Most process automation projects fail for a boring reason: someone automated the wrong process, or picked the wrong tool for it. This guide covers how to pick the process, which kind of tool fits which job, and how to work out whether it will pay for itself.
Step 1: Pick the Process
Score each candidate from 1 to 5 on four things, then add them up:
- Volume. How often does it happen? Daily scores 5, quarterly scores 1.
- Rules. Can you explain the steps to a new hire on one page? Yes scores 5.
- Pain. Real hours, real errors, or slow customers? Score honestly.
- Clean inputs. Do the inputs arrive in a consistent format? PDFs from ten different suppliers score lower than one form.
Start with the highest total. If you cannot get anything above 14 out of 20, the honest answer may be that you do not have a good automation candidate yet, and that is fine to hear.
Step 2: Pick the Right Kind of Tool
This is where projects go wrong, usually by reaching for the trendiest option. Here is how I decide:
| Situation | Best fit | Examples |
|---|---|---|
| Clean data moving between popular apps, if-this-then-that logic | No-code workflow tools | Zapier, Make, n8n |
| Work that has to happen inside an old desktop app with no API | RPA (software that clicks and types like a person) | UiPath, Microsoft Power Automate |
| Custom logic, high volume, several systems, or strict reliability needs | Custom code and integrations | Python services, scheduled jobs, APIs |
| Messy inputs that need reading or judgement: emails, PDFs, free-text tickets | An AI step inside one of the above | A language model that extracts or classifies, with a human checking edge cases |
Two rules of thumb. First, use the simplest tool that works: if a no-code flow does the job, do not build custom software. Second, add AI only for the step that needs it. A reliable rule-based flow with one AI step for reading the incoming email is usually better than an "AI agent" doing everything, because you can test and predict it. I wrote about this in it is all automation, not just AI. For a concrete example, see how an AI sales agent for freight brokers handles inbound enquiries around the clock.
Step 3: Work Out Whether It Pays
Payback is the check most people skip.
Yearly saving = hours saved per month x 12 x cost of an hour of that person's time.
Example, hypothetical numbers: a weekly sales report takes an analyst 6 hours, so about 26 hours a month. At Rs 600 an hour all-in, that is roughly Rs 15,600 a month, or Rs 1.9 lakh a year. A no-code automation that costs Rs 40,000 to set up pays back in about 3 months. A custom build at Rs 6 lakh pays back in over 3 years, so you should not build it for that reason alone.
Include savings people forget: fewer errors (what does a mistake cost you?), faster response to customers, and work that simply did not get done before. Also include costs people forget: your team's time to test it, and ongoing hosting and upkeep. For typical price ranges see how much AI consulting costs in India.
When Not to Automate
- The process is broken. If two people do it two different ways, fix that first. Automation hardens whatever you give it.
- It happens a handful of times a year. Setup time will never come back.
- Every case is an exception. If judgement is the whole job, you can assist a person but not replace them.
- An error would be expensive and nobody reviews the output. Add a human checkpoint or leave it manual.
What a Good Automation Project Looks Like
- Watch the process being done, several times, by different people. Write down what really happens.
- Build the smallest end-to-end version for one case, with a person checking every result.
- Run it alongside the manual process for a couple of weeks and compare.
- Add error handling, alerts, and a simple way for a person to take over.
- Hand over with documentation. If you cannot change it without calling the builder, it is not finished.
Measure the manual version first. Time it for two weeks before you build anything. It gives you the baseline to prove the project worked, and often it shows the process was smaller or bigger than everyone assumed.
Working With Me
I build automations and AI agents for startups and mid-size companies, and details are on the process automation and AI agents page. If you have a process in mind, describe it to me with rough volumes, and I will tell you which of the tools above I would use and whether it is worth doing.
Frequently Asked Questions
What is process automation consulting?
It is help identifying repetitive, rule-based work in your business and choosing and building the right way to automate it: a no-code workflow, RPA, custom code, or an AI step for unstructured inputs.
What should I automate first?
Choose a process that runs often, follows clear rules, costs real time or causes errors, and has consistent inputs. Score candidates on volume, rules, pain and input quality and start with the highest score.
Should I use Zapier, RPA or custom code?
Use no-code tools for clean data moving between popular apps, RPA when you must work inside an old desktop app with no API, and custom code for high volume, complex logic or strict reliability. Add AI only for steps that need to read or interpret unstructured input.
How do I know if automation will pay for itself?
Multiply hours saved per month by 12 and by the hourly cost of that person's time, then divide the project cost by that yearly saving. A payback of 12 to 18 months or less is usually reasonable.
About the Author: Utkarsh Gupta is an AI, Analytics & Automation consultant with 6+ years of experience. He helps companies across India implement practical AI solutions that deliver measurable business impact. See his AI consulting services in India or get in touch.