The enterprise conversation about AI has matured past the initial fascination with chatbots — the manual, high-friction loop of prompting a model and pasting the result somewhere useful. Real return doesn’t come from that friction. It comes from asynchronous, event-driven workflows that run behind the scenes, cutting the context-switching tax that defines most professional work. This post walks through one: an AI Travel Assistant that turns Gmail confirmations into a structured Google Sheets itinerary, built natively in Google Workspace Studio.
I’ve recorded the full build end to end. The video is the hands-on companion to everything below — worth watching alongside if you plan to build this yourself.
The problem it solves is a familiar piece of technical debt: travel data is fragmented across a dozen confirmation emails. Turning that unstructured Gmail data into a structured itinerary moves you from manual retrieval to a centralised, machine-readable data layer. That shift — from conversation to action — is the whole point.
Chatbots vs. true task automation
Building automation inside the Google ecosystem has a specific advantage over third-party platforms like Zapier or Make. Those tools are versatile, but every extra connector is another hop: more latency, more surface area, and data leaving your perimeter. Native flows keep IAM parity — the data stays inside the boundary you already govern, and the stack stays shorter.
| Feature | Manual workflow | Third-party automation | Native AI automation |
|---|---|---|---|
| Data entry | Manual copy-paste | Automated via API connectors | Native background process |
| Integration | None — high context-switching | External API dependencies | Direct (Gmail / Sheets / AI) |
| Intelligence | Human cognitive effort | Basic rules-based logic | AI-driven intent & extraction |
| Security | High human error risk | Data egress to third party | Ecosystem native (IAM parity) |
The native advantage is that AI becomes the connective tissue between your communication channels and your data stores without exposing anything sensitive to external middleware.
The automation blueprint
A blueprint isn’t ceremony here — it’s how you get predictable outcomes and keep token costs sane. Before touching the builder, deconstruct the logic into four asynchronous stages:
- The trigger. A new incoming Gmail message.
- The decision (gatekeeper). An AI classification step. Gmail has no native “travel” filter that catches every nuance, so an LLM evaluates the email body against specific categories — flight, hotel, car rental, train — and returns a boolean.
- JSON extraction. If the decision is true, a specialised agent converts natural language into machine-readable JSON.
- The action. Append the structured record to the target Google Sheet.
The decide step is what makes this efficient. A lightweight logic gate filters out irrelevant mail before the heavier extraction model ever runs, which optimises both latency and cost.
Phase 1: preparing the data foundation
An automation is only as robust as its destination. Configure the structured destination before deploying any AI logic, so field mapping is a lookup rather than a guess.
Create a Google Sheet titled AI Travel Itinerary and define the schema explicitly:
- Date
- Subject
- Trip Type
- Departure
- Arrival
- Price
- Raw Snippet — used for audit logging
Then source diverse test data: real confirmation emails, or AI-generated templates that mimic the awkward cases. With the foundation laid, the logic can be manifested inside Workspace Studio.
Phase 2: building the workflow
Google Workspace Studio positions Gemini as a co-developer, which meaningfully accelerates writing JSON schemas and logic frameworks.
- Trigger configuration. Monitor the Gmail inbox for new messages.
- The decide step. Prompt the AI to determine whether the email is a travel confirmation. It must return a binary true/false that governs the downstream branch.
- Extraction logic. Define the agent persona as a Travel Itinerary Extraction Assistant. Engineer the system prompt to force output into a specific JSON shape and strictly forbid conversational filler. Hallucinated prose is what breaks the Sheets integration, not bad data.
- Field mapping. Connect the JSON output keys to spreadsheet columns.
Phase 3: validation and live execution
Reliability comes from testing against edge cases, not happy paths. The Skyline Airways example in the video is a good one: a single email containing both onward and return legs. A robust workflow has to be tuned to decide whether that produces multiple rows or a nested JSON object.
- Thread selection. Pick a complex multi-leg email.
- Status monitoring. Verify that the Step 2: True condition fires and extraction executes.
- Audit logging. The Raw Snippet column is your observability tool — it lets you reconcile the original email text against the AI’s structured output and catch drift or missed fields.
Once accuracy is consistent, move the workflow from Test to Live.
Applying the pattern elsewhere
Trigger → extract → log is a modular template. It fits any operation that runs on unstructured communication:
- Invoice trackers. Vendor, due date and amount pulled from receipts into a financial report.
- Meeting summary logs. Calendar notifications and follow-ups turned into structured action items.
- Hiring trackers. Candidate details extracted from CV-heavy email threads into a recruitment pipeline.
Governance and optimisation
Long-term automation needs governance and observability, not just a working first run:
- Prompt consistency. Avoid vague instructions. Precise constraints stop the model reverting to conversational habits.
- Model drift. Both models and email formats evolve. Audit rows periodically to confirm the logic still matches current email structures.
- Data governance. In Google Workspace Studio, API-based calls typically don’t use customer data for model training — a meaningful reassurance for corporate privacy standards.
Your first AI transformation
Moving from manual copy-pasting to autonomous, event-driven systems is the practical marker of a modern engineering practice. This Travel Assistant framework is small, but it establishes the foundation: a trigger, a cheap gate, a hardened extractor, an auditable sink.
Pick one repetitive, email-based task today and apply the blueprint. The future of productivity isn’t in talking to AI — it’s in building AI that acts.
If you found this useful, the full walkthrough is on YouTube — and subscribing to CodeRunShip helps more of these get made.
Working on something like this? Let's trade notes.
I'm always up for a good conversation about enterprise architecture, agentic AI and the AI-native SDLC. Subscribe for field notes, or write if a specific problem is on your mind.