Strategies for Negotiating with AI Vendors
Insights from CAUCUS AI in Procurement discussions to provide actionable strategies for negotiating AI contracts.
AI Procurement Campfire
Session Recap & Key Takeaways
January 2026
AI companies are obfuscating pricing like crazy. Some aren't even real AI—just automation somebody built in their basement. Here's what members are dealing with:
Einstein, Copilot, Agent Force—they're all being sold as separate licenses instead of platform improvements
"Outrageously expensive" was how one member described Einstein pricing
Salespeople can't explain how tokenization works or what you'll actually pay
One team burned $250k in under four months on AI credits
For small deployments (<1000 users), vendors won't negotiate anything—not even typos
You can't negotiate what you don't understand. And right now, vendors are counting on that.
Say what?
AI is supposed to be the ultimate calculation machine. But vendors can't (or won't) calculate their own costs.
What you're running into:
84% markups on features that should be part of the platform
Sales reps pointing you to documentation that "says nothing essentially"
Zero negotiation on POCs unless you're spending serious money
Standard response: "We can't answer questions about cost scenarios"
This is deliberate. They want you flying blind so they can bill whatever they want.
Three Ways Vendors Are Charging You
1. Baked into the platform
They raise your renewal 10% and say it includes AI features. Simplest to forecast. You can usually toggle it off if you don't want it.
2. À la carte / Add-on license
Separate SKU on top of your enterprise agreement. Pay only for what you use. Easier to control costs but watch out for "you need this to make the other thing work" bundling.
3. Consumption-based (this is where it gets messy)
Pay per token, per API call, per interaction. Vendors can't or won't define what drives costs. This is the highest risk model and where you need the most protection.
Tokenization 101 (What They're Not Telling You)
Every conversation with an LLM has two parts that cost you money:
Input tokens: What you send in
Output tokens: What comes back—costs 3-4x more
But wait, it gets worse:
API calls cost up to 10x more than regular tokens
RAG calls (when it pulls from your internal docs) = extra charges
Different models = different pricing (Sonnet 3.5 vs Opus 4.5)
Every time the AI "forgets" context mid-conversation and has to reload = more tokens
Real Example: Salesforce Agent Force
Say you replace customer service reps with AI agents. Here's what actually happens:
Agent calls the LLM 50 times for ONE customer interaction
When it times out? Retry = you pay again
Error? Retry = you pay again
Needs to pull CRM data? API call = 10x cost
Without clear contract definitions of "conversation" and "interaction," you're writing them a blank check.
Why Traditional Tactics Fail
In SaaS, you'd say "Oracle, I'm buying SAP." In AI? They say "Cool, go ahead. We work well together."
Why this doesn't work:
AI vendors don't see themselves as competitors
Most are reselling LLM access from OpenAI, Anthropic, or Google
They literally integrate with each other
So you need different leverage. Here's what's working:
Hard Caps (Your Best Defense)
Since they won't give you transparent pricing, cap what you'll pay. Period.
What one member negotiated on a $20M Salesforce deal:
Three-year price caps on Einstein, Slack GPT, and Agent Force
Hard monthly cap—anything over that, vendor eats it
Quarterly reviews for overages, not surprise invoices
Annual true-up only
Rollover credits for unused allocation
Why Salesforce agreed? Because they want adoption. Your usage will eventually blow through that cap and they know it. But by then, you'll have real data to renegotiate.
POC Pricing (Yes, You Still Pay)
Forget what you learned in SaaS. AI vendors charge for POCs. But you can still win:
Buy small: 20-30 licenses, not 1,000
Push hard on discounts: Members have gotten 97% off because vendors need adoption numbers
Set success criteria: Treat it like an SOW with acceptance requirements
Structure earn-in: One member got Year 1 completely free with aggressive SLAs—vendor had to earn their way to Years 2-3
Lock future pricing: Contract Year 2-3 rates now, even if you don't commit
Build Your Own Cost Model
When they won't tell you their costs, calculate it yourself:
Go to Anthropic or OpenAI's website—they publish token costs
Estimate your input/output ratio
Add 10x multiplier for API calls
Present it: "I assume you're using Sonnet 3.5. Here's what that should cost. Tell me why I'm wrong."
