AI credits pricing explained: what your chatbot invoice is hiding

Published August 23, 2026

Credit-metered chatbot billing is per-resolution pricing wearing a different hat. Here's how credits actually convert to dollars, and what to ask before you sign.

Chatbot pricing pages increasingly quote a number that isn't money: 10 million AI credits, 30 million, 100 million. Credits sound generous — who could use ten million of anything? — but a credit is not a message, not a conversation, and not a customer helped. It's a metering unit the vendor defines, and the definition is where your invoice actually lives. The same customer question can cost a different number of credits depending on which model answered, how long the conversation ran, how much knowledge the bot retrieved to ground its reply, and whether it called any tools along the way.

The tell is the calculator. When a vendor publishes an 'AI credits calculator' next to its pricing page, that's an admission that the pricing page alone can't tell you what you'll pay. You are being asked to forecast your support bill by estimating conversation volume, average conversation length, model mix, and retrieval depth — four numbers your support team does not control and your finance team cannot audit after the fact. Per-resolution pricing at least billed you per solved problem; credit metering bills you per token-shaped unit of vendor-defined work, which is the same variable-invoice problem with less legible math.

Three specific questions cut through any credit scheme. First: what does one typical grounded support conversation cost in credits, on the model you'd actually run — not the cheapest one in the demo? Second: what happens at the ceiling — does the bot stop answering mid-month, degrade to a worse model, or silently start metering overage credits at a separate rate? Third: if you upgrade the model because answer quality demands it, what happens to the effective price per conversation? Vendors quote allowances on their cheapest model; teams discover the real burn rate after switching to the model that's actually good enough for customers.

There's a structural reason this billing model spread: it passes the vendor's model bill through to you with a margin, while keeping the margin invisible. That's not villainy — someone has to pay OpenAI — but it means the vendor profits when conversations get longer and models get pricier, which is a strange incentive for a product whose job is resolving things quickly. The clean alternatives are flat pricing (the vendor absorbs model variance and prices it into the subscription) or bring-your-own-key (you pay the model provider directly, read that bill yourself, and pay the vendor for the product layer).

Dchat's position, stated plainly: flat per-seat pricing with no credits, no per-resolution fees, and no meter that moves with model choice. Managed AI is $29 per AI seat per month with Dchat absorbing provider usage; customer-managed AI is $15 per seat and you pay OpenAI directly on your own key, so the AI line item on your budget is the provider's actual price with zero markup. A busy month costs the same as a quiet one. If a vendor's pricing needs a calculator, ask what the calculator is protecting.

Whatever tool you choose, run the audit once before signing: take last month's real conversation count, ask the vendor to price it in writing on the model you'd deploy, and compare that number — not the tier price — across candidates. Credit allowances are marketing; cost per grounded conversation on the model you'll actually run is the price.

A credit is a vendor-defined metering unit — the same conversation costs different amounts by model, length, and retrieval.

Three questions: cost of one typical conversation, behavior at the ceiling, and price impact of upgrading the model.

Flat per-seat or bring-your-own-key pricing makes the AI line item auditable; credit meters don't.