Document memory and AI assistant for a professional firm
A Conakry firm gathers its documents in a searchable memory, an AI assistant drafts replies, the partner approves from Telegram.
Project overview
This scenario illustrates what we build; it is not a client reference.
This scenario shows how a law or accounting firm in Conakry can find what it already knows. Its contracts, notes and past replies are gathered in a searchable memory. An AI assistant drafts replies to client questions. The partner approves from Telegram before anything is sent.
Client challenges & requirements
A firm with 2 partners and 6 staff in Conakry. Ten years of contracts, notes and letters are spread across Word files, mailboxes and WhatsApp conversations. When a client asks "what was the amount agreed in last year's amendment?", someone spends half a day searching.
If the staff member who knows the file is travelling, the client waits. Replies sometimes contradict each other from one person to the next, and nobody knows which version of a document is the right one.
Main objectives
Gather the firm's documents in a single memory that can be searched in plain language
Draft a reply for every client question, with the sources cited
Keep the partner in the loop: nothing leaves without their approval
Stay easy to use from a phone, even on a slow connection
Project solution
-
We start small
A shared folder where the firm drops its Word files, PDFs and mail exports. A Laravel application splits them, indexes them in PostgreSQL with pgvector and keeps track of every version. The AI assistant (Claude) answers staff questions and cites the document and passage it used. If a client question comes in by email or WhatsApp, the assistant drafts a reply and sends it to the partner on Telegram.
-
If the partner replies "ok", the message goes out. If they correct it, the correction is remembered for next time. If they do not reply, nothing goes out. The memory can be exported at any time and is hosted with a provider chosen with the client.
Indicators to track
- Time spent searching for a document
- Reply time to client questions
- Share of drafts approved without correction
- Number of documents indexed and versions found
Tech stack
- Application
- Laravel Livewire
- Mémoire documentaire
- PostgreSQL pgvector
- AI assistant
- Claude API Anthropic
- Canaux
- Telegram Bot API WhatsApp Business API IMAP
- Hosting
- Docker hébergeur choisi avec le client
Frequently asked questions
No. Every draft goes through Telegram and waits for approval. If nobody replies, the message stays pending. The firm keeps control of everything that goes out.
It stays in the firm's memory, hosted where the firm decides, and it is never used to train a model. Access rules (who sees which file) are defined with the firm and its counsel.
No. We start with the files that are already digital: Word, PDF, emails. Paper files are added later, at the firm's pace, with a simple photo from a phone.
A similar project?
Share your goals. After the scoping call, we propose a first scope and a timeline.
A question?
Call or write to us. We reply within one working day.
+224 627 293 377
hello@prisma-soft.comServices used
The Prisma Soft services this scenario draws on.
In the same industry
Client work and usage scenarios from the same industry.
