The AI Agent Stack Decision Cheat Sheet

Use this when the task is always the same, has a clear trigger, and needs no judgement. If you can write the rule in one sentence ('when X happens, do Y'), this is your tier. Examples: send a Slack message when a new lead comes in, copy a form submission to a spreadsheet, tag a HubSpot contact when they visit a pricing page. Cost to build: hours. Cost to maintain: near zero.

Use this when the task requires language, judgement, or generation but you are always the one starting it. You write the prompt, you paste the input, you review the output. Examples: drafting a follow-up email from call notes, summarising a report, rewriting a landing page section. This is a force multiplier for your time, not an autonomous system. Cost to build: minutes. Requires your attention every time.

Use this when the task spans multiple steps, requires decisions mid-way, reads live data or external tools, and should run without you watching. Examples: an agent that monitors your inbox, qualifies new leads against your criteria, drafts a personalised reply, and flags it for your review. Cost to build: days to weeks. Requires ongoing oversight and error handling. Do not build Tier 3 when Tier 1 will do.

If the task is fully predictable ('when X, do Y, every time, no exceptions'), you need Tier 1. Build a Zap or a Make scenario. Do not involve AI at all. Adding AI to a deterministic task adds cost, latency, and failure points for zero benefit.

If yes, you need Tier 2. Build a Custom GPT, a Claude Project, or a saved prompt with a clear template. This is the most underused tier. Most businesses jump straight to Tier 3 when a well-crafted prompt and 10 minutes of their time would do the job faster and more reliably.

If yes, you may need Tier 3. But first, ask yourself: can you break this into a Tier 1 trigger that feeds a Tier 2 prompt that returns output for your review? That hybrid approach is faster to build, easier to debug, and fails more gracefully than a fully autonomous agent.

If yes, do not go fully autonomous yet regardless of which tier fits. Add a human review step. Autonomous agents make mistakes. The question is not whether they will fail but when. Build the human checkpoint in from day one and remove it later once you have confidence in the output.

This is Tier 1. Your email platform already does this. Mailchimp, Kit, HubSpot, ActiveCampaign, all of them have native automations for welcome sequences. An AI agent here adds cost and complexity for a task that does not require intelligence.

Zapier can trigger the task. It cannot write the proposal. This is a Tier 2 task at minimum. Use a Zapier trigger to detect a new qualified lead, then pass the data to a Claude or GPT prompt that drafts the personalised proposal. Review before you send.

If you are getting fewer than 50 new leads a week, Tier 2 is almost certainly enough. Manual review of a GPT-drafted qualification summary takes two minutes per lead. A Tier 3 agent costs weeks to build and months to tune. Do the maths before you commit.

This is the architecture that works for 80 percent of business tasks. A Zapier trigger fires when something happens. It passes structured data to a Claude or GPT prompt via an API call or a tool like Make. The prompt produces a draft. A human reviews and approves before anything goes out. You get speed and accuracy without the failure risk of full autonomy.

Upgrade to a true autonomous agent only when: the task runs more than 20 times per day, the human review step is taking more time than the task saves, and you have at least 4 weeks of output you can evaluate for quality. If all three are true, build the agent. If not, stay hybrid.

Zapier, Make (formerly Integromat), n8n (self-hosted), native automations in HubSpot, Mailchimp, Kit, Notion, Airtable. Use the native tool first. Only build a Zap when your native tool cannot do it.

ChatGPT Custom GPTs, Claude Projects, saved system prompts in any AI interface, Notion AI for document tasks, Perplexity for research tasks. The key is a well-written system prompt that defines the task, the output format, and the constraints. A bad prompt in a great tool gives worse results than a great prompt in a basic tool.

AutoGen, CrewAI, LangChain (for developers), Relevance AI (no-code), Lindy, and custom builds via Claude or OpenAI APIs with tool-use enabled. These require technical setup, error handling, and ongoing monitoring. Budget at least 20 hours to build and test a simple agent well.

1. Does this task already exist in a tool I pay for? Check before building. 2. How often does this task happen per week? Under 10 times, consider whether it needs automating at all. 3. What does failure look like, and is that acceptable? 4. Am I solving a real bottleneck or a hypothetical one? 5. Can I describe the output in one clear sentence? If you cannot, the task is not ready to automate. Define it first.

If the task takes you 15 minutes and happens twice a week, that is 26 hours a year. An automation that takes 8 hours to build and maintain pays back in about 4 months. An automation that takes 40 hours to build breaks even in almost 2 years. Calculate break-even before you start, especially for Tier 3 builds.

Lead research pipelines (find prospect, enrich data, score, draft outreach), content repurposing at scale (take one blog post, produce 5 social formats, schedule), customer support triage (classify ticket, pull relevant knowledge base entry, draft reply for human review). The common thread: high volume, structured inputs, predictable output format.

Agents that watch for something and alert you when it happens. Competitor price changes, brand mentions, SEO ranking drops, email replies that match certain criteria. These run quietly in the background and only surface when action is needed. Low risk, high value.

Responding to a new inbound lead within 5 minutes increases conversion rates measurably. If you cannot respond that fast manually and you are getting consistent lead volume, an agent that drafts the first response for human approval is a legitimate Tier 3 use case. The human still sends it. The agent just makes sure it is ready before you have had your coffee.