The simplest form of software automation is a rule that fires on a trigger: an email filter that labels every message from a supplier, a cron job that runs a backup at 00:05, a recorded macro that formats a report. One condition, one action, no judgment.
From there the types of automation climb through scheduled scripts, workflow tools that connect apps, screen bots, multi-step process automation, AI steps that read and classify, and finally agents that plan their own steps. Each level handles more variety and costs more to supervise. Here’s the ladder with real examples, and a short guide to which rung you need.
Level 1: Rules inside the apps you already use
The entry point costs nothing. Gmail filters take a search (sender, subject, words) and apply an action to every matching message: send it to a label, archive, star, delete or forward it. Outlook rules, Slack reminders and Shopify’s order tagging work the same way.
Example: every email from your 3PL with “exception” in the subject gets labeled “Shipping issues” and starred.
Handles: one app, fixed inputs. Breaks when: the sender changes their subject line format.
Level 2: Macros and scheduled scripts
One step up is recording or writing a sequence and replaying it.
An Excel macro is “an action or a set of actions that you can run as many times as you want.” Microsoft’s own example: a monthly accounting report where the names of customers with overdue accounts need to be red and bold. Record it once, run it on any selection.
A cron job runs a command on a schedule you write in 5 fields (minute, hour, day of month, month, day of week). The man page’s example runs a daily job at 5 minutes past midnight. It has run server housekeeping for decades and still does.
Example: a script exports yesterday’s orders to CSV at 06:00 and drops it in a shared folder.
Handles: a fixed sequence on a fixed schedule. Breaks when: the spreadsheet layout or the file path changes.
Level 3: Workflow automation across apps
Now 2 or more tools talk to each other without code.
Zapier builds each workflow from a trigger, “the event that starts your Zap,” and actions, “what the Zap does after it triggers.” The free plan (October 2026) covers 100 tasks a month with 2-step Zaps. Microsoft’s Power Automate does the same for Microsoft 365 with cloud flows that run automatically on an event, instantly on a button press, or on a schedule.
Example: new paid Shopify order, add a row to a Google Sheet and post a line in a Slack channel.
Handles: structured data moving between modern SaaS tools. Breaks when: an app renames a field or an API key expires. Set an alert for days with zero runs.
Level 4: Robotic process automation (RPA)
RPA is a workflow that drives a screen: it clicks and types where a person would. UiPath describes it as software robots automating “repetitive, rule-based tasks like data entry and system integration.” IBM’s one-line distinction: “RPA is process-driven, whereas AI is data-driven.” The robot replays the clicks a person recorded.
Power Automate’s desktop flows are the budget version; UiPath is the enterprise one. RPA exists because banks, insurers and hospitals run systems with no API.
Example: a bot logs into an old accounting system every morning and keys in yesterday’s invoice totals from a spreadsheet.
Handles: legacy systems with screens. Breaks when: someone moves a button. For a company on modern SaaS, level 3 does the same job with far less breakage.
Level 5: Business process automation with branches and approvals
Levels 1 to 4 run a single path. Process automation adds branches, waits, conditions and human sign-offs, usually inside a system that knows about your customers.
HubSpot workflows enroll contacts by filter, event, schedule or webhook, then send email, assign records, create tasks and notify people, on Professional and Enterprise plans. Klaviyo flows are “a sequence of automated actions (e.g., sending messages) triggered by a behavior or event,” with triggers for list joins, segments, a placed order, a date, a price drop or low inventory.
Example: an abandoned cart flow that waits 4 hours, emails, waits a day, emails again with a code if the cart’s over $80, and stops the moment the customer buys.
Handles: multi-step journeys with timing and conditions. Breaks when: the flow has no owner and it keeps emailing a segment that stopped making sense.
Level 6: Automation for the software itself
Software teams automate the building and checking of their own product. GitHub Actions runs workflows on events like a code push or a pull request, on fresh virtual machines, to build, test and deploy. Playwright is “an end-to-end test framework for modern web apps” that drives Chromium, WebKit and Firefox so a test can click through your checkout on every release.
Example: every pull request triggers a test run that adds a product to the cart and completes checkout in 3 browsers.
Handles: anything a developer can script. Breaks when: the page changes and the test wasn’t updated. This level is for companies that ship software; a store on Shopify can skip it.
