For several months, reconciling our Evaneos commissions relied on a chain I had built piece by piece: 3 Zapier workflows, PDF.co for extraction, ChatGPT to structure the data, a custom Google Apps Script, Airtable as an intermediate database, an OneDrive Excel file for enrichment, and finally a PDF generated and automatically deposited back on OneDrive.
It worked. But it was opaque, expensive, and any single link could break independently of the others. When Zapier updated a connector or PDF.co had an outage, the workflow would fail silently.
The trigger: a working session with Claude
This morning, I described the problem to Claude (Anthropic's AI). By sharing screenshots of the Zaps and the relevant PDFs, we analysed the workflow together in a few minutes. Claude immediately understood the business context — not just the code, but the operational logic behind it.
Then we coded. Together.
What we built
A single /ventilation page in our Nortrip Toolbox back-office (Next.js 16 + Supabase + Vercel):
- Drop the Evaneos PDF invoice — Claude extracts file IDs and amounts automatically via the Anthropic API (Haiku model for speed), summing multiple fee lines per file
- The app connects to Microsoft OneDrive via MSAL and Microsoft Graph API — read-only token only (
Files.Readscope, so Microsoft itself blocks any write attempt at the token level) - The sales recap file loads automatically from OneDrive — the central company reference file, which changes constantly
- Each line is automatically enriched: Toogo code, client name, destination, display order
- Sorted table + one-click PDF export — same format as the output produced by the 3 Zaps
The result
Zero Zapier. Zero Airtable for this workflow. Zero Google Apps Script. Zero PDF.co. All of it built in one working session.
What struck me: Claude didn't just write code. It understood the business context, anticipated edge cases (multiple fee lines per file to sum, thousands separators breaking PDF rendering), and raised the right security questions before I did.
When I mentioned the Excel file was the company's central reference document, Claude immediately proposed the read-only firewall — restricting MSAL scopes to Files.Read so Microsoft blocks any write at the token level. That wasn't in my original specs.
What this says about AI in 2026
The AI-augmented developer isn't a myth. It's also not "AI replaces the developer". It's a collaboration: you bring the business context and the decisions, the AI brings execution speed and exhaustive technical memory.
In one morning, we replaced a chain of SaaS tools that cost money in subscriptions, was fragile, and too complex to evolve quickly.
And it runs on Vercel.
Benjamin Bertigny — Founder NBN IA. I build AI tools for travel agencies and SMBs.