DiscoveryVIP / Apps Script AcademyHelp

85. Reliable AI automation

Module 17 · 20 minutes · Lesson 84; earlier service and JavaScript fundamentals

Open Interactive Lesson

Discover

How can you use reliable ai automation without changing the wrong data?

Predict

Which values or resources will change, and which should stay unchanged? Write your prediction before execution.

Understand

Human review should precede applying a model-proposed change. Validate structure and content, retain the source, and record approval separately. A confidence number from a model is not calibrated proof. This example applies only explicitly reviewed proposals and logs those held back, illustrating a gate independent of model generation.

Example 1: Working automation

The normal path makes no write. Changing the review gate visibly changes the side effect to investigate.

function main() {
  const proposal = {category: "support", reviewed: false};
  if (!proposal.reviewed) { console.log("Held for human review"); return; }
  SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
}
Important line explanations
Line 1: function main() {
Define a named block of instructions. Run Script calls main; helpers run only when called.
Line 2: const proposal = {category: "support", reviewed: false};
Initialize a local binding. const prevents reassignment of that binding, but does not freeze a referenced object.
Line 3: if (!proposal.reviewed) { console.log("Held for human review"); return; }
Record diagnostic evidence. A log does not grant permissions or prove a later side effect succeeded.
Line 4: SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
Resolve the current workbook. A real standalone project needs openById with a configured ID instead.
Line 5: }
Close the current block. Indentation makes scope and the function boundary visible.
Exact expected output and resource
{
  "logs": [
    "Held for human review"
  ],
  "error": null,
  "resource": "logs",
  "value": null
}

Example 2: Investigate a reproducible failure

Predict the effect of reviewed: true and identify the first mismatch. This example is intentionally faulty.

function main() {
  const proposal = {category: "support", reviewed: true};
  if (!proposal.reviewed) { console.log("Held for human review"); return; }
  SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
}
Important line explanations
Line 1: function main() {
Define a named block of instructions. Run Script calls main; helpers run only when called.
Line 2: const proposal = {category: "support", reviewed: true};
Initialize a local binding. const prevents reassignment of that binding, but does not freeze a referenced object.
Line 3: if (!proposal.reviewed) { console.log("Held for human review"); return; }
Record diagnostic evidence. A log does not grant permissions or prove a later side effect succeeded.
Line 4: SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
Resolve the current workbook. A real standalone project needs openById with a configured ID instead.
Line 5: }
Close the current block. Indentation makes scope and the function boundary visible.
Exact expected output and resource
{
  "logs": [],
  "error": null,
  "resource": "logs",
  "value": null
}

Example 3: Separate setup from execution and inspect repeat behavior

Call the same named operation twice. Compare outputs and resource counts to learn whether this operation is idempotent. Repeated creation is deliberately visible; real exports require a rerun policy.

function lessonAction() {
  const proposal = {category: "support", reviewed: false};
  if (!proposal.reviewed) { console.log("Held for human review"); return; }
  SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
}
function main() {
  lessonAction();
  lessonAction();
}
Important line explanations
Line 1: function lessonAction() {
Define a named block of instructions. Run Script calls main; helpers run only when called.
Line 2: const proposal = {category: "support", reviewed: false};
Initialize a local binding. const prevents reassignment of that binding, but does not freeze a referenced object.
Line 3: if (!proposal.reviewed) { console.log("Held for human review"); return; }
Record diagnostic evidence. A log does not grant permissions or prove a later side effect succeeded.
Line 4: SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
Resolve the current workbook. A real standalone project needs openById with a configured ID instead.
Line 5: }
Close the current block. Indentation makes scope and the function boundary visible.
Line 6: function main() {
Define a named block of instructions. Run Script calls main; helpers run only when called.
Line 7: lessonAction();
Follow this operation using the surrounding named values. Compare the actual recorded state with the lesson’s specific explanation below.
Line 8: lessonAction();
Follow this operation using the surrounding named values. Compare the actual recorded state with the lesson’s specific explanation below.
Line 9: }
Close the current block. Indentation makes scope and the function boundary visible.
Exact expected output and resource
{
  "logs": [
    "Held for human review",
    "Held for human review"
  ],
  "error": null,
  "resource": "logs",
  "value": null
}

