What Customer Service Actually Costs a Bangladesh Online Store
Everywhere else in the world, the case for AI customer service is a wage story: an American support ticket costs eight dollars of human time, the machine does it for one. Bangladesh breaks that story on first contact with a job listing. The most common advertised salary for a customer service executive in Dhaka is Tk 12,000 to 20,000 a month, and at typical throughput of around 21 tickets a day, a human-handled conversation here costs Tk 18 to 63. Imported AI platforms charging a dollar per resolution do not beat Dhaka wages on arithmetic.
So the honest question is not whether AI is cheaper per ticket. It is why stores that can hire this cheaply still lose money on support. The answer is that salary is the only cheap part.
The arithmetic of always-on
A week has 168 hours and the Labour Act caps an agent at 48, so one always-answered seat needs 3.5 agents before anyone takes leave, in practice four to five. Load the salary honestly, two festival bonuses a year, gratuity after year one, night hours at double rate as the law requires, and each agent costs Tk 14,000 to 29,000 a month. One genuinely 24/7 human seat therefore runs Tk 56,000 to 145,000 a month, for a store whose entire support volume might be twenty conversations a day.
Then the roster decays. Call-center turnover runs 40 to 45 percent a year, with first-year attrition near 70 percent, and a new agent takes six to eight months to reach full competence. The store is perpetually training someone, and the gap always falls on the night shift, which is exactly when an unanswered message turns into a refused cash-on-delivery parcel. Pathao's published example prices that failure mode at Tk 117,000 a month for a hundred-order-a-day store.
What the machines actually charge
The 2026 market prices for AI resolution are public: HubSpot charges $0.50 per resolved conversation, Intercom's Fin $0.99 per resolution, Salesforce's Agentforce $2 per conversation, Tidio's Lyro about $0.50 to $0.65. In taka, Tk 62 to 247 per conversation, at or above the human cost. What changes the equation is not imported per-ticket pricing but structure: an AI layer is the whole roster at once, every hour of the week, with no attrition, no ramp time and no overtime multiplier, and local pricing has already adjusted to local wages, with Bangladeshi AI plans starting under Tk 4,000 a month, less than three days of one human agent.
The best-documented deployment shows both edges. Klarna's assistant handled 2.3 million conversations in its first month, the work of 700 full-time agents, cutting resolution from 11 minutes to under two. A year later Klarna publicly rehired humans for judgment calls, its CEO conceding customers must always be able to reach a person. The mature reading, which Gartner's projection of 80 percent autonomous resolution of common issues by 2029 supports, is a division of labor, not a replacement.
The model for a Dhaka store
Run it for a store doing a thousand orders a month. E-commerce benchmarks put support volume at 20 to 46 tickets per hundred orders, higher for small stores, so call it 200 to 500 conversations, a third of them where-is-my-order. Today that is one drowning part-timer or a Tk 56,000-plus always-on payroll. The honest AI model resolves 30 to 65 percent of that band autonomously, not the 80 to 90 percent the marketing claims, and hands the rest, with context, to the person you already employ. That person stops being a night shift and becomes an escalation desk.
This is the shape Omnistra runs for Bangladesh: agents priced for this market that answer in the customer's own language on every channel at 3am, confirm the cash-on-delivery orders, and flag the ten conversations a day that need human judgment. And the metric we would tell you to track is not cost per ticket, where Dhaka wages already win, but cost per delivered parcel. In a 75 percent COD market, the support layer's real job is getting the parcel accepted at the door, and that is the number that moves.
