From what your team writes to answers it can stand behind.
Favel reads the material your team already produces, prepares the decisions in it, and a person approves them. The assistant answers from what was approved. This page walks through each step.
8 min read
It starts with what your team already writes.
Favel reads the source material, from meeting transcripts to the latest budget table, and surfaces what matters for your company.
It reads each source for what it could add, and keeps the passage the addition came from. Then it sets what is new against what already holds.
What comes out is not a summary. It is a decision, prepared for the person it belongs to. It arrives with the material already extracted and a title, so you start from something readable, not a transcript.
Only a person decides what holds.
Nothing the company relies on gets there without approval. The decision shows the contradictions and gaps: you choose a direction, or add what is missing.
The decision type says what is being asked. A choice puts two directions side by side. An addition names what is missing, such as an owner nobody has named yet.
Why keep a person in the loop at all? Because removing the decision is what every other assistant promises, and it is what produced the problem you probably have today: an answer nobody can stand behind.
The bottleneck is the decision, not the reading. A document does not record whether a decision still holds, and only a person can decide. That is why the gate is part of the product.
Decisions come from where the work happens.
Not only from a document someone uploads. A finding, a resolution, or a conclusion reached with the assistant can be submitted as a decision.
It is prepared like any other, it goes to the person who owns it, and nothing the assistant reads changes until they approve it. Results produced in the assistant go back for a person to approve.
The assistant answers from what was approved.
The assistant knows your Canon, the knowledge your team approved, and the record every answer carries.
Canon is the name for that set: the approved decisions, kept as one maintained set, where each entry names who approved it and what it came from. The context is set before you ask, so you do not rebuild it in every new chat.
A document is written for a person, so a model reads it and derives the decision from it again on every question. Canon holds the decision as an entry: what holds, who approved it, and when. The model reads the decision rather than deriving it from a document.
The assistant has no write access, and what it proposes goes back through the gate. It can write a clumsy sentence, and an answer can be wrong. It cannot answer from a decision nobody approved.
An answer traces back to the person who approved it.
Each answer shows the decision it comes from, the passage, and who approved it. Followed all the way, that is five links, from the answer back to the source it came from.
The middle link is the one a search cannot give you. A search can point at a document. It cannot point at the person who decided that what the document says still holds.
A change goes through the same approval.
A change is a source, and a source goes through the same approval as everything else. Favel sets what it adds against what already holds and prepares a decision for the person it belongs to.
Nothing changes in what the assistant reads until a person approves it, and the earlier version stays in the record.
A document cannot say whether it is still true. Nothing supersedes it, so an outdated answer looks exactly like a current one. Canon records what still holds, entry by entry, with what each one replaced.
What holds is not a snapshot that ages quietly. It is a set a person keeps current.
More context does not make a better answer.
The usual way to give an assistant your company's knowledge is to let it search. A search pulls in your documents, and every one of them becomes context in the prompt. That line only grows.
Answer quality does not: it peaks and then falls.
"even when models can perfectly retrieve all relevant information, their performance still degrades substantially (13.9%–85%) as input length increases …"
Favel works from Canon instead, a small set where every entry is a decision a person approved, so the assistant reads less and all of it still holds.
Why not just use what you already have?
A document can contain a decision. It cannot record whether it still holds, who approved it, or what it replaced. Only a maintained set records that: Canon.
A general chatbot, a search that indexes everything you wrote, and the connected storage behind it fail the same way: with no context, or with too much context to handle effectively.
| Your chatbot, your search, and your storage | Favel | |
|---|---|---|
| The source | The document is where the answer comes from. | The document is input. The answer comes from a decision. |
| The evidence | The answer quotes the first passage it finds. | The answer comes from what a person explicitly approved. |
| What holds | A document cannot say whether it is still true. Nothing supersedes it, so an outdated answer looks exactly like a current one. | A set records what still holds, entry by entry, with what each one replaced. |
| The timing | The decision comes later, if it comes at all. | The decision happens before the answer. |
| The context | You rebuild the context in every new chat. | The context is set before you ask. |
Your chatbot, your search, and your storage: The document is where the answer comes from.
Favel: The document is input. The answer comes from a decision.
Your chatbot, your search, and your storage: The answer quotes the first passage it finds.
Favel: The answer comes from what a person explicitly approved.
Your chatbot, your search, and your storage: A document cannot say whether it is still true. Nothing supersedes it, so an outdated answer looks exactly like a current one.
Favel: A set records what still holds, entry by entry, with what each one replaced.
Your chatbot, your search, and your storage: The decision comes later, if it comes at all.
Favel: The decision happens before the answer.
Your chatbot, your search, and your storage: You rebuild the context in every new chat.
Favel: The context is set before you ask.
Favel is for the company that decides.
Favel asks one thing of you: a person has to decide. That fits some teams better than others, so here is who we build it for.
Fits you
- A leadership team that owns what the company has written down and runs a formal strategy layer: a plan, OKRs, an offsite cycle.
- AI already in use: a company-wide AI search, or a chatbot you use ad hoc.
- You want a centrally governed, company-wide AI system.
Does not fit you
- A company with no written state. If nothing has been decided, there is nothing to approve.
- A team that wants the assistant to take the decision off its hands. That is the one thing Favel will not do.
- Anyone shopping for document search, a meeting notetaker, or a wiki. Those answer a different question.
If that sounds like your team, join the waitlist below. We build with selected design partners, and we review the entries and hand-pick who fits best.