AI Document System for a Coaching Practice
Rescue, Full Build and a Pre-Delivery AI Quality Check
A half-built AI system that could not be trusted to deliver
A coaching and counselling practice sells personalized documents. A client pays, answers a detailed questionnaire, and should receive a document written specifically for them. A previous developer had started the system in Make but left it unfinished. Payments came in, but documents often failed to generate or never reached the client, answers were being stored in a form the AI could not read, some emails went out with blank names, and the AI step timed out on the longer documents. On top of that, the documents deal with personal topics, so the practice has strict rules about what each document may and may not say, and nothing was enforcing those rules.
One connected system, from payment to a checked, branded PDF
I took over the existing build, traced every failure to its root cause instead of patching symptoms, and then built the complete system around it. A Stripe purchase now triggers a Make scenario that routes the order to the right product and tier and sends a questionnaire linked to that exact order. When the answers arrive, Claude prepares a safe brief, writes the personalized document, and a second Claude step checks it against the practice's rules and corrects anything that breaks them. Docupilot turns the result into a branded PDF, Google Drive files it, and Resend delivers it after a timed hold. A separate scheduler handles reminders and fallbacks, so no order is ever left waiting. The practice does not touch anything unless it wants to.
Every layer of the system.
Ten connected components, from checkout to a delivered, quality-checked document.
One master Make scenario with its own route for every product and tier, 4 product lines and 12 versions in total. Checkout fields are read by their key, not their position, which fixed orders that were silently sent down the wrong path.
Every Typeform link carries hidden values that tie it to its order, so each set of answers attaches to the right purchase, even when one order includes more than one questionnaire.
Typeform answers are flattened into clean text before they are stored, so the AI always receives real answers instead of empty objects.
Claude, connected through the API inside Make, writes each document as structured JSON with more than 200 fields. Long documents were split into coordinated calls to stay inside Make's 300-second module limit.
When two related documents are created for one order, a shared data store keeps their themes, headings and structure identical while the personal sections stay unique.
Before anything is written, a Claude step prepares a brief for each reader that contains only what that reader is meant to see. The writer never receives what it must not say.
A second Claude step reads the finished document, catches any sentence that breaks the practice's rules, and rewrites only that sentence. It returns clean JSON every time, so the PDF build never breaks.
Docupilot templates produce the branded PDF, Google Drive stores it, the order record updates, and Resend delivers it after a timed hold. Products that include a session send the document and the booking link in one email. Kit tags each buyer by product.
A scenario runs every 30 minutes, sends reminders at 12 and 24 hours when a questionnaire is still missing, starts generation anyway when time runs out, and still delivers a late questionnaire's document when it arrives.
A separate paid study of Typeform's save-and-continue feature, tested by hand across seven real-world scenarios. It proved the feature safe for this system and removed the one real risk with dedicated forms.
The system in full detail.
Client details anonymized. Shown with permission.
What made this project different.
Three outcomes that changed how the practice delivers its documents.
What started as fixing someone else's broken scenario became three contracts: the rescue and full build, a save-and-resume study, and the AI quality-check layer.
A safe brief plus a two-stage AI check on the documents that need it. More than 100 test documents were generated and independently audited, and the client approved every document type before launch.
Every purchase becomes a personalized, branded PDF in the client's inbox, with reminders and fallbacks handled automatically and nobody in the loop.
The stack behind the system.
From a stalled build to a system the practice trusts.
The practice went from a half-finished automation that failed silently to a complete document system that runs on its own. Every purchase now becomes a personalized document, checked against the practice's own rules before anyone reads it, and delivered as a branded PDF without anyone touching it. The client reviewed and approved every document type before launch, and has hired FormLogic Studio again for each next stage of the work.
Have a similar project in mind?
Describe what you need and we will come back with a plan within 24 hours. No commitment required.
Or work with us directly on Upwork ↗
Let's build your AI document system.
Whether it's personalized reports, AI-written client documents, or a full intake-to-delivery pipeline, tell us what you need and we'll map out exactly how to make it work, with quality checks built in.
← Back to All Projects