I've spoken to about 160 agent builders in the last half year, and my estimate says 75% of them are leaving money on the table. They tell me so, and they tell me they're worried about it.
Last week, on a discovery call with a company building a support agent for automotive use, the founder told me "Our agent handles 50,000 customer support tickets monthly, but we're losing $3 on every interaction".
What's missing in my opinion is a bit of education about what you can and can't do for agentic monetization.
So after helping 40+ AI agent companies implement profitable pricing models and having spoken to at least 100 more, I'm going to share the pattern that I see: You've built something incredible, but you're struggling to turn it into sustainable revenue.
The problem: Normal SaaS pricing doesn't work for AI agents
This one is hard to unpack but let me explain:
- LLM costs are unpredictable: One customer saw costs spike from $500 to $8,000 in a single month due to INPUT token lengths changing
- Value varies wildly: The same agent might save one customer $100/month and another $10,000/month
- Usage patterns are non-linear: Unlike SaaS, AI agent usage doesn't scale with team size. There is no seat to price for!
Because we're all used to SaaS, tons of founders default to per-seat pricing because it's familiar. But when one user can deploy 50 agents doing the work of an entire department, seat-based pricing leaves lots to be desired...
The 4 AI agent pricing models that actually generate profit
After extensive testing with real customers, these four models consistently outperform traditional pricing:
1. Agent-based pricing: The FTE replacement model ($2K-$20K/month)
When it works: The agent replaces a specific job function or part of it
Real example: One of our legal tech customers charges $8,000/month for an AI contract reviewer that replaces a $120,000/year paralegal.
The math
Human paralegal: $10,000/month fully loaded
AI agent: $8,000/month all-in
Customer saves: $2,000/month
Your margin: ~65% after infrastructure costs
My take: You can position against headcount budgets, not software budgets. HR budgets are 10x larger than IT budgets and it may work in your favour.
What to track today: Start documenting how many human hours your agent replaces. You'll need this data to justify your pricing.
2. Action-based pricing: The consumption model ($0.10-$5.00/action)
When it works: Competing with BPOs or call centers
Real example: A voice AI company charges $0.12/minute for calls, undercutting call centers by 70%.
Actual breakdown from a customer support agent company we work with:
Voice infrastructure: $0.03/minute
LLM costs: $0.02/minute
Other APIs: $0.01/minute
Total cost: $0.06/minute
They charge: $0.12/minute
Margin: 50%
The problem with this model is it can be "crunched" with margin. When things get commoditized, you will race to the bottom on pricing.
What to track today: Calculate your true cost per action including ALL components: LLM, APIs, infrastructure, even that Google Maps geocoding service you may be using.
3. Workflow-based pricing: The process automation model ($50-$500/workflow)
When it works: Multi-step processes with clear deliverables
Real example: An SDR agent that charges:
Lead research: $2/lead
Email personalization: $1/email
Meeting booking: $8/meeting
Even a simple workflow could involve:
- 3 data enrichment API calls
- 2 LLM summarization steps
- 1 quality verification check resulting in a total cost of $0.47/lead
By bundling into workflows and charging $2/lead, one of our customers achieved 76% margins while still being cheaper than human SDRs.
My tip: Bundle workflows into packages. "500 leads/month for $750" converts better than "$1.50 per lead" even though it's the same price.
What to track today: Map out EVERY step in your agent's workflow. You'd be surprised how many hidden costs lurk in "simple" processes.
4. Outcome-based pricing: The results model ($500-$5,000/outcome)
This is my favourite and it's the holy grail, but it can be tricky to get to it!
When it works: Clear, measurable business results
Real example: A recruiting AI that charges:
- $500 per qualified candidate
- $1,000 per scheduled interview
- $5,000 per accepted offer
- $0 for everything else
As the agent success rate improves (say, from 10% to 25%), the revenue increases and the cost can remain rather flat.
Yes, you need rock-solid attribution. We have one customer that spent 3 months building attribution before switching to this model.
Attribution tracking with Langfuse (if you're already using it):
# Track outcome attribution in your observability
generation = langfuse.generation(
name="interview_scheduled",
trace_id=trace.id,
metadata={
"outcome_type": "interview_scheduled",
"candidate_id": candidate_id,
"attribution_confidence": 0.95, # How sure are you?
"billable": True,
"amount": 1000
}
)
Tracking is one thing. Actually billing for it, handling disputes, and managing different pricing tiers? That's where things get messy - and I suggest you look away from the usual suspects for handling it :)
What to track today: Start measuring success rates and outcomes NOW, even if you're not charging for them yet. This data is so important...
Recommended decision framework for agentic monetization models
Which model should you choose?
Here's my framework:
What budget are you targeting?
→ Headcount budget (10x larger)?
Use Agent-based pricing
→ Outsourcing/BPO budget?
Start with Action-based, plan your escape
→ ROI/Innovation budget?
Use Outcome-based if you can prove attribution
→ Operational efficiency budget?
Use Workflow-based pricing
What should you do today?
Margins will kill you if you don't know what they are. You can't price well if you don't know what your inputs are.
I recommend adding a cost calculation now! If you're using something like n8n, add a cost calculator node after each service:
// Cost aggregator
const costs = {
openai: $('OpenAI').first().json.usage.total_tokens * 0.00003,
whisper: $('Whisper').first().json.duration * 0.006,
elevenlabs: $('ElevenLabs').first().json.characters * 0.00018,
// Add all your services here
};
const totalCost = Object.values(costs).reduce((a, b) => a + b, 0);
// Log it somewhere (or don't, I'm not your boss)
console.log(`Workflow cost: ${totalCost.toFixed(4)}`);
If you're using Langfuse, you can track costs per trace:
# Add cost calculation to your traces
from langfuse import Langfuse
langfuse = Langfuse()
# When creating a trace
trace = langfuse.trace(
name="customer_support_workflow",
metadata={
"customer_id": customer_id,
"token_count": completion.usage.total_tokens,
"estimated_cost": completion.usage.total_tokens * 0.00003
}
)
The above will work for understanding costs, but they won't help you with billing, margin analysis, or scaling.
What you should do next week
After you've started tracking your costs, I highly recommend you talk to 10 customers about value. Don't ask "What would you pay?", instead understand:
- What were they doing before?
- How much time/money did that cost?
- What specific outcomes matter to them?
After that, you can start to figure out which models (as I outlined above) could match.
Test the water with a few friendly customers (simulate it!), and see if one of them leaves you more profitable.
I'm very happy to discuss: What's your biggest pricing challenge?
I'm curious about your experience. Are you currently:
- Still figuring out your costs?
- Struggling to communicate value?
- Worried about customer pushback?
- Already profitable and scaling?
Drop a comment below. I respond to every question and often write follow-up posts based on common challenges.
Top comments (2)
SaaS pricing always felt off to me for agent-based features too - these models are much more practical. What’s the earliest sign for you that a company should switch pricing strategy?
There's so many!
Plus a ton of others.