AI drafts with the wrong numbers? Fill documents from your own data
§Work with me →
AI was supposed to draft the reports, quotes and replies your team fills from tickets, CRM records and spreadsheets. The drafts read well, but one has the wrong price, another has a different customer's date, and the layout changes every time. Now someone checks every line anyway. The reliable fix: your records supply the facts, and a program checks the model's.
Why AI drafts go wrong
- Facts: a language model retypes the numbers it's given and can change, round or invent them. A New York federal court fined two lawyers and their firm $5,000 for a brief citing court cases ChatGPT had made up (source).
- Format: free text changes every time, but a quote or report needs the same sections and tables, in PDF or Word.
- Integration: the data lives in your helpdesk, CRM or shared files, and copying it into a chat and back eats the time saved.
- Review and cost: the reviewer must see what came from where, and whole ticket histories sent to a large model get expensive.
What a checked draft looks like
The program copies every name, date, amount and line item into the template. The model writes the sentences, as structured output (a JSON schema it must follow), so a program can check its figures against the record. In a delivery reply:
| Line in the draft | Written by | Check |
|---|---|---|
Total: {{quote.total}}, filled as "Total: $4,180.00" |
The program | None: copied from the record |
| "Your 3 units ship on 14 November." | The model | Every number and date is in the record: passes |
| The same line with "12 November" | The model | The record says 14 November: flagged |
An empty required section is flagged too. The draft lands where your team works (a draft reply, a note, an attachment), and nothing goes out until a person approves it. For replies sent with no reviewer, see stop your chatbot from making things up.
Ways to fix it
| What it does | Good for | Limits | |
|---|---|---|---|
| Templates and mail merge | Word or Google Docs fills fields from a spreadsheet | Documents where only names, dates and amounts change | Can't summarize a ticket or tailor a paragraph |
| Your helpdesk's AI add-on | E.g. Zendesk's Copilot add-on suggests replies from your help articles and procedures; agents approve them (source) | Ticket replies inside that tool | Paid extra; check it can reach your other systems and templates |
| A custom integration | The setup above: facts from your systems, wording from the model, every figure checked | Recurring documents whose figures must be right | Built for your templates; a new one takes work |
What I've worked on
I've built an AI tool inside a helpdesk that drafts PDF documents, other LLM features that return structured data instead of free text (see the LLM case), and a tool that builds reports from Excel, Word, CSV and text files (see the report case).
What I need from you
Two or three real examples (personal data removed, if you prefer), the template, where the data lives, which fields must be exact, who approves the drafts, and how many your team writes a week.
How to start
Email me at cdesivo92@gmail.com with a few lines about the documents your team writes, or hire me on Freelancer.com if you prefer the platform's contracts and payment protection. I'll tell you honestly whether I'm the right person for it, and how I'd approach it.