The pitch this year is a two-line spreadsheet. Average digital agency net margin in 2025: 13%. White-label AI agents: 60 to 80. Rebadge the bot. Keep the retainer. Watch the P&L heal.
Those numbers are both real. They do not sit on the same line.
Promethean Research asked 119 agency leaders for the 2026 State of Digital Services. After-tax net was 13%, down from 14% in 2024 and a long-run 15% since 2015. Average revenue in the sample was $4.43 million. About $575k left over. Not a crisis. A slightly compressed normal.
The 60 to 80 figure is vendor-claimed gross on AI-delivered work. Platform fees of $49 to $500 a month against retainers of $2,500 to $10,000. Support, onboarding, sales time, and churn are usually off the slide. That is the current sales script for agency AI margins in 2026. Gross next to net, on purpose.
The leak already has a name
Promethean already measured this inside ordinary delivery. Among the 59% of shops that track it, average project margin was 35%. Firm net was 13%. Twenty-two points vanished into overhead, non-billable hours, and the cost of finding the next job.
White-label AI does not close that gap. It moves the same overhead onto a new SKU. Labor in cost of goods becomes a platform bill, a token bill, and your support queue. Vendor decks stop at the platform bill.
Size is what eats the net line, not the model. Studios with 0 to 9 people ran 19%. Shops at 50 and up ran 8%. Nick Petroski, same report: some of the recent compression is clients expecting cheaper work because AI made production cheaper. They asked you to hand the savings over.
Adding a line made it worse. Agencies that cut services in 2025 ran 30% net and grew 13%. Agencies that expanded the menu ran 10% net. Industry average: 13% net, 7.5% growth. An extra "AI agents" offer on a messy menu is the expander path.
Value-based pricing fell from 31% of agencies in 2024 to 18% in 2025, and those shops grew slower than everyone else. Specialists still won the project line (37% vs 27%) and lost the firm line for the first time in that dataset: 12% net against 16% for generalists. "We are an AI agency now" has the same shape.
Put support on the page
Here is a published year-1 P&L for a cart-recovery agent sold into an existing book. Forty SMB ecommerce retainers. Six weeks and $25k to stand it up. Price: $400 a month. Half the book bought it.
Year 1 revenue: $96k. Costs: $25k build, about $3k in tokens, and $15k of agency support time. Gross: $53k, 55%. Years 2 and 3, build paid, more of the book on it, they put margin around 75%.
Without the $15k support line, year 1 looks like the vendor pitch. With it, you are in the mid-50s. Still good. Not 80.
Platforms print the same story when they show a real table. One operator: 40 to 55% gross at 5 clients, 55 to 65% at 10, 65 to 75% at 20. A voice vendor's single-client sheet lands at 28 to 64% once you load the full monthly fee plus usage against a $497 to $997 price. The 80 lives at the far end of a full book, after you ate the onboarding.
Even native AI software companies do not hold classic software gross. Bessemer put them at 50 to 60% in February 2026, because inference and a human in the loop sit in cost of goods. If the people who own the model cannot keep 80, a reseller who inherits the ticket queue will not either.
Three businesses, one sales page
Mixing these is how shops lose money.
Rebranded software: you sell access, you own support. A common white-label operations platform at $497 a month, resold around $297 per client, looks like 83% gross at 10 accounts -- before support.
Metered messages and add-on seats push real spend 30 to 50% above the sticker. Agencies report 6 to 8 weeks before the team can stand up a client account from scratch, then a portal queue that never leaves. One of those platforms has 60,000-plus resellers making the same product decisions. Software resale is a support business.
AI-delivered services: you keep service prices, the agent does the production. This is the 60 to 80 slide. The real win is the long tail: small accounts nobody could watch every day.
Per-agent SKU: a few hundred a month for the platform, $300 to $500 per deployed agent, then a 3x to 5x markup on usage. Fine as a product. Painful if you price it like a retainer and staff it like a helpdesk.
Onboarding does not shrink like software. Operators who timed it got 4 to 8 hours per client for a basic agent, and volume did not help much. The client calls you when a conversation goes sideways. Price power dies the week a prospect gets three identical proposals and the honest answer is that you run the same platform as the shop down the street.
Churn is the quiet line. At 10% a month you replace the book every year. On a $500 cost to land a client and $300 monthly profit, life-of-client value is about $3k. At 2% a month it is about $15k. A 70% gross offer with 10% monthly churn is a 13% business wearing a software costume.
Simon-Kucher surveyed 182 business-services leaders over the 2025-2026 turn. Ninety percent were on AI or within a year of it. About 18% were actually ahead. Promethean had 34% of digital agencies implemented across the firm by early 2026. Adoption is common. Capture is not. Their line: "Efficiency gains that stay as margin are not a strategy. They are a deferred reckoning."
The shops that moved 2025 numbers did something boring. They raised rates inside the model they already had and grew engagement size in current accounts. Those shops: 13% growth, 37% project margin. Cheaper than a platform migration.
One client, 90 days
Before you quote 60% on a white-label agent, run this on one real account.
revenue_90d = monthly_price * 3
cogs_90d = platform_fees + token_usage
labor_90d = (config_hrs + qa_hrs + support_hrs) * loaded_hourly
churn_hit = monthly_price # if they cancel at day 90, else 0
gross_90d = revenue_90d - cogs_90d - labor_90d - churn_hit
gross_pct = gross_90d / revenue_90d
If gross_pct still clears your current project margin (35% among shops that track it), the SKU is real. If the math only works when support hours are zero, you inherited a helpdesk with a new logo.
The tool can work. We run production agent systems, including white-label work where we stay invisible under someone else's brand. The delivery model is the product. Design support, onboarding, and a reason you are not the shop down the street, or the 80% never shows up on the net line.
If you want a second set of eyes on a live client P&L, first conversation is free. Start here. Members get the PDF version of this math at kief.studio.