Example 4: Add a diagnostic boundary around the operation

Observe start, successful completion or a rethrown failure. This optional professional pattern adds error context without swallowing the error or rolling back changes.

function lessonAction() {
  const proposal = {category: "support", reviewed: false};
  if (!proposal.reviewed) { console.log("Held for human review"); return; }
  SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
}
function main() {
  console.log("Start: Reliable AI automation");
  try { lessonAction(); console.log("Completed operation"); }
  catch(error) { console.error("Failed operation: " + error.message); throw error; }
}
Important line explanations
Line 1: function lessonAction() {
Define a named block of instructions. Run Script calls main; helpers run only when called.
Line 2: const proposal = {category: "support", reviewed: false};
Initialize a local binding. const prevents reassignment of that binding, but does not freeze a referenced object.
Line 3: if (!proposal.reviewed) { console.log("Held for human review"); return; }
Record diagnostic evidence. A log does not grant permissions or prove a later side effect succeeded.
Line 4: SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
Resolve the current workbook. A real standalone project needs openById with a configured ID instead.
Line 5: }
Close the current block. Indentation makes scope and the function boundary visible.
Line 6: function main() {
Define a named block of instructions. Run Script calls main; helpers run only when called.
Line 7: console.log("Start: Reliable AI automation");
Record diagnostic evidence. A log does not grant permissions or prove a later side effect succeeded.
Line 8: try { lessonAction(); console.log("Completed operation"); }
Record diagnostic evidence. A log does not grant permissions or prove a later side effect succeeded.
Line 9: catch(error) { console.error("Failed operation: " + error.message); throw error; }
Record diagnostic evidence. A log does not grant permissions or prove a later side effect succeeded.
Line 10: }
Close the current block. Indentation makes scope and the function boundary visible.
Exact expected output and resource
{
  "logs": [
    "Start: Reliable AI automation",
    "Held for human review",
    "Completed operation"
  ],
  "error": null,
  "resource": "logs",
  "value": null
}

Build and test

Repair the supplied script so that: The normal path makes no write. Changing the review gate visibly changes the side effect to investigate. Preserve other source data. Checks compare observable output and the named resource, not exact source text.

function main() {
  const proposal = {category: "support", reviewed: true};
  if (!proposal.reviewed) { console.log("Held for human review"); return; }
  SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
}

Graduated hints

Hint 1

Inspect the operation containing reviewed: true. Compare it with the stated rule and the first mismatch.

Hint 2

The repair concerns reviewed: false. Change only the relevant argument, operator or method; rerun with fresh input.

Hint 3

The reference uses reviewed: false where the broken source uses reviewed: true. Explain why before applying it.

Reference solution

function main() {
  const proposal = {category: "support", reviewed: false};
  if (!proposal.reviewed) { console.log("Held for human review"); return; }
  SpreadsheetApp.getActiveSpreadsheet().getActiveSheet().getRange("B2").setValue(proposal.category);
}

Explain and reflect

Which operation proves the intended result in reliable ai automation? Which wrong input would expose a weakness, and what would remain unchanged?

Adapt the calculation or validation to a second realistic case. Write its expected result before running. Demonstrate one invalid case and explain whether the side effects should be rolled back or retried.

Troubleshooting

If the console is empty, inspect the logs preview: this script may work through state changes. If the function is missing, select main. For a service error inspect the first failed timeline entry and its arguments. Use Manual unless investigating a trigger. Changing input data also changes the expected result; restore sample data before the authored checks.

Knowledge check

Explain the key method or operator. Give an exact output for a second input, plus one invalid case. Verify the result in the interactive lesson; written explanations need self or educator review.

Optional AI practice

Explain this concept using the source above. Help implement one small extension and define its tests. Review/debug the source using its actual expected output; cite real method references and never claim execution.

Transfer to Google and authorization guide

Next lesson: Syntax, runtime and logic errors