How FedM8 uses AI
Some FedM8 features use AI models. This notice lists them, names the companies that run the models, says what each feature sends and what its scores mean, and explains how AI output can be wrong and what to check. FedM8 LLC ("FedM8," "we," or "us") publishes it. The name "FedM8 AI" is the name of the service; not everything in it uses AI.
1. Two kinds of AI in FedM8
- Language models write text. They write summaries, requirement lists, first drafts, and the pinned FedM8 AI note. FedM8 sends these requests through its own AI gateway (LiteLLM, which runs on Railway) to OpenRouter, which passes them to OpenAI's GPT-4o mini and GPT-4o models.
- Jev makes judgments. Jev is a model run by TypeSafe. It does not write text. It gives scores, rankings, matches, and answers to narrow questions, such as whether a draft answers a requirement.
FedM8 can turn AI features on or off. If FedM8 changes the companies or the models named here, it updates this notice.
2. Features that use AI
This list includes places where the screen does not say AI, such as the one-line summary at the top of a grant record and the Match-M8 chips.
- AI summaries of notices and grants. A language model writes a short read of a published notice and its attached files. The box is labeled "AI summary". A grant summary is written from the Grants.gov listing, not from every file attached to it.
- The pinned FedM8 AI note. In contract and grant conversations, the first note is a short summary written by a language model. It is labeled "FedM8 AI", "Pinned", and "AI brief", and it is never a member.
- Win Kit (Pro). A language model reads the notice and its files and lists what they ask for. Win Kit keeps a line only when its quote is found word for word in the source. A language model also writes a first draft from that list, your company profile, and your Firm Library, and marks each fact only you can supply as NEEDS_HUMAN. Your contact details stay out of the model call. The bid decision uses written rules first, then Jev; no language model. When you ask Win Kit to check a draft, Jev reads it against each requirement. Win Kit has a daily limit per account.
- Best fit (Pro). Jev scores notices against your company profile in the background. The page shows Fit, Worth pursuing, a group such as Strong or Check eligibility, and when the scores were made. Its four tools also use Jev. Shred labels the statements in a notice. Red-team a draft shows which requirements look unanswered. Check a CO question says whether the notice likely answers your question already. Teaming partners suggests firms from public award data.
- Match-M8. Chips such as "Chase", "Pass", and "Needs a human", with a percentage, come from the same kind of stored score. Your home page can rank work by Match-M8 using your NAICS codes, states, certifications, and stored scores. The weekly market email can show the same score as "pursue" with a percentage.
- Team Up matching. Jev scores how well two companies fit, from their public team profiles: shared NAICS codes, role fit, certifications, places, and capability words. If you upload a resume, Jev reads part of its text to find work you can do. Other members never see your resume text.
- Proposal Studio. The requirement list and bid brief that Proposal Studio loads are built by written rules from the notice's files. When FedM8's AI gateway is available, a language model may refine them.
3. Features FedM8 can turn on later
These features are built but run only when FedM8 turns them on. The plans page shows which credit tools are available.
- FedM8 AI Assist chat. A language model answers your questions with a fixed list of tools for looking up notices, awards, and your saved records. It has no tool to send, sign, or submit anything. Jev can rank the search results it finds.
- Notice reviews and coverage suggestions that spend AI credits.
- AI search. Jev reads your words to work out what you are looking for, such as a set-aside.
- Decision briefs and catch-up summaries of long conversations, written by a language model.
4. Tools that do not use AI
- Check the record: a plain search of a notice's official record or a grant's public synopsis.
- Win Kit's past award prices.
- The posting check in Chat and private messages, which is a set of written rules.
5. What the scores mean
- Fit and Worth pursuing. A percentage is a model's estimate of how well a piece of work fits your company, based on the data FedM8 has. It is not a chance of winning an award, and it is not a promise of eligibility.
- Groups. Best fit sorts notices into Strong, Possible, Weak, Check eligibility, or Not eligible, from the scores and the profile data FedM8 has. A group is not an eligibility ruling. Only the official notice and the agency decide who may bid.
- Low confidence. This tag means the model was unsure. Read the notice yourself.
- Chase, Pass, and Needs a human. A stored score of 70% or more shows Chase. A score of 30% or less shows Pass. Anything between shows Needs a human. These are suggestions, not decisions.
- Team Up matches. A match score says how much two profiles have in common. It is not an endorsement of either company.
Scores and summaries are made at a point in time. FedM8 shows when Best fit and Match-M8 scored a notice. A notice can change after that.
6. How AI output can be wrong
AI output can be wrong or out of date. It can:
- miss an amendment, an attachment, or a file FedM8 could not read;
- misread a scanned page or a table;
- state something with no source, or cite the wrong source;
- rely on data that is out of date, for example when a government source has stopped updating;
- get a date, a deadline, or a time zone wrong;
- leave out a required clause, form, or step;
- misjudge eligibility, a set-aside, or a cost; or
- be steered by text hidden in a notice. FedM8 tells its models to treat notice text as information, not instructions, and screens text it collects for instructions, but this is not perfect.
FedM8 does not promise that any AI output is complete or correct.
7. Check the official source
Before you rely on AI output, check it against the current official notice on SAM.gov or the listing on Grants.gov, every amendment, and the agency's instructions. The official source controls. Ask the contracting officer about a requirement in writing, or the agency contact listed on Grants.gov about a grant. AI output is drafting and research help, not legal, financial, or other professional advice. You are responsible for everything you submit, sign, or certify.
8. AI does not act for you
AI features in FedM8 cannot sign, accept a team or an agreement, certify, submit a bid or application, send a message, or commit funds for you. Win Kit does not send anything. Members write their own posts and private messages; AI never posts as a member.
9. What FedM8 sends to AI providers
To run an AI feature, FedM8 sends the text that feature needs. Depending on the feature, that can be:
- the text of public notices and grants and their files;
- the business profile fields you entered, such as company name, UEI, NAICS codes, certifications, places, and capability statement;
- your past performance entries and your company's, when Win Kit writes a draft, and a draft you ask Win Kit to check;
- your Team Up profile and part of your resume text, for matching;
- your words, when you use the Assist chat or AI search, and the records the Assist chat looks up for you, such as notices, awards and your past performance entries; and
- a conversation's recent posts, for a catch-up summary.
Veteran confirmation details are never sent to AI models. FedM8 keeps the text of your FedM8 AI prompts and the replies in your activity history for 90 days, as section 1 of the Privacy Policy explains.
What to leave out. Do not put anything into an AI feature that the Community & Acceptable Use Rules say to keep out of FedM8. That includes classified information, CUI you are not allowed to share, export-controlled data, credentials, Social Security numbers, and medical, tax, or banking records the task does not need. If you do not want your resume read for matching, do not upload one, or remove it in Team Up.
10. Training
FedM8 does not use your content to train AI models, and does not sell it or give it to anyone for that purpose. FedM8 sends data to OpenRouter, OpenAI, and TypeSafe only to produce results for FedM8's features. This notice does not describe what those companies do with data on their own systems; their own terms and privacy policies cover that.
11. Changes to this notice
FedM8 updates this notice when its AI features, or the companies or models named in it, change. Each version has a number and stays published, unchanged, at its own address on fedm8.com.