A knowledge base for AI sales replies

AI replies are only as accurate as the facts they read. Write down one offer at a time: what you sell, who it is for, the price you are willing to quote, the links you share and the rules you work by, then keep objections and proof as separate linked notes. In AY Inbox this knowledge base is a campaign brain, and every reply drafted for that campaign reads it.

A brain is a named set of facts and material about one offer. It never sends anything on its own. It is what the inbox agent and auto-reply read before they write, so a gap in it becomes a vague reply and a mistake in it becomes a confident wrong one. The product reference is the campaign brain guide; this page is about what to put in it.

One knowledge base per offer

Write one brain per offer, not one per company. A workspace can hold up to 50, one of which can be the workspace default, and a campaign reads the brain assigned on its Brain tab, else the default.

The offer-specific fields (offer, audience, pricing, booking link, links and material) come only from the brain that applies. A field left empty stays empty, so a campaign brain with no pricing never borrows the default brain's price: a reply on that campaign says it will check, and the Brain tab warns you. That is deliberate. Two offers with two prices should never leak into each other's replies.

Some fields do fall through, because they describe how you work rather than what you sell. Tone comes from the campaign's brain, then the default brain, then your voice profile. Internal rules come from the campaign's brain, then the default. Do-not-say phrases combine both lists.

The fields that reach a prospect

Every field at its limit is read whole by every reply, so short, exact fields beat long, hedged ones. If you would not say a sentence to a prospect, it does not belong in these fields.

Automatic replies are stricter still: their links and figures must also be on the campaign's approved lists. When the brain holds a link or figure that is not approved yet, the auto-reply panel offers it for approval. See safe automatic replies.

The fields that never reach a prospect

Internal rules (up to 800 characters) shape what the AI proposes and are never quoted: the lowest discount you accept, who you do not sell to, when to hand a deal to a founder. Do-not-say holds up to 15 phrases, each up to 60 characters, that replies must not use.

After a draft is written, AY Inbox checks it and flags a figure or a run of words from your internal rules, and a phrase on your do-not-say list. Automatic replies go further and never read internal rules at all. Reading or editing them needs the Manage campaigns permission.

Material and pages

Material is everything that is not a field: case studies, objections, proof, an FAQ. A brain holds up to 30 sources and 500,000 characters in all, each source up to 200,000 characters, as pasted text, a file (.txt, .md, .csv, .html or .docx, up to 10 MB) or a public https web page. Pinned sources are always read first.

Pages turn that material into a small vault of linked notes: one page per persona, objection, case study, proof point, competitor, FAQ or step of how you work. Write [[Page title]] to link one page to another. A brain holds up to 300 pages and 400,000 characters of pages, each page up to 40,000 characters.

The links matter at reply time. A page that matched the prospect's message also brings in the opening of each page it links to, so an objection page pulls in the case study it points to. To build an objection library that works this way:

1. Write one page per objection, titled the way a prospect says it ("Too expensive").
2. Put your best answer on that page, in a few sentences.
3. Link it to the case study or proof page that backs the answer.
4. Link each persona page to the objections that persona raises.

Public or internal

Material and pages are internal by default: the AI may use them as background but is told never to quote them. Turning on OK to quote makes a source or page public.

Internal is the default because material is usually written for your team, not for a prospect. A case study draft, a pricing spreadsheet or a sales call note can hold customer names and numbers nobody approved for outside use, and quoting one should be a decision someone made, not an accident.

Automatic replies read only public facts: the offer, audience, pricing, links, booking link and public material, up to 6,000 characters. They do not read pages yet.

Let AI organize it, not invent it

Organize with AI reads what you sell, who it is for, the pricing and the material, and writes linked pages from it: every objection links to the proof that answers it, and every persona to the objections they raise. It is instructed to use only facts in your material and never to invent a customer, figure, price, quote, date or link. A run reads up to 30,000 characters of material and writes up to 24 pages; the rest waits for the next run.

Pages it writes are internal. A page a person edited last is never overwritten, and running it again updates only the pages the AI wrote. Every save keeps a version: History lists the last 25 and Restore writes an older one back as the newest. Read what it wrote before you mark anything OK to quote.

Check what a reply would read

Do not guess what the AI sees. The Try it view on the brain page shows what a reply would read for a sample message, and costs nothing; ayinbox brain preview does the same from the CLI. A reply reads the fields whole, then the campaign step the prospect answered, then pinned material and pages, then the rest ranked by how well it matches the message, up to a reply budget of 8,000 characters by default.

Every agent draft in the unibox shows a Used line listing the campaign, the brain and each source by title, and says when the brain changed after the draft was written. Flags for an unknown link, figure or phrase appear above the send button. They are a heuristic: they catch specific mistakes and miss others, and they never prove a reply is true.

What never goes in

Never put a password, API key or other credential in a brain. Its text goes to your AI provider with each draft. Text read from your material is marked as data in the prompt, so instructions inside it are not followed; that lowers the risk of a hostile web page or document steering a reply, but it does not remove it. A playbook that applies across offers, such as how you qualify a lead, belongs in AI skills instead. For how these facts become a reply, see AI replies to cold email.

Frequently asked questions

How do I train AI on my product for sales emails?

You do not retrain a model; you give it a written knowledge base it reads on every reply. In AY Inbox that is a campaign brain holding what you sell, who it is for, the price you will quote, the links you share, your internal rules, and material such as case studies and objections.

What should an AI sales assistant never be given?

Passwords, API keys and any other credential. Everything in an AY Inbox brain is sent to your AI provider with each draft, so treat it as text that leaves your workspace.

Can AI turn my sales material into a knowledge base?

Yes, if it only rearranges facts you gave it. In AY Inbox, Organize with AI reads a brain's offer and material and writes linked pages, such as objections linked to the proof that answers them, using only facts in your material. It writes up to 24 pages a run and never overwrites a page a person edited last.

How many offers can one workspace describe?

An AY Inbox workspace can hold up to 50 brains, one per offer, and mark one of them as the workspace default. Each campaign reads the brain assigned to it, or the default when none is assigned.