The SaaS AI Frankenstein Problem
When Every Tool Has Its Own Broken Brain
I was in Jira yesterday trying to use Rovo to find a ticket. It gave me three results—none of them relevant. So I opened GitLab, where Duo confidently suggested code that would break our CI pipeline. Then I jumped into Slack, where their AI summarized a critical thread as "general discussion about project updates."
Three different AI assistants. Three different failures. Three separate $20-50/month charges on our invoice.
And I had a revelation: We're living through the Marvel Multiverse problem, but for AI.
The Spider-Man Multiverse Problem: Too Many Spider-Copilots
Remember Spider-Man: No Way Home when three different Spider-Men showed up and spent the first 10 minutes pointing at each other trying to figure out who was the "real" Spider-Man?
Tobey Maguire Spider-Man worked completely differently from Andrew Garfield Spider-Man, who had nothing in common with Tom Holland Spider-Man. Same name. Same costume. Same powers (sort of). But completely different universes, different villains, different web-shooters, and zero ability to coordinate.
That's exactly what's happening with "Copilot."
The Copilot Multiverse (All Real Products, All Called "Copilot"):
- GitHub Copilot (Tobey Maguire): AI code completion ($10/month) - Uses OpenAI Codex/GPT-4. Lives in your IDE.
- Microsoft Copilot (Andrew Garfield): AI for Word/Excel/Outlook ($20-30/month) - NOT the same as GitHub Copilot. Lives in Office apps.
- Microsoft 365 Copilot (Tom Holland): ALSO different from Microsoft Copilot ($30/user/month) - Yes, Microsoft has TWO different products both called "Microsoft Copilot."
- GitLab Duo (Miles Morales): DevSecOps AI suite (bundled with Premium) - NOT called Copilot, NOT related to any Copilot, but does the same job as GitHub Copilot.
- Salesforce Einstein Copilot (Spider-Gwen): CRM AI assistant ($50-125/month) - Different AI entirely, different universe (CRM), happens to also use "Copilot" in the name.
Plot twist: Paying for GitHub Copilot does NOT give you Microsoft Copilot. They're completely different products from the same parent company. Like how Sony owns Tobey/Andrew Spider-Man but Marvel owns Tom Holland Spider-Man.
Now imagine explaining this to your CFO:
CFO: "We're already paying for Microsoft Copilot. Why do we need GitHub Copilot?"
You: "Those are different Copilots."
CFO: "Both from Microsoft?"
You: "Yes. But they don't talk to each other."
CFO: "What about this 'Microsoft 365 Copilot'? Is that the same as Microsoft Copilot?"
You: "No. Different product. Same company. Also called Copilot."
CFO: "And Salesforce Einstein Copilot?"
You: "That's Salesforce. Completely different company. But also called Copilot. Does CRM stuff."
CFO: "Can they work together?"
You: "No. Different universes."
CFO: "This is insane."
You: "Welcome to 2025."
Just like in No Way Home, all these Spider-Copilots are pointing at each other, claiming to be the real AI assistant, but none of them can actually coordinate.
And unlike the movie where the three Spider-Men eventually team up to fight villains, these Copilots will never work together. Each one lives in its own universe (GitHub, Office, CRM, DevOps), has its own powers (code completion, document editing, ticket management), and charges separately for web-slinging.
This is the first symptom of the SaaS AI Frankenstein problem: We have a multiverse of AI assistants with the same name doing completely different things, and nobody—not even the companies selling them—can explain which Spider-Man you actually need.
The Cable TV Bundle Death Spiral
Remember cable TV? You wanted HBO and ESPN, but had to buy the "Gold Premium Ultra Package" that included 200 channels you'd never watch. Then they'd raise prices annually while the content got worse.
Welcome to SaaS AI in 2025.
The Bundling Bait-and-Switch Timeline:
Slack AI
2024: $10/user/month add-on
August 2025: Add-on discontinued, now bundled into Business+ plan (20% price increase)
Result: Everyone pays for AI whether they use it or not
Notion AI
2024: $8-10/month add-on
May 2025: Bundled into Business plan ($20/user/month required)
Result: Free/Plus users get 20 AI responses, then paywall
Atlassian Rovo
October 2024: $24/user/month
Late 2024: Free with Premium/Enterprise
2025: Coming to Standard tier
Result: Pricing model changed twice in 3 months
GitLab Duo
2024: Bundled into Premium/Ultimate at no extra cost
Result: The only vendor not playing pricing games (yet)
Here's the pattern: Launch AI as an add-on to test the market. Realize customers won't pay extra. Fold it into higher-tier plans. Raise base prices. Claim it's "included at no extra cost."
Translation: You're paying for it. You just don't get to opt out anymore.
The Blockbuster Late Fees Problem: Token Economics
Blockbuster didn't die because Netflix had better movies. It died because people hated late fees—those surprise charges that made renting a $3 movie cost $15.
SaaS vendors are speedrunning the same mistake with AI token costs.
