When we rolled out the new subscription tiers in 2026 with AI capabilities bundled in, every other agency I talked to told me I was going to bleed money.
"AI work has variable cost. You can't flat-rate it."
A year in, we're fine. Better than fine, actually. Here's the economic thinking behind bundling AI into a subscription — and why I think the conventional wisdom is wrong.
The objection
The worry goes: AI engineering is highly variable. One project might take 20 hours; another 200. If you cap the subscription tier at a fixed price, the 200-hour projects destroy your margins. Therefore: price AI per-project, not per-subscription.
This is exactly what most agencies do. It's also wrong at our scale.
Three reasons flat-rate actually works
1. Portfolio averaging
If we had one client, flat-rate would be insane — all the variance hits us on each engagement. But at ~30 active subscriptions, the variance averages out. Clients don't need 100 hours of AI work every month; they need bursts. One month a client uses 40 hours, the next month 8. Across 30 clients, the aggregate lands near the per-tier cap we set.
This only works if you have enough clients that variance cancels out. If you're a smaller shop, per-project AI pricing does make sense. At our scale, flat-rate is the cleaner product.
2. Scope discipline becomes the feature
When AI hours are capped per tier, every engagement starts with a scope conversation: "Here's how many hours of AI work fits this month — what's the highest-leverage use of them?"
This is the opposite of the unconstrained project-based engagement where scope inflates forever. The cap forces prioritization — which actually produces better outcomes than unlimited budgets.
Clients end up shipping smaller, more focused AI projects. Those projects work. Unlimited-scope AI projects are where most of the field is failing right now.
3. Predictability is the product
Mid-size businesses don't want "AI on a variable cost." They want predictable monthly cost for access to AI capabilities, so they can plan without fear of a $180K bill for a three-week agent project.
Flat-rate lets them say yes to exploring AI in a way variable-cost pricing doesn't. The psychological friction is lower. They try more things. They get more outcomes.
The clients where it breaks
It's not infinite. Some engagements genuinely don't fit:
- AI-first startups that need only AI work, at scale, with no design need. They should not be on our subscription; a specialist engineering shop fits them better.
- Enterprise-scale AI platforms needing custom model work, full teams, year-long roadmaps. Again, wrong shape — Custom tier handles these specifically.
- Regulated industries requiring dedicated infrastructure, custom hosting, specialized compliance. Custom engagement, not subscription.
For everyone else — AI-curious businesses between $500K and $50M revenue — the flat rate is the right shape.
What the numbers look like
I'll spare the exact math, but the shape is:
- Pro tier ($1,895/mo): 20 hours AI engineering cap. Average utilization lands closer to 12.
- Enterprise tier ($3,495/mo): 60 hours cap. Average utilization ~40.
- Overage rate ($150/hr) kicks in rarely. Most clients don't hit the cap.
The bundling lets us sell a qualitatively different product from "buy design, buy AI separately." The integration is the moat.
The strategic point
We priced AI into the design subscription not because it's cheaper to deliver, but because it's easier to sell as one thing than two. Buyers don't want to evaluate "do we need AI help?" as a separate decision — they want to know their design partner can handle it when the question comes up.
Flat-rate is a product decision, not an accounting decision. The fact that the math works is a bonus.
What I'd tell other agency founders
If you have fewer than 15 active clients, price AI per project. If you have more than 25, start experimenting with bundling. Run the math on your historical AI hours. You'll probably find, like I did, that the variance is smaller than the anxiety suggests.
The anxiety is real. The math is fine.