Pricing White-Labeled Agents for Platform Partners: The Margin Math Nobody Talks About

Table of Contents

• Why White-Label Agents Break the Old Pricing Playbook
• The Three Wholesale Models and How Each One Fails
• The Inference Cost Problem Is Your Problem, Not Theirs
• Setting the Wholesale Floor: A Worked Example
• Contract Terms That Decide Whether You Make Money
• Who Owns the Outcome When the Logo Isn't Yours
• Insights Most People Overlook
• References

Why White-Label Agents Break the Old Pricing Playbook

White-labeling software is a solved problem. You license your code, your partner slaps their logo on it, you charge a per-seat or per-instance fee, and your marginal cost to serve one more customer rounds to zero. The CFO sleeps fine. That comfort is exactly what trips up companies moving into Agentic AI-as-a-Service.

An agent isn't software in the margin sense. Every task it runs burns tokens, and tokens cost real money, money that scales linearly with usage rather than flattening out the way SaaS infrastructure does. When a platform partner resells your agent to ten thousand of their customers and those customers run it hard, your variable cost doesn't disappear into the noise. It becomes the dominant line item. This is the structural reason the broader GaaS pricing taxonomy, per-task, per-outcome, per-seat, exists in the first place, and white-labeling inherits all of it plus a second markup layer.

Here's the part that makes it genuinely hard. In a direct-sale relationship, you set the price the end customer pays and you eat or pass through your costs as you see fit. In a white-label deal, your partner sets the retail price. You only control the wholesale rate. So you're forecasting your own gross margin based on usage patterns of customers you've never met, sold to by a partner whose pricing page you don't write. You are, in effect, underwriting someone else's business model.

The partners know this asymmetry and they will push. A platform doing the reselling wants the lowest possible wholesale rate and the most predictable terms, ideally flat, because their whole pitch to their customers depends on bundling your agent into a clean, simple offering. Your incentive runs the opposite direction: you want the wholesale rate to track usage, because usage is what costs you money. That tension is the entire negotiation.

The Three Wholesale Models and How Each One Fails

There are really only three coherent ways to price a white-labeled agent to a platform partner. Each is defensible. Each has a specific failure mode, and the failure modes matter more than the upside because the upside is what gets you to sign and the failure mode is what shows up on the income statement six months later.

Flat OEM License

You charge the partner a fixed annual or monthly fee, say $200,000 a year, for the right to embed and resell your agent, often with a usage ceiling baked in. Partners love this. It's predictable, it's easy to bundle, and it lets them price their own product however they want without checking a usage meter.

The failure mode is obvious once you've lived it: you've sold an uncapped (or generously capped) call on your most volatile cost. If the partner's customers run the agent harder than you modeled, and successful agent products always get used harder than the model assumed, because that's what "successful" means, your variable inference cost can blow straight through the flat fee. You've turned a profit center into a usage subsidy. Flat OEM works only when you can put a hard, enforced cap on consumption and you've stress-tested that cap against the partner's most aggressive realistic growth.

Wholesale Per-Outcome (or Per-Task)

You charge the partner a wholesale rate per completed task or per successful outcome, say $0.40 per resolved support ticket, and they mark it up to their customers at, say, $1.20. This is the cleanest model from a margin-defense standpoint because your revenue and your costs move together. When usage spikes, so does your top line.

The failure mode here is the audit and attribution fight. Who counts a "resolved" ticket? If the partner's customer disputes a resolution, does the partner still owe you? The question of who defines and audits the outcome is hard enough in a direct deal; in a white-label arrangement there are now three parties who can disagree about whether the work happened. You need contractual, instrumented agreement on the billable event before launch, not after the first invoice dispute.

Revenue Share

You take a percentage of what the partner charges their end customers, commonly 15% to 30%. This aligns incentives beautifully on paper: you both win when the end customer pays more.

The failure mode is that you've coupled your margin to a number you can't see and can't control. If the partner discounts aggressively to win logos, and partners discount, especially early, in what looks a lot like a discounting death spiral, your share shrinks even as your inference costs hold steady or rise. Revenue share also creates a reporting dependency: you're now relying on the partner's billing data to compute your own revenue, which is an audit headache and a trust exercise rolled into one. Revenue share works best when the end-customer price is high and stable, and badly when it's low and competitive.

