A seller I was helping had two tabs open at half past midnight. One was her inbox: 61 unread messages. The other was a spreadsheet where she was trying to decide whether to pay her cousin part-time to answer them.
She kept staring at a single cell. The monthly salary. As far as she could tell, that was the whole decision — pay it, or don't.
She was staring at the wrong cell.
The question was never "can I afford one more person." It's "what does one customer conversation cost me right now, and what does that conversation bring back?" Almost no seller has that number written down anywhere. Once you do have it, the hire-or-automate argument stops being a mood and turns into arithmetic you can actually win.
Start with your cost per conversation
Add up everything you pay in a month to answer customers. Salary or hourly pay for anyone doing replies. Your own hours, priced at what you'd otherwise earn: this is the line sellers skip, and it's usually the biggest one. Any tools you already pay for. Then divide by the number of conversations you actually handled.
Say a part-time helper costs you $900 a month and handles roughly 1,200 conversations. That's $0.75 per conversation. Feels cheap. It isn't, and here's why: most of those 1,200 conversations were never going to buy anything.
That's the part the spreadsheet hides. When we looked at what first-contact DMs actually contain, only about 15% came from someone with real buying intent. The rest were price checks, size questions, people who wandered in from a story, and a steady drip of "hi" with nothing after it. So your $0.75 conversation is really costing you about $5 per conversation that could become a sale. Now the number stings a bit.
Do the same math on the automation side
Automation gets sold as though it's free. It isn't. Be honest with the cost line or you'll make a bad decision in the other direction.
- The subscription. Whatever your plan costs, divided by conversations. This is the obvious one.
- Setup time. Someone has to write down your products, prices, shipping rules, and return policy in a form the assistant can use. Budget a real afternoon, not ten minutes.
- Maintenance. Prices change. A product sells out. A new colour arrives. If nobody updates the assistant, it confidently tells customers something that was true last month.
- The cost of getting it wrong. A bad automated answer is more expensive than a slow human one, because it also costs trust.
Run those four lines and you'll typically land somewhere between five and fifteen cents per conversation at volume. That's the honest comparison: roughly $0.75 against roughly $0.10. The gap widens the more messages you get, because the human number scales linearly and the software number mostly doesn't.
Volume is what flips the decision
At 200 conversations a month, this is not a real debate. Answer them yourself. You'll learn more from reading your own DMs than from any dashboard, and the money you'd spend isn't worth the setup.
Somewhere around 800 to 1,000 conversations a month the picture changes. That's roughly 30 a day, which sounds manageable until you notice they arrive in two clumps: a small one at lunch and a big one after 9pm, when you are cooking, driving, or asleep. The volume isn't the problem. The timing is.
Past 2,000 a month, doing it with people only means you're now running a shift schedule, which is a different business than the one you meant to start.
The cost that never shows up on the spreadsheet
Here's what actually drains money, and it has no line item: the conversation you answered nine hours later.
Buying intent in a DM has a short shelf life. Someone messages you at 11pm about a jacket, gets nothing, and by morning they've either forgotten or bought it from the account that replied. We've written before about how sharply conversion drops with reply delay, and the curve is steeper than most sellers expect. It doesn't decline politely over a day. It falls off in the first hour.
So try this. Take the DMs you answered more than four hours late last month. Count them. Multiply by your average order value, then by a conservative conversion rate, say 10%. That's your night-shift bill. For a seller with 300 late replies a month and a $40 average order, it's roughly $1,200 in sales that quietly went somewhere else. More than the salary she was agonising over.
Nobody puts that on a spreadsheet, because it never appears as money leaving. It appears as money that never arrived.
When hiring is genuinely the better buy
I'm not going to pretend software wins every time. It doesn't. Hire a person when:
- Your average order is large and the conversation is consultative. Furniture, jewellery, custom work, anything where the customer wants to be talked through a decision. A good salesperson earns their salary in two closes.
- Every order is bespoke. If the answer to "how much?" is genuinely "it depends," there's no knowledge base to write.
- You're dealing with angry customers, refunds, or anything reputational. Give that to a human. Every time. It's not a cost question.
- Your product changes weekly. Maintaining an assistant that's always half-wrong costs more attention than it saves.
Notice that none of those are about volume. They're about the kind of conversation. Which is the actual insight here: you shouldn't be choosing between people and automation. You should be sorting your inbox by conversation type and routing each type to whichever is cheaper at doing it well.
The hybrid setup, and why it costs the least
The sellers with the best support economics I've seen all run some version of the same thing:
- The assistant takes first contact, always. Instantly, at any hour. Price, sizing, stock, shipping time, return policy, "is this still available": the questions that make up most of the inbox and none of the value.
- It qualifies quietly while it answers. By the time a human reads the thread, they know what the person wants and roughly how serious they are.
- It hands over on clear triggers. Complaint language, a discount request, a bulk order, anything it doesn't know. And it says so plainly instead of improvising.
- A human works the short list. Not 40 conversations. The six that are worth a person's attention.
The economics of that are hard to argue with. Your per-conversation cost falls because the cheap layer absorbs the volume, and your revenue per conversation rises because the expensive layer only touches deals. You're not replacing anyone. You're stopping a person from spending their day typing shipping times.
If you want the longer version of how this works in practice, we've laid out what actually holds up in AI customer service for online stores, including where it reliably breaks.
Run the audit this week
Twenty minutes, a notebook, no consultants:
- Count last month's conversations. Rough is fine.
- Add up what you paid to answer them, including your own hours at a real hourly rate.
- Divide. That's your cost per conversation.
- Read 30 random threads. Tag each one: routine question, real buyer, or complaint. The routine share is your automation ceiling.
- Count the ones answered more than four hours late. That's your leak.
Most sellers finish this and find two things at once. They're spending more than they thought on questions that don't convert, and they're losing more than that again to replies that arrived too late to matter.
That's the case for automating. Not that it's cheap, but that it's cheap at the specific thing that eats most of your inbox.
If you want to see your own numbers instead of mine, Vardast connects to your Instagram, WhatsApp, Telegram and website, answers from your product knowledge in your own tone, and passes the conversation to you when it should. Start with one channel for a month and compare the cost per conversation to what you're paying today. The math either works for you or it doesn't, but at least you'll be looking at the right cell.