Pricing in the AI era can’t be guesswork
As a CFO, I’ve learned the hard way that getting pricing wrong doesn’t just cost you margin; it puts growth at risk. In AI, that risk compounds.
On November 4th, join Andreessen Horowitz General Partner Martin Casado for a live session on how AI is reshaping SaaS monetization. He’ll share lessons from startups and public companies building from the Monetization Operating Model for this new era. Whether you’re exploring usage-based pricing, hybrid models, or outcomes, this session will show how to design with confidence.
For the last three months I’ve been in the lab.
I’ve spoken to founders and execs who are supercharging the tools we love and use each day with AI - from ERPs to Close Management to Invoicing and Billing to Procurement - to figure out wtf is actually real and game changing.
The output is a comprehensive report, with market maps built for each distinct layer of the cfo tech stack.
We. Love. Market. Maps.
It’s like Gartner, without the librarian tone or the software-like price tag (my report is free AF)
I’m sending the report early to a select group of finance leaders. Sign up to be one of them.
When to use AI in Finance
Left to right: CTO, CFO, CEO bout to run a probabilistic model without auditability
👋 Yoooo, it's CJ and last week I told you when NOT to use AI in finance.
This week? The flip side.
Where agents actually earn their bread… too complex for RPA (Robotic Process Automation - Zapier, UiPath etc.), but too structured to leave fully manual. The messy middle where AI genuinely delivers value that your ERP and planning tools could benefit from.
(And look, after last week's post, I got alotta replies. Half of you said "thank the accounting Gods someone's saying this." The other half work at AI startups and called me a hater. Both groups are reading this, so let's find some common ground.)
Last week I laid out the red flag question: "Can I accomplish this with a deterministic tool or simple workflow?"
If yes → skip AI.
But what about when the answer is no?
What about processes where:
Rules exist, but they're fuzzy
Data lives in 47 different places
Context matters more than formulas
You currently have a human doing CSI detective work for 10 hours a week
That's where we need to talk.
Because here's the thing: I'm not anti-AI. I'm anti-bullshit. And there ARE use cases where agentic AI is really the right tool for the job.
(Shocking, I know. It's almost like nuance exists.)
The Spectrum Framework
Think of finance workflows on a spectrum:
SIMPLE → Traditional automation (RPA/workflows)
MODERATE → Agentic AI (the sweet spot)
EXTREME → Human-led (with AI assist)
The magic happens in that middle zone. Let me show you what I mean.
SIMPLE → AUTOMATE IT
Characteristics: Deterministic, rule-based, low variability, high volume.
Examples:
Basic invoice processing (standard POs)
Expense report approvals (within policy)
Simple GL reconciliations
Amortization and depreciation schedules
Standard month-end journal entries
Tools: Your ERP workflows, basic RPA, existing finance platforms
If you can write it as "if this, then that" → don't pay AI prices for it. Big cat don’t play those games.
MODERATE → THE AGENTIC SWEET SPOT
Characteristics: Variable inputs, requires context and judgment, multi-source synthesis, high volume.
This is where it gets interesting.
Tier 1 - The Proving Ground:
Invoice matching (when POs vary) - Same vendor, different amounts, timing variations
Duplicate invoice detection - Legitimate vs. erroneous duplicates
Renewal alerts and tracking - Auto-renewal clauses, usage thresholds, price escalations
Purchase requisition prep - Automating mundane workflow initiation
Compliance flagging - Understanding policy, context, exceptions, and intent
Tier 2 - High Value:
Collections & AR management - Customer behavior, economic conditions, relationship history, dispute resolution
Countless variables impact outcomes. Like, do you always email your biggest customer an invoice directly rather than sending it from your billing system?
Pipeline forecasting - Synthesizing signals from your CRM, email sentiment, call transcripts, rep behavior, macro trends
Churn prediction - Multiple behavioral signals, usage patterns, support tickets, payment history
Procurement end-to-end - Requisition → negotiation → renewal
Market conditions constantly changing
Month-end close orchestration - Involves accounting, FP&A, ops, product teams, requires context-switching and prioritization
Cash flow forecasting - Requires input from AR, AP, treasury, sales (deals closing)
Why these work:
High data velocity from disparate sources
For churn analysis you’re looking at rep inboxes, call logs… and 17 other places
Too much variability for simple rules (but patterns exist)
Require fuzzy logic and contextual interpretation
Currently manual, time-consuming, and high-volume
Key differentiator: RPA needs explicit rules ("if this, then that"). Agentic AI handles fuzzy logic, context, and judgment calls (“use this context and data to go accomplish end goal”).
EXTREME → STILL HUMAN-LED
Characteristics: High stakes, high complexity, strategic implications, regulatory risk
Judgment-heavy but agent-assisted:
Revenue recognition (complex deals with services components)
Get it wrong, maybe go to jail, idk
Tax planning and strategy
Note: It is indeed cool to use AI to kickoff research for applicable tax and gaap accounting advice. But double check it.
M&A financial modeling
“AI, go buy that guy’s company… ”
Capital allocation decisions
Pure human domain (agent support only):
Final contract negotiations
Note: I have been told I negotiate like a robot. Little do they know, it’s part of the strategy.
Organizational design and headcount planning
Investor relations and fundraising
My general standard is when someone could get fired or do three to five in camp cupcake due to an error or oversight → keep humans in charge.
TL;DR What Makes Something "Agent-Ready"?
