Connect AI to Your Tools

How to Use This Template

Answer the questions in each section. Replace every [[Token]] with your real value. Where you see a table, add or remove rows to match your actual app stack. The result is a working integration plan you can hand to a developer, run through Zapier/Make yourself, or use to brief an AI agent.

Section 1: App Inventory

List every app involved in the workflow you want to automate. Be specific: "email" is not enough; name the provider.

Primary goal for this integration: [[One sentence: e.g., "Automatically qualify new leads in HubSpot and send a personalized first-touch email within 5 minutes of form submission."]]

Current pain point: [[What breaks, takes too long, or gets missed without automation?]]

Section 2: Integration Approach Options

Choose the method that matches your technical comfort. Pick one approach per workflow: start with the simplest that meets your goal.

Your technical comfort level: [[Non-technical / Somewhat technical / Developer]]

Recommended approach for your comfort level:

• Non-technical → Start with native integrations for any two apps that already support it, then add a no-code tool (Zapier or Make) for anything that spans more than two apps. Avoid API and agent frameworks until you have a no-code automation running reliably.

• Somewhat technical → Native integrations first, then use Make (more powerful than Zapier for branching logic) or an AI-native tool with a system prompt. A simple webhook can unlock a lot without full API development.

• Developer → MCP / API or an agent framework gives you the most control. Start with the OpenAI or Anthropic API for a single-step augmentation before building a multi-agent pipeline.

Selected approach for [[Workflow Name]]: [[Approach from the table above]]

Section 3: Trigger → Action Workflow Maps

Map every automation as a trigger (the event that starts it) → action chain (what happens next). One row per step.

Workflow 1: [[Workflow Name]] (start here: your highest-value automation)

Workflow 1 trigger condition: [[Exact condition: e.g., "Form submitted on [[Landing Page Name]] between 6am–10pm local time"]]

Workflow 1 expected outcome: [[What should be true 5 minutes after the trigger fires?]]

Workflow 2: [[Second Workflow Name]] (add after Workflow 1 is stable)

Workflow 3: [[Third Workflow Name]] (Phase 2 or later)

Section 4: Data Flow and Where AI Fits

Describe how data moves between apps and at which point AI does work. Use the format below.

Data origin: [[Where the data starts: e.g., "Customer fills out a Typeform"]]

Data path:

1. [[App 1]] captures [[data type, e.g., contact details + inquiry text]]

1. [[Automation tool, e.g., Zapier]] triggers on [[event]] and passes [[specific fields]] to AI

1. AI receives [[inputs: e.g., inquiry text, CRM history]] and produces [[output: e.g., a classification tag and a draft reply]]

1. [[App 2]] receives AI output and [[what it does with it: e.g., updates the lead record and saves the draft]]

1. [[App 3 if applicable]] is notified [[how and with what info]]

Data that AI sees (and must be allowed to see): [[List the fields: e.g., "first name, company, inquiry text, past purchase history"]]

Data AI must never see: [[e.g., "payment card numbers, full SSN, medical records"]]

Personally identifiable information (PII) in this flow: [[Yes / No]]: if yes: [[name the fields and the rule for handling them, e.g., "Email address passes through but is not stored by the AI provider per their DPA"]]

Section 5: Auth, Credentials, and Security

Credential map

Security rules

• Least-privilege access: Each API key or OAuth connection should request only the permissions the workflow actually needs. Do not use admin tokens for read-only steps.

• Key rotation: Rotate API keys every [[90 / 180]] days or immediately after any personnel change. Record rotation dates in [[your password manager or internal doc]].

• No keys in code or prompts: Never paste an API key into a prompt, a Zap description, or a shared document. Store in the automation platform's built-in credential vault or a secrets manager.

• Scope review: Before connecting any AI model to a data source, confirm the AI provider's data processing agreement (DPA) covers your data type. OpenAI, Anthropic, and Google Gemini all have enterprise DPAs available: check [[your provider's trust page]] before processing customer PII.

• Audit log: [[Automation tool]] should log every run. Enable run history and keep logs for [[30 / 90]] days minimum.

Section 6: Cost Considerations

Cost-control levers:

1. Use the cheapest AI model that passes your quality bar: run a 20-call spot check on a smaller model first.

1. Cache or skip AI for records that don't need it (e.g., internal notes, test submissions).

1. Set a hard spend cap in your AI provider dashboard: [[OpenAI / Anthropic / Google Cloud]] all support monthly spend limits.

1. Measure cost-per-outcome, not cost-per-call, to validate ROI.

Section 7: Error Handling and Human Review

Every automated workflow fails eventually. Define how failures are caught and what happens next.