Level 7: AI steps inside a workflow
Everything above runs on rules, which means it fails on input outside its rules. An AI step fixes that for text, images and documents: a language model classifies, extracts, summarizes or drafts, and rules take over again on either side.
Anthropic’s engineering guide calls these workflows, “systems where LLMs and tools are orchestrated through predefined code paths,” and names the patterns: prompt chaining, routing (classify, then send to a specialized step), parallelization and evaluator loops. Its advice is to start here and only add complexity when it measurably improves outcomes.
Example: every incoming support email is classified into “order status”, “return”, “wholesale” or “other”, routed by a rule, and a draft reply is generated from the order record for a person to approve.
Handles: variable input. Breaks when: facts aren’t in the prompt. A September 2025 paper from OpenAI researchers, Why Language Models Hallucinate, argues training rewards guessing over admitting uncertainty. Ground every step in your data and check numbers with code.
Level 8: AI agents
At the top, the model decides the steps. Anthropic’s definition of agents: systems where “LLMs dynamically direct their own processes and tool usage.” The agent reads a goal, plans, calls tools through connectors (the Model Context Protocol is the open standard for that), checks its work and comes back with a result.
Products at this level in 2026 include ChatGPT Work (launched 9 July 2026), which gathers information across connected apps, produces documents and runs scheduled tasks, pausing for you to approve important actions; Power Automate’s generative actions (in preview), where you give the intent and the AI picks the steps; and managed agents installed around a specific business process.
Example: an agent reads the inbox, looks up each order, checks the carrier, drafts the reply, and queues any refund for a human click.
Handles: multi-step jobs with judgment in the middle. Breaks when: a tool fails mid-plan, an email contains hidden instructions, or the agent loops. Stanford’s 2026 AI Index reports agents still fail about 1 in 3 attempts on structured benchmarks, and Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027. The fix at this level is procedural: cap steps per run, log every tool call, and keep a person approving anything that sends, pays or publishes.
Which type of automation do you need?
| Level | Pick it when | Setup effort | Supervision |
|---|---|---|---|
| 1-2: Rules, macros, scripts | The job lives in one app and inputs look the same | Minutes to an hour | Set and forget |
| 3-5: Workflows, RPA, process flows | 2+ tools must talk; journeys have steps and timing | Hours to days | Alerts on failures |
| 7: AI step in a workflow | Inputs are free text, images or PDFs | Days, plus a test set | Spot-check drafts |
| 8: AI agent | Several steps and a judgment call per run | Days to weeks | Approve consequential actions |
Climb one rung at a time. The companies that get burned usually jump from level 1 to level 8 because a demo looked good. The ones that get value run the lower levels for months first, because that’s where they learn which jobs have the volume and the clean data an agent needs. Our automation strategy guide lays out that sequence for a 10 to 200 person company, and AI vs automation explains the difference between levels 7 and 8 in more depth.
Where AutomationHero fits
We install level 8 for ecommerce and SaaS teams, with level 7 discipline: each agent is scoped to one job, grounded in your data and installed in one session, and you approve anything that sends, refunds or publishes. The reviews agent and the store morning brief are typical first installs because the inputs are repetitive and the output is easy to check. The full agent catalog is here.
FAQ
What is the simplest form of software automation?
A rule on a trigger: when this happens, do that. An email filter, a scheduled cron job or a recorded Excel macro. It needs no code beyond the condition and the action, and it’s the right first step for most small teams because it costs nothing and never guesses.
What are the main types of automation?
Eight levels in rough order of complexity: rules in apps, macros and scripts, cross-app workflows (Zapier, Power Automate), RPA (UiPath), process automation with branches and approvals (HubSpot, Klaviyo), build and test automation (GitHub Actions, Playwright), AI steps inside workflows, and AI agents that plan their own steps.
Is RPA the simplest form of automation?
No. RPA is level 4 on this ladder: a recorded sequence of screen actions, which is fragile and mostly needed for systems without an API. Email filters, macros and cron jobs are simpler, cheaper and more reliable. Some vendor pages call RPA “the simplest” because it’s the easiest to buy, which is a different claim.
When should I use an AI agent instead of a workflow?
When the job needs several steps with a judgment call in the middle, the inputs vary, and you have enough volume to justify supervising it for a month. If you can write the rule, write the rule; if only one step needs a model, put the model in a workflow (level 7) and keep the rest deterministic.