Either they accept your pricing or they have to explain their actual model. Both work in your favor.
Contract Terms Worth Fighting For
Beyond pricing, here's what you actually need in the contract:
1. Know Where Your Data Goes (and How to Get It Back)
Sit with engineering. Understand:
How does data flow? Application → middleware → LLM?
Can you trace it in a standard format like JSON or XML?
What happens to "anonymized" training data if you leave?
Will they give you a deletion certificate?
Negotiate the divorce before the marriage. You want help migrating to a competitor AND proof they destroyed your data.
2. Don't Pay for Their Screw-Ups
AI agents retry until they succeed. Every retry costs you money. Fight for:
Zero cost for vendor technical errors
Automatic credits when error rates hit a threshold
Maximum retry limits before the system stops
Outcome-based pricing where possible—pay for results, not attempts
3. Accuracy Requirements
LLMs hallucinate. That's your liability if it tells a customer the wrong thing. Get:
Monthly accuracy measurements (look up MAPE if you need a metric)
Service credits if it drops below your threshold
Right to terminate if accuracy degrades month over month
No model changes without your approval
4. Who Owns What
Standard vendor language: "We won't give your data to others, but we'll use it to improve the service."
Push for:
You own all inputs and outputs (getting harder to get)
High indemnity caps for IP breaches and data leaks
If you let them train on your data, demand transparency on how
Use your data as leverage: "I'll let you train on my anonymized data IF you give me X"
5. Price Increase Caps
This is critical for 2027-2028. Chip shortages are coming. Vendors will see 40%+ cost increases and pass them through to you.
Cap annual price increases (shoot for single digits)
Require 90+ days notice on any pricing changes
Build in termination rights if increases blow past your cap
What to Do Monday Morning
If You're Just Starting…
Don't panic. Seriously. Nobody has this figured out yet.
Get technical. Book time with your engineering team. Understand how data flows, what tokenization means, how the models work.
Use AI to fight AI. Feed vendor whitepapers and marketing BS into ChatGPT or Claude. Ask it to build cost models and spot red flags.
Build scenarios. Conservative, moderate, aggressive. Account for costs shifting on vendor side.
Make them show their work. "Model what it costs to replace our customer service team with your agents." Force transparency.
If You Already Have AI Contracts
Audit quarterly. QBRs aren't optional in consumption models. Track actual vs forecast constantly.
Control vendor access. Sales teams will evangelize new features directly to your business units to bypass you. Manage this.
Watch for shadow AI. Departments are buying tools you don't know about. Get visibility.
Renegotiate early. Terms from six months ago are already outdated. Market's moving too fast.
When to Walk Away
You need to be willing to walk when:
They won't define pricing after multiple asks
Terms are non-negotiable below enterprise spend
No protection for cost overruns or errors
Business case doesn't justify the risk
Walking away IS leverage. Tell them you need SME review, legal approval, digital risk assessment. Vendors get nervous when they realize it might not close.
One More Thing
This is where we were with SaaS in 2017. Immature market, vendor-favorable terms, deliberately opaque pricing. Took years to mature.
The difference? AI costs can spiral WAY faster than SaaS ever could. That $250k in four months? That doesn't happen with seat licenses.
We're learning this together. Share what's working. Share what isn't. That's the whole point of this community.
Key Quotes from the Session
"They're not gonna give me their cost per token, so I'm building in financial guardrails. That's the hard cap."
"I asked them to model it out for me. It was really hard for me to believe the model was accurate because it did not cost nearly as much as 10 customer service people, which I knew was a lie."
"They cannot or will not answer questions regarding scenarios or how to calculate costs."
"It's deliberately obfuscated. They want uncertainty so they can bill you whatever they feel like."
"Negotiate the divorce before you're even married."
"This isn't adopt or die. We'll still be breathing air if we don't sign by 5pm yesterday."
"We are not running, rocking and rolling with AI right now. Nobody knows what they're doing and everyone's buying it."