Atiq Israk has shipped AI products in emerging markets where SMS fees, device costs, and thin margins make token spend visible on every P&L. This guide explains how to price AI products when you cannot subsidize inference forever.

Key Takeaways

• Price AI products on value per successful task, not per seat or per token—customers and finance both understand outcomes.
• In thin-margin markets, every AI feature needs a cost ceiling tied to the metric it moves (handle time, stockouts, conversions).
• If unit economics do not work at pilot scale, the problem is scope or retrieval—not "wait for models to get cheaper."

Why Is AI Pricing Harder in Emerging Markets?

US-centric pricing advice assumes customers tolerate usage-based bills and high ARPU. In Bangladesh and similar markets, I have shipped products where operators watch every dollar—restaurants on Neoshift POS, retailers on AssetIQ, hospitals on UHL. Token costs that look trivial in a Series B budget can erase margin on a $29/month SKU.

That forces a discipline Silicon Valley often skips: name the cost per successful task before you name the price.

What Should You Measure Before Setting Price?

Build a simple triangulation table internally before customer pricing:

1. Cost per successful task — tokens + latency + human fallback, averaged over your eval set.
2. Value per successful task — revenue gained or cost removed (minutes saved × labor rate, fewer stockouts × margin).
3. Adoption threshold — minimum usage for the customer to see ROI in one billing cycle.

When Navbot automated 85% of restaurant inquiries, we priced against handle-time savings—not "AI premium." If the bot cost more than the staff hour it replaced, the feature was dead on arrival.

How Do You Set a Cost Ceiling for an AI Feature?

Use the Find the Money step first: what dollar metric must move? Then work backward:

• If the feature saves $500/month in labor, your fully loaded inference cost should stay under ~20–30% of that at scale (rule of thumb, not law—adjust for your gross margin).
• If it does not clear that bar, narrow scope (fewer intents, smaller context window, cheaper model on the happy path).
• Document the triangulation in the PRD/eval doc—not a footnote for finance to discover post-launch.

AssetIQ's inventory reconciliation justified higher per-store spend because stockout reduction drove 264% revenue growth for the business using the platform. The AI line item was small against recovered sales.

Should You Pass Token Costs to Customers?

In emerging markets, transparency beats surprise. Options that work:

• Included quota in base price with overage aligned to outcomes, not raw tokens.
• Tiered plans by successful tasks per month (matches operator mental models).
• Offline-first fallbacks that reduce inference when connectivity or budget is constrained—see emerging markets PM.

What Kills AI Pricing Decisions?

Three anti-patterns I see repeatedly:

1. Pricing the demo — wow factor in a pitch, no eval at production volume.
2. Ignoring fallback cost — human review eats the savings you promised.
3. Copying US list prices — without local willingness-to-pay and payment rails.

Explore: case studies · frameworks · emerging markets PM