Here's your checklist before you even take a vendor call (I shamelessly stole it from Ali Hussain, Co-Founder and CEO at Tabs):
✅ High variability in inputs - Not the same thing every time
✅ Multi-source data synthesis required - Pulling from 3+ systems
✅ Context and judgment needed - But patterns exist in the chaos
✅ Currently manual, time-consuming, high-volume - Someone's doing this 5+ hours/week
✅ Can validate outputs against known answers - You can check the AI's homework
If you check 4 of 5 boxes → worth exploring.
If you check 2 or fewer → it's either too simple (use automation) or too complex (keep it human).
Real World Examples (Because Theory is for the Birds)
Let's talk about companies actually doing this well:
Tropic - Procurement Intelligence
The moat: $15B data management asset, 30,000 suppliers, 150,000 contracts, multiple negotiation cycles with each of them.
Why it works:
Started as a negotiation service company (humans first)
Gathered data along the way from actual negotiations
Now uses that dataset to power agent-assisted procurement
The agentic capabilities:
Invoice matching and duplicate detection
Compliance flagging (what's not in policy)
Purchase prep assistance
Renewal agents (alerting before renewals come up)
Negotiation support (benchmarking, market intelligence)
Procurement has high variability (every supplier is different), requires context (is this duplicate legitimate?), and benefits from massive datasets (what should I actually pay for this?). And then there’s a human in the loop to handle the final negotiations.
Numeric - Close Management
What they nail: Audit trails and explainability (I think that’s a word but spell check is angry at me) in close management
Why it works:
Month-end close is like nailing jello to a tree
Requires orchestration across accounting, FP&A, ops
High context-switching and prioritization needs
Every decision needs to be traceable (no black boxes)
They're facilitating agent interactions across the close process while maintaining the audit trail CFOs actually need.
You can’t just say go Pinkman and yell:
Multi-Agent Systems (Already Here)
This isn't future-state. It's happening now:
Brex and Navan agents interact today (expense management ↔ travel)
Tropic agents will interact with ERPs (procurement ↔ financial systems)
Numeric's agents facilitate cross-functional close processes
The future is modular and open.
(At least, we hope so. Closed ecosystems can go kick rocks.)
The Questions You Should Ask Vendors
When you're on that demo call and the sales rep is doing their thing, ask:
"Is there actual math to accomplish this outcome, or is it just picking from a haystack?"
"What's the reasoning chain? Can I audit decisions?"
"What's your error rate on financial processes?" (If they say "95% accurate" → run)
"Can this integrate with my existing stack, or am I locked in?"
"Show me a customer deployment story with real metrics." (Time saved, accuracy gains, ROI—not testimonials)
If they can't answer these clearly, they're selling you a chatbot with good marketing.
“It’s” here, but not all of “It”
The future is here. It’s just not evenly distributed.
The agentic sweet spot exists. It's real.
You can find it in processes that are:
Too dynamic for RPA (customer behavior, market conditions changing)
Too cross-functional for simple automation (close, forecasting, procurement)
Too complex for simple rules (requires fuzzy logic and context)
But structured enough to learn patterns (you can validate results)
Companies doing it right:
Start with data cleanup
Test in sandboxes with known answers
Demand <1% error rates on financial processes
Require audit trails and explainability
Measure everything obsessively
Companies doing it wrong:
Deploy AI to solve problems they don't have (hammer looking for a nail, software looking for a solution)
Skip the data foundation work (crap in, crap out)
Accept "95% accurate" on financial processes (and go to jail)
Can't explain decisions to auditors
Don't track ROI
Which one are you?
What's Next
Next week: Finance in 2030 - From Transaction Processor to Control Tower
Picture this: You wake up to a briefing from your finance agent: "While you were sleeping, I found five business trends you should look into, reconciled 847 transactions, flagged 3 compliance issues, and proposed 2 policy changes."
This isn't the dark arts. This is where we're headed.
I'll break down:
The realistic timeline (2025 vs 2028 vs 2030)
What CFOs need to prepare for NOW
The uncomfortable question: Will your role exist in 2030? (Yes, but it looks different)
Multi-agent systems and what they mean for finance teams
This is part of my research for "The State of the Agentic Financial Stack" report, launching in November. Full market map, case studies with real metrics, evaluation frameworks, and expert commentary from CFOs and founders building this future.
Sign up here to get the report when it drops - it's free, and it's comprehensive as hell.
Run the Numbers Podcast
In this episode of Run the Numbers, I sat down with one of smartest people in software pricing, Michael Stanisz of Revenue Managment Labs.
For decades, software pricing was predictable: sell more seats, add more margin. But with AI, every usage prompt has a real cost, and the old economics no longer hold. We dive into:
What happens when your product looks like software but behaves like infrastructure
How pricing power erodes as features become free
Why value-based pricing might not be as customer-friendly as it sounds.
The chaos of real-world pricing: from Microsoft cloning your startup
“We still run this on DOS” war stories.
Mostly Growth Podcast
The death of the middle manager is upon us. What are you doing to stay alive out there?
Me and Kyle discuss the rise of title deflation (are we just not going to call people CFOs any more?) and the latest trend of calling people “Founding BDRs” and “Founding Marketers.”
Plus, we talk about momentum as a moat… or a boat… or a… IDK as doing good?
Looking for Leverage Newsletter
Why Unit Economics Matter (again)
This post is a deep dive on how to build a customer level P&L.
Yes, that’s right!
Arriving at a reliable and accurate contribution margin by customer is Nirvana for a PE backed company. It exposes the levers you can pull to ensure you are serving the most profitable customer profile, rather than treating all your customers the same.
Quote I’ve Been Pondering
“The great thing about dead or remote masters is they can’t refuse you as an apprentice.”
Hoping you don’t pay AI prices for RPA,
CJ