A practical note from watching these deals: the strongest agreements blend models. A modest flat platform fee covers your fixed integration and support cost, and a per-outcome wholesale rate on top covers your variable inference cost. That hybrid structure mirrors the base-plus-usage logic that's becoming standard across GaaS, and it protects you from the single worst outcome in each pure model.

The Inference Cost Problem Is Your Problem, Not Theirs

This deserves its own section because it's the thing that separates white-label agent pricing from white-label everything-else, and it's the thing partners least want to hear about.

Your underlying model costs are volatile and outside your control. Providers reprice, sometimes dropping costs 50% or more, useful research from a16z on the falling cost of LLM inference has documented order-of-magnitude declines over short periods. That sounds like good news, and over a long horizon it is. But within a single annual contract, you can also face the reverse: you architected your agent around a premium model, the partner's customers demand the quality only that model delivers, and your cost-per-task is higher than you quoted.

In a flat OEM deal, you absorb every cent of that volatility. In a revenue-share deal, you absorb it too, because the partner's retail price doesn't move when your token costs move. Only the per-outcome model gives you a clean pass-through, and even then, only if your contract lets you adjust the wholesale rate when underlying costs shift materially.

The discipline that keeps you solvent here borrows directly from FinOps. You need to know your fully loaded cost per task, not just the headline token price, but retries, tool calls, failed attempts, the cheap-model-when-possible routing you've built, and the human-in-the-loop review your enterprise partners insist on. The FinOps Foundation's framework for cloud cost management is increasingly being adapted for exactly this kind of AI workload accounting, and treating your agent costs with that rigor is the difference between a wholesale rate that's a number you guessed and one you can defend in a renewal negotiation.

One more thing partners don't volunteer: model routing is a margin lever you own and they don't see. If you can serve 70% of tasks with a cheaper model and reserve the premium model for the hard 30%, your blended cost drops and your margin on a fixed wholesale rate expands. That optimization is yours to keep, which is a quiet argument for per-outcome wholesale pricing over revenue share, because in per-outcome your efficiency gains flow to you, not split with the partner.

Setting the Wholesale Floor: A Worked Example

Abstract advice is cheap, so here's concrete math. Say you run a white-labeled support agent. Your fully loaded cost per resolved ticket is $0.18, that's the premium model for hard tickets, a cheaper model for easy ones, retries, a small human-review sampling cost, and your platform overhead allocated per ticket.

Your minimum viable margin, the floor below which the business isn't worth running, is, say, 60% gross. That puts your absolute wholesale floor at $0.45 per resolved ticket ($0.18 cost is 40% of $0.45). You would never go below that for a partner deal, no matter how big the logo.

Now the partner needs room to make money too. If they want a healthy markup, they'll retail your agent at maybe $1.20 to $1.50 per resolution. That leaves them a comfortable spread on a $0.45 wholesale rate. Everyone is solvent. The deal closes.

The trap is the volume discount. The partner brings scale and asks for $0.30 per resolution at high volume. At $0.30 against an $0.18 cost, your margin collapses to 40%, below your floor. The right move isn't to refuse the discount outright; it's to tie the discount to a cost-down commitment. Offer $0.30, but only at a volume tier where your blended inference cost has dropped to, say, $0.10 because routing and caching get more efficient at scale. Now $0.30 is still a 67% margin. The discount is real, the partner is happy, and you didn't sell below cost to do it. This is where floor-and-ceiling thinking, capping the partner's per-unit cost while protecting your own floor, earns its keep.

The number that should never appear in a white-label contract is a wholesale rate that assumes today's inference cost will hold for the contract's full term. Build the rate around a cost band, with a clause that lets you revisit if the band breaks.

Contract Terms That Decide Whether You Make Money

Pricing models live or die on contract language. A few terms do most of the work.

A usage cap or true-up, always. Whatever model you pick, the contract needs a mechanism for what happens when usage exceeds the modeled range. Either a hard cap that throttles, or a true-up that bills overage at a defined rate. Without one, your flat or discounted rate is an unlimited liability. McKinsey's analysis of how companies are capturing value from generative AI repeatedly lands on the same point: the firms that profit are the ones that instrument and govern usage, not the ones that ship and hope.

