Most failed automations fail before any code is written: the process was unclear, the task was wrong for AI, or nobody owned the result. The steps below deal with those first.
The examples in this guide are public and none is ours. We name no client and claim no result.
Step 1: Pick one process
Choose a process that happens often, follows clear rules, costs little if it goes wrong and can be checked quickly: filing incoming forms, sorting requests, drafting replies for approval, assembling a weekly report. Avoid anything with legal, financial or safety consequences for a first project. Our guide to AI automation examples lists public patterns to choose from.
Step 2: Write the process down
List every step as a checklist a new colleague could follow: what triggers it, what information is used, what decision is made at each point, what the output is, and who gets it. Note the exceptions, such as unusual requests and missing data. If you cannot write the steps, you cannot automate them. Time the process and count how often it runs, so you have a baseline.
Step 3: Decide which steps need AI
Not every step needs a model. Moving a field from a form to a sheet needs a rule. Reading a free-text request and deciding its topic needs language understanding. Mark each step as a rule, a lookup or a language step, and use AI only for the third kind.
Anthropic's post Building effective agents (19 December 2024) advises finding the simplest solution possible and increasing complexity only when needed. It separates workflows, which follow predefined paths, from agents, where the model chooses its own steps, and notes that agentic systems often trade latency and cost for better performance. Choose a workflow unless a fixed path truly will not do. See our AI agents page for the cases that need more.
Step 4: Build the smallest version
Build only the steps from the trigger to the first useful output. Use an established automation tool where you can, because it already handles connections, retries and logs. The tools price by different units: Zapier, for example, says on its pricing page that a task is counted whenever it successfully completes a unit of work, while other tools count credits or executions. Our guide to n8n, Zapier and Make compares them.
If the process touches several of your own systems, plan the connections early. Our AI integration page covers connecting systems and data.
Step 5: Run it beside the manual process
For a few weeks, let the automation produce its result but do not act on it alone. Compare it with what a person did. Count agreements, misses and surprises, and read the misses closely, because they show where the rules or the instructions to the model are weak. Only when the match is good enough for the task, and a mistake is cheap, let it run on its own with sampling.
Step 6: Add checks, a log and a stop switch
- A human check on anything that reaches a customer or changes a record that is hard to undo.
- A log of each run: what came in, what the model was given and what it returned.
- A stop switch that the person responsible can use at once.
- Minimal access: the fewest permissions, the least data and the narrowest actions the task needs.
- Clear wording to customers where an automated reply could be mistaken for a person's, and honest descriptions of what the automation does.
The checks are not optional politeness. In February 2024 a British Columbia tribunal held Air Canada responsible for wrong information its website chatbot gave a customer, treating the chatbot as part of the website and awarding about $650. Torkin Manes' summary (7 March 2024) advises companies to train AI systems to give accurate results and to warn consumers about the limits of AI tools.
In the United States, the FTC said in September 2024 that there is no AI exemption from the laws on the books. Make claims about your automation that you can prove.
Step 7: Measure and decide
After a few weeks, compare against the baseline: time saved per run, runs per month, errors caught by review and cases a person had to fix. Count the time spent building, checking and maintaining. If the numbers are not clearly better, change the automation or stop it. If they are, keep the weekly review of a sample, and only then think about the next process. Our guide to measuring digital marketing ROI uses the same logic for costs and results.
What should you plan for?
- Change. Tools, models and your own process will change, so assign an owner.
- Exceptions. A small share of cases will not fit. Decide where they go.
- Privacy. Know what data goes to which tool, and where. Check your agreements and the laws that apply to your customers.
- Skills. People who review the results need to know what a good result looks like.
Do you need help?
Many teams can run a first automation themselves. Help is useful when the process spans several systems, when the data is sensitive, or when you want a second view on where AI fits. Our AI consulting page covers planning, and our AI automation page covers building and checking.
Conclusion
Choose one process, write it down, use AI only where language is involved, build small, run it beside the manual version, add checks and measure. Keep the person responsible for the result.
PromotEdge International is an international strategy, technology, marketing and design company, and our delivery team is in India. This guide is not legal advice. See our AI automation page.
Facts we rely on: see our company facts for what we state about ourselves and where each fact comes from.