Yes. You build a small, structured library of your actual results, then wire it into your agent so it retrieves the right case before it writes a single sentence. The library is the easy part. The wiring is where most founders quit.
Why does your AI agent write proposals that prove nothing?
The short answer is that it has nothing to prove with. Every new chat starts from zero. You paste a job description, maybe a name, and the model reaches for the only thing available when it has no facts: adjectives. Proven track record. Trusted partner. Results-driven. It is not lying to you. It is filling a vacuum you left.
I have watched this play out across a lot of businesses now. The proposal reads fine. It sounds like a proposal. Then the client asks the one question that matters, which is who have you done this for, and you end up answering it yourself because the AI had no idea.
What's becoming clearer to me is that this is an AI agent context problem, not a prompt problem. People keep rewriting the prompt. The prompt was never the issue. This is the same wall most AI for founders and operators runs into in month two: the model is capable, the context window is empty.
What actually goes into a client case study library for AI?
Not a testimonials page. A testimonials page is marketing. A client case study library for AI is raw, structured evidence, and the structure matters more than the prose.
One entry per client. Every entry has the same seven fields:
- Client name and what they do
- The starting point, with a number attached to it
- What we actually built or did
- The result, with a number and a timeframe
- A quote in their words, with their name and role
- What did not work, or what we would change
- Industry and company size tags
That last pair of tags does more work than you think. When your agent writes to a 40-person agency, it needs to pull the case that looks like a 40-person agency. If it cannot filter, it picks at random and the proof lands flat.
Where to put it: a Notion database with those fields as properties, an Airtable base, or 20 markdown files in a Drive folder called /context/cases. I have set this up both ways and the plain folder is honestly faster to maintain. Start with 12 to 15 entries. You already have them. You just never wrote them down.
How do you teach AI your business so it retrieves the right proof at the right moment?
This is the part people skip. Dumping a PDF into a chat window gives the model facts but no instruction about when to use them.
Context engineering for founders comes down to three layers. Layer one is a single page on who you are, who you serve, what you sell, and what you refuse to do. Layer two is the case files. Layer three is three to five proposals or sales emails that actually closed. Not the prettiest ones. The ones that worked.
Then you wire it. Concretely: if you want to be running in an hour, upload the folder into a Claude Project or a custom GPT knowledge base. If you need this inside your own product, embed the files with OpenAI's embeddings endpoint, store the vectors in Supabase pgvector or Pinecone, and retrieve the top three matches on industry and company size before the model writes a word.
The instruction block matters as much as the files. Mine says something close to: before writing, select the single case study closest to the prospect's industry and size. Quote one specific number from it. Quote one line from the client in their own words. Do not invent numbers. If no case fits, say so and ask me.
That last line is the one that saves you. It is also the difference between an agent that is busy and an agent that is useful, which is the whole point when you build AI agents for business.
How do you test whether the AI proposal writing actually improved?
Pull five real proposals you sent in the last year that won. Run them through the new setup cold, as if the prospect just landed in your inbox.
Score each output on four things: does it contain at least one proper noun from your library, does it contain a number that is real and checkable, does the case match the prospect's size and industry, and would you send it without editing.
If an output has zero proper nouns, it failed. That is my blunt rule. Generic prose is what happens when the model has nothing named to hold onto.
Keep the scores in a sheet. Re-run it whenever you add cases or change the instruction. It takes me a minute, and it is the only way to know your AI proposal writing is getting better rather than just longer.
What is the mistake that quietly kills the whole system?
Stale libraries. You close a new client, you get a great result, and it never makes it into the folder. Six months later your agent is still citing the 2023 case and you cannot figure out why the proof feels old.
The fix is boring: a Friday habit where every new result becomes one markdown file. Or automate it. I have a small agent that watches my project management tool and drafts a case file whenever a project closes, then pings me to approve it. Approval takes ten seconds. Memory is permanent.
FAQ
Can I do this without a vector database?
Yes. A Claude Project or a custom GPT with 15 to 20 files uploaded gets you most of the way. Vector retrieval starts to matter once you cross roughly 50 documents, or when you need the whole thing running inside your own application.
How many case studies do I need before the output improves?
Five well-structured ones is a visible jump. Twelve to fifteen is where the matching starts to feel sharp, because you finally have enough spread across industries and company sizes for the agent to find a real comparison.
Will my AI leak client names to the wrong prospect?
Only if you let it. Add a confidentiality field to each case file, mark which ones are name-cleared and which are anonymized, and tell the agent to check that field before it quotes anything. Mine defaults to anonymized when it is unsure.
I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step. You can get it at a.mastermindshq.business/ai-os-book.