A cost-adjustment clause. Give yourself the right to revisit the wholesale rate if underlying model costs move beyond a defined threshold, in either direction. Partners will ask for the downward protection (they want your cost savings passed to them). Fine, make it symmetric. If costs drop 30%, you'll share. If they rise 30%, you adjust up. Symmetry is the only version a partner can't reasonably refuse.

Clear definition of the billable event. For per-outcome deals, the single most disputed line will be what counts. Write it precisely, instrument it independently, and make your telemetry, not the partner's, the source of truth. The alternative is invoice disputes that poison an otherwise good partnership.

Minimum commitments. White-label deals carry real integration and support cost on your side. A minimum annual commitment ensures a partner who underperforms on sales doesn't leave you having staffed up for revenue that never arrived. This matters more in white-label than direct, because you've often done custom branding and integration work that doesn't transfer to another partner.

Exclusivity, and what it costs. Partners often ask for category or geographic exclusivity. Exclusivity has a price, and that price should be a higher minimum commitment, not a lower wholesale rate. Never trade your unit economics for exclusivity, trade your guaranteed revenue floor.

Who Owns the Outcome When the Logo Isn't Yours

There's a quieter dimension to white-label agent pricing that pure margin math misses: liability and reputation are now split in a way that should affect what you charge.

When your agent runs under your partner's brand, their customers blame the partner when it fails, but the partner blames you. If your agent makes a costly mistake in an autonomous workflow, the question of who pays, and whether there's a refund or SLA credit owed, has to be settled before launch. And here's the pricing implication people miss: an agent operating under someone else's brand, with the partner's reputation on the line, often demands a higher reliability tier than the same agent sold direct. Higher reliability means more guardrails, more human review, more expensive models on the critical path, all of which raise your cost per task.

So the white-label wholesale rate is frequently higher per task than your direct retail unit cost would suggest, because you're pricing in the reliability premium the partner's brand requires. Counterintuitive, but real. A partner who wants your agent flawless under their logo is asking for a more expensive product, and the wholesale rate should reflect it. Don't let the volume of a platform deal pressure you into pricing the premium-reliability version at the commodity rate.

The deals that endure are the ones where both parties understand they're underwriting a shared outcome. You're not selling a license and walking away. You're entering a multi-year relationship where your costs, their prices, and a shared brand risk are all moving at once. Price for that reality, not for the clean SaaS world the term "white label" makes you remember.

Insights Most People Overlook

The reliability premium inverts the usual wholesale logic. Most wholesale pricing is lower per unit than retail, that's the whole point of wholesale. White-label agents often break this, because the partner's brand on the line demands a higher-reliability, higher-cost configuration than you'd ship direct. Smart vendors charge a wholesale rate that's above their own direct unit cost and justify it with the reliability tier. Partners who understand brand risk will pay it.

Revenue share is a trap disguised as alignment. It feels collaborative, but it hands your margin to a number you can't see and can't control, and it couples your fate to the partner's discounting discipline, which is usually weak in the land-grab phase. Per-outcome wholesale keeps your costs and revenue mechanically linked and lets you keep the gains from your own efficiency work. Choose revenue share only when retail prices are high and stable.

Your efficiency improvements should be a secret, not a giveaway. If your wholesale rate is per-outcome and you improve model routing to cut your cost per task in half, that gain is entirely yours, the partner never sees it. Under revenue share or cost-plus, you'd surrender half of it. This is a structural argument for outcome-based wholesale that almost nobody makes explicitly: it privatizes your R&D wins.

The flat OEM fee is a short option you're selling for free. A fixed annual license against uncapped usage is financially identical to writing an uncapped call option on your most volatile cost and collecting a tiny premium for it. The only safe version has a hard, enforced usage cap stress-tested against the partner's most optimistic growth case. If you can't enforce the cap technically, don't sell the flat fee.

Volume discounts should be earned by cost-downs, not granted by hope. The discipline that saves white-label deals is refusing to discount the wholesale rate faster than your actual cost per task falls. Tie every volume-tier discount to the blended cost you can actually achieve at that volume. A discount that outruns your cost curve is just selling below margin with extra steps.

References

• a16z, Welcome to LLMflation: LLM Inference Cost Trends
• FinOps Foundation, What is FinOps?
• McKinsey, The State of AI: How Organizations Are Rewiring to Capture Value