AI disclosure
Version of September 11, 2026.
Outreach and support emails from Findingbase are prepared with software assistance and reviewed under the policy set out below. This page says exactly where that assistance sits, and where it does not.
The product contains no AI
The renewal score is arithmetic. Each contract starts at zero and gains a fixed number of points for each of five conditions found in the federal award record — a run of consecutive awards to the same vendor, an option already exercised, limited competition or two or fewer offers, a set-aside, and a current end date that is also the potential end date. The total is capped at 100 and every point is reported in words alongside it.
No model ranks your alerts, chooses which contracts you are shown, writes your digest, or sees your dashboard. The digest email is a template filled with rows from a SQL query. If a contract has not been through scoring yet, it is labeled not yet scored, because printing a zero would be a judgment we have not made.
Where a language model is used
In one place: sorting replies to the outreach email we send to federal vendors who are not customers. When someone replies, a language model reads that reply and classifies it — interested, not interested, unsubscribe, out of office, wrong person, bounce — and may draft a suggested response for a person to review.
Two engines can serve that step, and which one runs matters for where your words go:
- A hosted model, within a weekly budget. The provider is Anthropic, PBC (San Francisco, United States), and the model is one of its Claude models, used under Anthropic’s commercial terms and privacy policy. When it is used, the text of your reply is sent to Anthropic for classification. It is sent on its own: no name, no email address, no record about your business, and nothing about any customer goes with it. Anthropic’s retention of that text is governed by those terms; we do not grant it any further use of what we send, and we keep the number of such calls under a weekly ceiling that we set ourselves.
- A model running on our own server. When the weekly budget is spent, or the hosted call fails, classification falls back to an open-weights model we host ourselves, on the same French server as the application. Nothing leaves our infrastructure on that path.
The model is given the reply and nothing else. It has no access to the contract database, to any customer account, or to the alerts customers receive, and it cannot send email: it returns a label and a draft, both of which land in a queue.
How outreach email is written
Outreach messages are assembled from templates we wrote, merged with facts a business published itself in its SAM.gov registration — its name, its NAICS codes, its state — and the expiring contracts that match them. The sequence is approved by a person before it can run, and every message carries a real sender identity, a physical mailing address, a one-click unsubscribe link, and a line saying it was prepared with software assistance.
The review policy
- A person approves every outreach template and every change to one before it sends.
- No reply drafted with model assistance is sent without a person reading it first. The draft is a starting point, not an outbox.
- A classification never triggers an irreversible action on its own. Unsubscribes and suppressions are applied by rule, from the link you clicked or the words you used, not from a model’s opinion.
- Model output is never used to decide a price, a plan, a refusal of service, or anything about a customer’s account.
- The person responsible for that review is Styvan Hauterville, the operator of Findingbase, who reads every outreach template and every drafted reply before it goes out. There is no one else it could be delegated to: this is a one-person business, and that is who answers for it.
If you would rather not be included
Reply to any message and say so, or use the unsubscribe link. The address goes on a suppression list and is not written to again. You can also write to contact@findingbase.com and ask us to delete the reply we stored, and we will.
Where this stands today
As of September 11, 2026, the scoring described above is running and no prospecting email has yet been sent from this system. This page describes what runs when outreach starts, and will be updated if any of it changes. The privacy policy covers what is stored either way.