Almost every AI tool being pitched to Indian small businesses right now automates the half of the job your customer can see, and leaves alone the half that is actually eating your week.
That was already the wrong trade. Since the first of this month it is also a more expensive one.
One question sorts every tool on the market, including the ones that don’t exist yet. Does a human read the output before a customer does?
If yes, buy it, because the worst case is wasted effort. If no, you are not buying efficiency, you are buying unsupervised publishing, and in some categories you are buying regulatory exposure as well.
And before you subscribe to anything, go and check what you already have for free. Most of the businesses I talk to are paying a platform fee for something sitting inside an app they already use.
The change that landed on 1 October
WhatsApp service messages, meaning your replies inside the 24 hour window after a customer writes to you, used to be free.
From 1 October they are charged. Business solution providers and resellers are quoting roughly Rs 0.115 per message, with a free allowance that most of them describe as 1,000 service messages per number per month.
Worth being careful here: those figures come from BSP and aggregator pages rather than from Meta’s own rate card, and your provider’s markup sits on top of whatever Meta charges. Go and read your own provider’s pricing page before you plan around any number, mine included.
Now think about what that does to an AI chatbot.
A bot does not reply once. It asks a clarifying question, confirms it understood, offers three options, then asks whether that helped. Where a human sent one message, the bot sent four. Past your free allowance, you are now paying for the bot’s conversational manners.
An AI that talks more than a person would used to be merely annoying. Now it shows up on the invoice.
The two lists, and why the principle matters more
Here is the sorting rule applied, but the rule is the point rather than the list, because the list will be out of date by March.
Worth paying for, because somebody checks the output first. Drafting ad variants you then choose between. Summarising a month of enquiries into what people actually asked. Turning a messy spreadsheet into something a human can read. Transcribing and summarising sales calls. Translating a draft that a fluent person then reviews. First pass keyword clustering before you apply judgement.
Not worth it at any price, because nothing stands between the machine and your customer. Auto-published blog posts. Auto-replies to public reviews. An unsupervised chatbot answering questions about price, stock or delivery timelines. Auto-generated case studies about clients who did not agree to the numbers.
The pattern is not creative versus operational, and it is not cheap versus expensive. It is whether there is a person in between.
What it is genuinely best at, which nobody is selling
This is the part I would actually spend money on, and it is boring, which is presumably why no vendor leads with it.
Consider the overnight pile for a typical business in Bengaluru or Pune. Messages in English, in Hindi, in Hinglish typed in Latin script, in Kannada or Tamil or Marathi. Voice notes. Payment screenshots. Three people asking the same question, one serious buyer, and a wedding invitation from a cousin in the same inbox.
Sorting that is genuinely difficult human work, and it is genuinely good machine work. Group by intent, flag the one that mentions a deadline, surface the two that have already paid, summarise the voice notes so you can read instead of listen, and tell you which four need a human before lunch.
Your customer never sees any of it. Nobody can be embarrassed by a triage summary.
This is the same conclusion I reached when I refined the argument about plugging the timezone leak in export enquiries, and I would now state it as a general rule rather than a nuance. Automate the sorting and the receipt. Do not automate the answer.
The one that could genuinely get you in trouble
If you run a clinic, read this part twice.
The NMC Code of Medical Ethics prohibits a registered practitioner from soliciting patient testimonials, and from amplifying content testifying to a treatment outcome. I went into what that rules out in why Indian doctors cannot follow the standard local SEO advice.
Now picture the review automation tool a vendor will happily sell you, replying publicly to a five star review that describes a successful surgery, thanking the patient warmly, by name.
That is amplification. Performed by software, at scale, in public, with a timestamp on it.
Which points at something broader than clinics. Automation does not only scale your output, it scales your compliance exposure, and nobody audits what a bot has been posting until a complaint arrives. If your category has rules about what you may say, a tool that says things without you reading them first is the single most expensive subscription on your list.
I am not your compliance advisor. If you are in a regulated category, somebody who knows your rules properly should look at your automation before it runs, not after.
Check what you already pay for
Meta launched Business AI for small businesses in India in May this year. It sits inside the WhatsApp Business app, it answers queries, captures leads and books appointments, it hands over to you when something needs a person, and at the time of writing Meta has not put a price on it.
Eligibility is limited and not every business will have it. But a surprising number of owners are paying a third party platform fee, on top of Meta’s per message rates, for something adjacent to a feature they have not opened.
Two habits worth adopting regardless.
Price your stack annually rather than monthly, because monthly pricing is designed to be invisible and annually is how your accountant will see it. Then put that annual figure next to what part time human help would cost for the same hours, and make an actual choice rather than a default one.
And cancel anything you have not opened in a fortnight. You will not miss it.
Where I come out
The tool list stops being accurate quickly. The sorting question doesn’t.
A human between the machine and the customer, every time, and the money spent on the invisible work rather than the visible work. Get that right and you can evaluate next year’s tools yourself in about ninety seconds each.
If you’d like to think through which parts of your own operation are worth automating, drop me a line on email, WhatsApp, phone or LinkedIn and we can have a quick chat, and my notes on how I work across these problems cover the rest. I’m taking on a small number of contracts at the moment, so I’ll tell you straight whether it’s something I could help with.