Pricing an AI agent platform is one of the hardest things I've done. Share a bit about it below
Per-agent? Per-seat? Per-query? Usage-based? Flat rate?
I've changed my pricing model 4 times in 12 months. Not because
I wanted to. Because the market kept telling me I was wrong and
I kept not listening fast enough.
This is the part of building nobody puts on their launch page.
Model 1: Per-agent subscription
I thought this was it. Clean. Predictable. I was genuinely proud
of it, showed it to my team like "we figured it out."
Then reality hit: "Can we get unlimited agents to test first?"
Every. Single. Discovery call. The pricing model I loved was
literally the reason deals died. I held onto it for 3 months
longer than I should have. Ego, probably.
Model 2: Per-seat (user license)
I switched to something enterprise buyers already knew. Per-seat.
No explanation needed. Easy budget line item. Smart, right?
Nope. AI agents aren't people. One agent processing 10k documents
has one "user," the system. A VP of procurement looked me in the
eye on a Zoom call and said: "I genuinely don't understand what
I'm paying for."
I laughed it off on the call. Didn't sleep that night.
Model 3: Usage-based (per API call / per token)
Ok so this time I went "fair." You use more, you pay more.
Startups loved it. I thought I finally cracked it.
Then I walked into an enterprise finance review. Two words
killed everything: "budget cycle." They need a fixed number
in January. For the whole year. My "fair" model gave their
CFO a headache. Deal gone. Again.
I remember sitting in a cab after that meeting thinking,
am I even solving the right problem here?
Model 4: Platform fee + outcome-based tier
Where I am now. Fixed fee so finance can sleep at night.
Plus a success tier tied to deployment milestones, so I only
win when they win.
Is it perfect? No. Honestly I still wake up some mornings
wondering if model 5 is coming. But for the first time,
procurement calls don't end with "let us get back to you."
They end with "send us the contract."
That felt like progress. Real progress.
Here's the thing nobody told me when I started:
You're not selling to the person who wants your product.
You're selling to the person who has to justify the invoice.
Those are two completely different humans with completely
different fears.
I don't think I've cracked enterprise AI pricing. Tbh I'm
not sure anyone has.
If you've been through this, what model survived contact
with enterprise procurement? And what blew up in your face?
I could use the honesty rn.
Replies
That part about selling to the person who has to justify the invoice really resonates.
For enterprise AI, pricing feels tied to proof. Procurement is not only asking “how much does it cost?” They’re also asking what claims they can safely defend internally: time saved, reliability, security, deployment milestones, and actual usage boundaries.
The pricing models that survive may be the ones that make those claims easier to audit: fixed enough for finance, outcome-based enough for buyers, and clear about what evidence counts.
Outcome-based pricing sounds right, but only if both sides agree what the outcome means and how it will be proven.
@shangyin_wei Totally agree, especially the "both sides agree what the outcome means" part. That's where most outcome-based pricing quietly falls apart in practice. I've seen deals stall for weeks because the buyer defined "time saved" differently than the vendor did. One side counted hours reduced, the other counted tickets resolved. Same metric label, completely different math.
The fix that worked for us was defining the success criteria together before the contract, not after. Sounds obvious but almost nobody does it. When you let procurement fill in their own definition of ROI after signing, you're basically handing them a reason to churn. Fixed base + outcome bonus with pre-agreed KPIs has been the least painful model so far. Not perfect, but at least both sides know what they're measuring from day one.
The fixed base + outcome bonus framing is probably the least bad version I’ve seen, but I’d separate “pricing metric” from “proof packet.”
The buyer needs the pricing model to feel predictable; procurement needs the proof packet to be defensible later. For AI agents, that packet might be: baseline workflow, agreed KPI, measurement window, exclusions, and who owns verification. Without that, outcome pricing can turn into a debate about attribution instead of a shared upside model.
The mistake I see a lot is making the pricing page explain the whole thing. The page should reduce anxiety; the proof packet can carry the complexity.