AI sounds useful. Everyone around you seems to be using it. But when you look at your own business, you can’t quite see where it fits, and every demo you watch assumes knowledge you don’t have time to learn.
You don’t need to automate the whole business to make a good first decision. Choose one recurring task you already understand, keep a person checking the result, and count the checking time honestly. This guide walks you through that, with a worked rupee calculation.
Founder field note
Why I start with the workflow
By 2018, growth at iGaud was asking more of me than I could handle through my own hours. I began putting the agency’s work into systems and building leaders who could run it. That experience is why I start an AI conversation with the process and the person responsible for it.
Write down the task, the inputs and what a good result looks like first. Then decide which part a tool can help with. The same clarity helps the business keep working when a person is away or software changes.
List the work before you choose a tool
For each task you’re considering, note how often it happens, who does it, what information it needs, what a correct result looks like, and what goes wrong today.
Good first candidates are often small and repetitive: summarising enquiries, sorting messages into categories, drafting a first reply from your approved service details, or writing up internal meeting notes. In each case a person checks the draft before it changes a record or reaches a customer.
Some tasks don’t need AI at all. “Copy this confirmed order number into that sheet” is a job for simple automation or a template. AI earns its place when the task involves understanding varied, messy language. Start with the simplest thing that works, and add complexity only when you need it.
Set clear boundaries for the trial
Take an enquiry-summary assistant as an example. Its input is a customer message (with private details removed) plus your approved service sheet. Its output is a short summary, a list of missing information and a suggested reply. A person checks the facts and the tone. Nothing is sent automatically.
Write down what should stop the workflow for human review: an invented price, a promise you don’t make, private information exposed, or an urgent request missed. Decide who fixes mistakes and how you’d return to the manual process.
Test with made-up or properly approved examples first, and read the tool’s data settings and terms before you use real customer information. A well-known AI brand name doesn’t mean you have permission to upload your customers’ messages.
Include the checking time in your calculation
Scroll across to compare every column.
| Monthly workload | Illustrative calculation |
|---|---|
| 100 summaries done by hand | 100 × 8 min = 800 min |
| AI draft plus human check | 100 × 4 min = 400 min |
| Extra time on awkward cases | 10 × 6 min = 60 min |
| Net time saved | 800 − 400 − 60 = 340 min |
| Value at an assumed ₹600/hour | 340 ÷ 60 × ₹600 ≈ ₹3,400 |
| Assumed software cost | ₹1,500 a month |
| Monthly value after software | ≈ ₹1,900 |
If setting up and testing takes eight hours at the same ₹600 rate, that’s ₹4,800 of time. At about ₹1,900 a month, you’d recover it in roughly two and a half months, provided the workload and quality hold steady.
Separate time capacity from cash savings. Count a financial saving when a real cost is avoided; count freed hours separately when you use them to serve customers or develop the business. Include maintenance, training and exception handling in both cases.
Test normal cases and awkward ones
Run a small batch that looks like your real work, including unclear messages, missing details and requests outside your services. Ten or twenty cases will show the obvious problems. They won’t tell you how often rare failures happen, so keep watching after launch.
Compare the manual and AI-assisted versions on the same kind of task. Track total handling time, corrections, invented statements, and whether the reviewer can see where each fact came from. Save the failures so you can check the next version against them.
Try the no-AI version too: a checklist and a saved reply. If that solves most of the problem, keep it. A simple fix that works beats a clever one you have to babysit.
Decide what would make you stop
Before you expand, name the three possible decisions: keep it as a draft assistant, change the workflow, or go back to doing it by hand. Stop if checking eats the benefit, if the output is too unreliable, or if mistakes would hurt in ways your team can’t supervise.
Expand only when the measured result is useful and someone owns the ongoing review. A workflow that works with today’s price list can break the day your prices change.
If the numbers don’t work out
You still gain something. The exercise usually produces a clearer service sheet, a saved reply or a better enquiry form, and those reduce work even without AI. If the task is rare, a better manual process may be enough. If it’s frequent but every AI draft needs rewriting, find out why before you add volume.
For a first project, I’d always choose work you can inspect over a dramatic demo. A workflow that quietly lightens a known load is easier to trust than a promise to run the business for you.
The pilot brief to write before opening a tool
Worked example: enquiry summaries
Task: turn a message into an enquiry summary and missing-information list.
Inputs: approved service details and a test enquiry.
Reviewer: the person who normally handles enquiries.
Measure: total handling time, factual corrections and missed requests.
Pause when: the system invents a price or commitment, or review takes longer than the manual task.
Run the manual and assisted versions on comparable cases. Keep the approach that delivers a useful result at an acceptable cost, and revisit the test when your services or prices change.
Questions you may have
Which AI tool should a small business start with?
Start with the task, not the tool. Once you know the input, the output and how you’ll check it, most mainstream assistants can run a small trial. Choose the one whose data settings and price you’re comfortable with.
Is it safe to paste customer messages into an AI tool?
Only after you’ve read the tool’s data settings and terms, and preferably with private details removed. Test with made-up examples first.
How do I know if AI is actually saving time?
Time the task by hand, then time the AI draft plus the human check, plus the extra time on awkward cases. If the difference is small, a template may serve you better.
Ask me about one task
Describe it in plain words, no tool names needed: “I spend time on ___ about ___ times a month. The hard part is ___. Is this a sensible first task?” A made-up example is fine.
Read next: before you pay someone to build it, ask to see what happens when the automation fails.
Prefer a conversation on social? Message me on LinkedIn.