Error types and responses

Human review queue: [[Where reviewed items go: e.g., a Notion database, a Slack channel named #ai-review, a shared inbox]]

Review SLA: AI-generated output that touches a customer must be reviewed within [[Review Window, e.g., 24 hours]]]] before it is sent or published.

Auto-send threshold: Only auto-send without review if confidence is [[e.g., above 90%: or define your own signal]].

Section 8: Phased Rollout

Start with one workflow. Automate, stabilize, then expand.

Phase 1: [[Phase 1 Name, e.g., "Single High-Value Workflow"]] (Weeks [[1–2]])

Scope: Workflow 1 only: [[Workflow 1 Name]]

Success metric: [[e.g., "Workflow fires on 100% of qualifying triggers with <2% error rate for 5 consecutive business days"]]

Completion gate before Phase 2: Error rate below [[2%]], zero PII incidents, human review SLA met.

Phase 2: [[Phase 2 Name, e.g., "Expand to Second Workflow"]] (Weeks [[3–4]])

Scope: Add Workflow 2: [[Workflow 2 Name]]

Prerequisite: Phase 1 gate cleared.

Success metric: [[e.g., "Both workflows running; total AI spend within budget"]]

Phase 3: [[Phase 3 Name, e.g., "Full Stack + Monitoring"]] (Week [[5+]])

Scope: Add Workflow 3, set up monitoring dashboard, document runbook.

Prerequisite: Phase 2 gate cleared.

Success metric: [[e.g., "All three workflows running; on-call runbook written; monthly cost review scheduled"]]

Section 9: Testing Plan

Before any workflow goes live, run these checks.

Pre-launch checklist

1. Dry run with test data: fire the trigger using a clearly labelled test record (e.g., name = "TEST USER DO NOT CONTACT"). Confirm each step passes data correctly.

1. AI output spot-check: review 10 AI outputs manually before enabling auto-send. Check for: accuracy, tone, hallucinations, PII leakage.

1. Error path test: deliberately break step 3 (e.g., send a blank input to the AI). Confirm the error handler fires and the alert reaches [[the designated person or channel]].

1. Duplicate test: fire the same trigger twice in 30 seconds. Confirm deduplication works and only one output is created.

1. Auth expiry simulation: revoke and re-grant the API key. Confirm the workflow fails gracefully and alerts instead of silently dropping records.

1. Cost check: run 50 test records and calculate actual cost-per-run. Compare against the estimate in Section 6.

1. Human review queue test: confirm at least one record routes to the review queue and the reviewer can approve/reject it.

Sign-off

Section 10: Maintenance

AI integrations drift. Models get updated, APIs change, app connections expire.

Monthly tasks

• Review automation run logs: check error rate trend; anything above [[3%]] needs investigation.
• Review AI spend: compare against the Section 6 budget; adjust model tier or volume if needed.
• Spot-check 5 recent AI outputs for quality: note any degradation vs. the pre-launch baseline.

Quarterly tasks

• Rotate all API keys (see Section 5 security rules).
• Re-run the pre-launch testing checklist on each live workflow.
• Review AI provider changelog: new models or deprecated endpoints can break integrations silently.
• Update prompts if AI output quality has drifted: model updates can change behavior.

Annual tasks

• Full audit: re-evaluate the integration approach (Section 2): better no-code tooling may have launched; costs may have dropped.
• Re-confirm DPA compliance for any PII in the data flow.

On-call runbook

Who owns this integration: [[Owner Name and role]]

Where run logs live: [[URL or path]]

How to pause the workflow in an emergency: [[Steps: e.g., "Log in to Zapier → Zaps → Toggle off [[Workflow Name]]"]]

Who to contact for each app: [[App 1]]: [[Contact or support URL]] · [[App 2]]: [[Contact or support URL]] · [[AI provider]]: [[Support URL]]

Worked Example

A concrete fill-in showing what a completed plan looks like for a common scenario.

Scenario: A small sales team wants to automatically qualify new inbound leads from their website form and send a personalized first-touch email: without the rep having to do it manually.

App inventory:

Technical comfort: Somewhat technical: will use Make (no custom code).

Workflow trigger → action:

Phase 1 goal: Run this single workflow for 2 weeks. Target: <2% error rate, all qualified leads touched within 5 minutes.

Cost estimate (100 leads/month): Make Starter ($9/mo) + OpenAI API (~$0.40/mo at GPT-5 Instant pricing) = ~$9.40/month total.

This template is a planning guide, not a guarantee of specific tool behavior. Verify API terms, pricing, and data processing agreements directly with each vendor before processing customer data.