineteen months. Line 4 is what that actually costs, and it is the sort of thing an organisation normally discovers the week after the leaving party.
Embeddings and search run locally on ordinary hardware — no GPU, no external model,
no API key. Run with --airgap and any outbound connection raises at the
socket layer.
One file per person. Readable without us, portable away from us, diffable in whatever you already use. If Bellwether disappears tomorrow you still have everything your people wrote.
Authored and edited inside the HR portal your people already use — Oracle HCM or anything else. Their words, their entry, their call about what goes in it.
This matters more here than almost anywhere else. A registry of what your staff know is not just sensitive — it is the distilled competitive advantage of the business, written down in one place for the first time. It should never sit in a vendor's cloud. With Bellwether it never does.
Markdown registry, local semantic + keyword hybrid search, ranked people with evidence, succession-risk detection, registry health stats, air-gapped operation. Working now, measured on a sample registry.
The Oracle HCM write-back is designed, not built — today Bellwether reads a folder. Search ranking is good on domain terms and merely decent on vague conceptual queries; we show evidence precisely so you can tell the difference at a glance. No access-control model yet, which a real deployment will need.
Bellwether doesn't replace your talent review. It gives the weakest half of it something to stand on — and brings a 1970s grid into the era of the technology that can actually read it.
The 9-box does two jobs. One of them it does reasonably well: performance is observable and managers can assess it. The other job — what happens to this organisation if this person leaves — it answers with a manager's guess about "potential", recorded once a year in a room the employee never enters.
| The succession question | 9-box today | With Bellwether |
|---|---|---|
| Where the answer comes from | A manager's opinion | The employee's own written knowledge |
| What it measures | “Potential” — undefined, and it tracks visibility | Specific things known, in specific words |
| How often | Annual calibration meeting | Continuous — edited whenever they learn something |
| Who sees it | Usually not the person it describes | They wrote it. They own it. |
| What it tells the board | Who looks promotable | Which knowledge has exactly one owner |
Keep the 9-box for performance and promotion. Let Bellwether carry the succession axis — because “what walks out the door with this person” is a question about knowledge, and knowledge is now something a machine can actually read, rank and search.
The grid was built when the only way to assess a workforce was to ask its managers. That constraint is gone. Every person can now describe their own expertise in their own words and have it be genuinely findable — which is what the 9-box was always reaching for and never had the tools to do.
One thing we watch, and you should too: a registry written by employees rewards people who write well, which is the same visibility bias that undermines the 9-box, arriving through a different door. Bellwether pushes back on it with structured prompts instead of blank pages, and by treating a thin entry as a sign nobody explained the point to that person — not as evidence they know less.
Bellwether is early and being built alongside real organisations with deep institutional memory and a retirement wave coming. If you can name the person whose departure would hurt most, that is the conversation.
