People message me convinced their business needs "AI". What they actually have is one task that annoys everyone. Sometimes automating it saves fifteen hours a week. Sometimes it saves twenty minutes and costs more to maintain than it's worth. Ten minutes with these questions tells you which.
- Automate what is frequent, predictable and boring.
- Leave alone what is rare, judgement-heavy or still changing.
- Count hours per week, not "it's annoying". Annoying isn't a business case.
- The first automation should be the one with the clearest payback, not the most impressive one.
The five questions
1. How often does it happen?
Daily or many times a day: strong candidate. Weekly: probably. Monthly: only if it takes a whole day. Twice a year: don't. A task that happens twice a year is cheaper to do by hand forever than to build and maintain.
2. Are the steps the same every time?
Write down what someone does, step by step, for a normal case. If you can, it's automatable. If every second sentence starts with "it depends", it's either two different tasks (fine, split them) or it's a judgement task (keep a person on it, maybe with an AI draft to speed them up).
3. What does it cost today?
Hours per week, times the hourly cost of whoever does it, times 50. That's the annual cost of not automating it. A task that takes one person ninety minutes a day costs your business somewhere around Rs 1.5 to 3 lakh a year in salary alone, before you count the mistakes and the leads that slipped.
4. What happens when it goes wrong?
Some tasks fail quietly: a report has one wrong number and nobody notices for a month. Some fail loudly: a customer gets a wrong invoice. Automation removes the tired-human errors but adds a new failure mode: the tool stops and nobody notices. That's why every build I ship sends an alert if a run fails. If the task is high-stakes, budget for a human approval step. It costs seconds.
5. Do the tools involved have doors?
Google Sheets, most CRMs, WhatsApp Business, Shopify, Razorpay, Gmail: all have "doors" (APIs) an automation can walk through. A desktop-only accounting package from 2011 with no export button does not. If the data lives somewhere with no door, the fix might be moving the data first, not automating around it.
Tasks that pay back fastest
| Task | Why it pays back | Typical time to live |
|---|---|---|
| Replying to new enquiries within a minute | Speed to first reply is the single biggest driver of whether a lead converts | About 10 days |
| Copying leads between WhatsApp, sheets and CRM | Pure copy-paste. Zero judgement. Errors are common by hand | About a week |
| Payment reminders | Directly tied to cash. Nobody enjoys chasing | Under a week |
| The weekly or daily report | Same numbers, same format, every time. Perfect for a schedule | About a week |
| Order or booking confirmations | High volume, fully predictable, customers expect it instantly | Under a week |
| Answering the same ten questions | An AI draft plus a human approval step handles most of it | Two to three weeks |
Tasks to leave alone, for now
- Anything you do fewer than a dozen times a year. Write a checklist instead.
- Negotiations, complaints from important clients, hiring decisions. Use AI to prepare, not to decide.
- Processes that change every month. Let the process settle for a quarter, then automate the version that survived.
- Tasks where nobody can describe the steps. That's a documentation problem first. Sometimes the documentation alone fixes it.
- "Fully autonomous" anything customer-facing. Not for a small business in 2026. Keep a person on the send button.
What it should cost to run
A build has a one-time cost and a running cost. The running cost is what people forget. For a typical first automation in India you're looking at an n8n plan (or a small server), WhatsApp per-message charges from Meta if messages are involved, and a small AI usage bill if there's a summarising or drafting step. Anyone quoting you should put all three in writing before you say yes. If they won't, that's your answer.
A rule I use: if the running cost is more than a quarter of the salary time it saves, the task isn't ready. Simplify it or wait.
How to describe it to a builder
You don't need a document. You need three things: what triggers the task (an email arrives, it's Monday, someone pays), the steps a person does today (messy is fine), and what "done" looks like (a row in a sheet, a message sent, a PDF in a folder). A two-minute screen recording of someone doing it beats any brief I've ever received.
A worked example
A coaching business gets around 30 enquiries a day on WhatsApp. One person copies each into a sheet, replies with the same three lines, and forwards the serious ones to the founder. Ninety minutes a day, seven days a week.
The check: daily (yes), same steps (yes), cost (about ten hours a week), failure mode (missed leads, which is expensive), doors (WhatsApp Business API and Google Sheets both have them). Strong candidate. The build: enquiries land in one sheet, get a first reply within a minute, and anything mentioning budget or a timeline pings the founder. Live in about ten days. The person who used to do it now handles the conversations that need a human.
Frequently asked questions
Do I need to buy new software?
Usually not. Most first automations connect the tools you already pay for. The one addition is normally an automation platform like n8n, which is inexpensive.
Will my team resist it?
Less than you'd think, if the first automation removes the task they hate most. Start there, not with the one that looks best in a presentation.
What if the task changes later?
Good automations are built so small changes are small edits. Ask whoever builds yours to record a walkthrough and write notes so you or anyone else can adjust it.
Can I do this myself?
For simple flows, yes; n8n and similar tools are learnable in a weekend. The parts people get stuck on are the WhatsApp approval process, error handling and edge cases. That's usually where a builder earns their fee.
Deepanshu Singh co-founded LZ Worth, a Delhi studio that builds AI automations, apps and content systems. He has tested more than 150 AI tools and would rather tell you which ones to skip. More about him.