The Token Cost Nightmare:
Salesforce Einstein
• Base: $50/user/month ("limited GPT credits"—definition unclear)
• Agentforce: $125/user/month for "unlimited" usage
• Total cost: Base CRM ($200-250/month) + Einstein = $500+/month per sales professional
ServiceNow Now Assist
• Pay-as-you-go consumption pricing (2025)
• Calls exceeding 1,000 tokens consume extra "assists"
• 3,000 tokens = 3 assists consumed
• Problem: Impossible to forecast monthly costs
HubSpot Breeze
• Credit-based system (changed June 2025)
• Each credit now $0.01 (down from $0.08)
• AI actions consume 8x more credits to compensate
• Additional credits: $10 per 1,000
• Problem: Credit complexity makes budgeting impossible
Monday.com AI
• 500 free credits/month (Standard+)
• Additional credits: $200/month ($2,400/year)
• Basic plan users can't purchase additional credits at all
Zendesk AI
• Base: $115/month/agent (Professional plan required)
• Advanced AI: +$50/agent/month
• Usage fees: $1.50-2.00 per automated resolution
• Triple billing: License + AI fee + per-use fee
Industry Reality Check:
- • Only 15% of companies can forecast AI costs within ±10%
- • 1 in 4 companies miss forecasts by more than 50%
- • Minor prompt changes can spike costs 100x overnight
- • Real costs frequently exceed estimates by 30-50%
This is Blockbuster late fees on steroids. Except instead of "you forgot to return the DVD," it's "your AI assistant used 10% more tokens than expected this month."
The Restaurant Impossible Problem: Everyone's a Chef
Remember Kitchen Nightmares? Gordon Ramsay would walk into failing restaurants where every cook had their own "secret recipe" and nothing worked together. The menu was 47 pages long, the food was mediocre, and customers were confused.
That's your SaaS stack right now.
Problem: Only knows Jira data. Can't see GitLab, GitHub, or Google Docs
Problem: Doesn't integrate with Jira, your actual project management tool
Problem: "Walled garden"—can't access Confluence, Google Docs, or external knowledge bases
Problem: Free plan only keeps 90 days of history; older messages permanently deleted after 1 year
Problem: Unreliable with database queries; struggles with large datasets
Problem: Knowledge base updates take 1+ week to sync to AI; can't test on historical data
See the problem? Every tool has its own AI brain. None of them talk to each other. All of them are half-baked.
It's like having seven chefs who each only know how to make one dish, refuse to share recipes, and all charge separately for ingredients.
Why Is This Happening? (The Uncomfortable Economics)
Let's be honest about why SaaS vendors are shipping AI that feels like a beta test stapled onto production software.
1. They Can't Afford to Train LLMs on Their Own Data
Training a custom LLM costs millions. Fine-tuning a foundation model on your SaaS platform's specific workflows, features, and edge cases? Still hundreds of thousands of dollars.
So what do vendors do instead?
They bolt OpenAI's GPT-4, Anthropic's Claude, or Google's Gemini onto their platform via API and hope the general-purpose model "figures it out."
Real example: Jira Rovo uses foundation models but hasn't deeply trained them on Jira's specific query language, custom fields, or workflow patterns.
Result? It can summarize a ticket, but ask it "Find all P1 bugs assigned to Frontend team in the last sprint" and it hallucinates results.
Source: The average cost to integrate ChatGPT into a SaaS product ranges from $40,000-$250,000 for implementation alone, plus $0.002 per 1,000 tokens ongoing. Custom training? 10-100x more expensive.
2. Bring-Your-Own-LLM Is a Technical Nightmare
Imagine if you could connect your own ChatGPT Pro account or Claude subscription directly to Jira, Salesforce, or GitLab.
You'd get better AI, more tokens, and full control. Perfect, right?
Except nobody's figured out how to make this work at scale.
The Real Problems with External LLM Integration:
1. MCP Servers Are Still Experimental
Model Context Protocol (MCP) was supposed to solve this—let your external AI talk to any SaaS tool. But MCP servers are in flux, implementations are inconsistent, and they burn through tokens like crazy because every interaction requires full context re-transmission. You can easily hit your Claude Pro token limit in an afternoon of "connected" work.
2. Enterprise Security Is Real (Not Theater)
Legal and compliance teams have legitimate concerns: GDPR, HIPAA, SOC 2, data residency requirements. Sending customer PII, trade secrets, or proprietary data to external LLM providers—even with encryption—violates most enterprise security policies. It's not paranoia; it's liability.
3. API Rate Limits and Coordination Chaos
Even if you could connect your Claude Pro account to Salesforce, what happens when 500 employees try to use it simultaneously? OpenAI and Anthropic don't have enterprise-grade API coordination for personal accounts. You'd hit rate limits constantly, and troubleshooting "why did my AI stop working?" becomes a support nightmare.
4. The "Who Pays?" Problem
If your external LLM integrates with your SaaS tools, who pays for overage? If a buggy automation loop burns through $10,000 in tokens overnight, is that on you or the SaaS vendor? Nobody wants that liability, so vendors bundle AI and eat the cost variability themselves (which is why they're adding token limits).
The honest truth: BYO-LLM sounds great in theory, but the infrastructure, security model, and economics don't exist yet to make it work reliably for enterprise customers.
So SaaS vendors default to embedded AI—not because they're trying to lock you in (though that's a nice side effect), but because it's the only model that actually ships and doesn't create massive support headaches.
3. The Market Is Too Squeezed to Add AI Properly
SaaS companies are stuck in a bind:
- • Customers expect AI features (because competitors have them)
- • Customers won't pay significantly more for AI
- • AI integration costs are high and margins are thin
- • Pricing experiments (add-ons, bundles, credits) confuse customers and erode trust
The Margin Compression Crisis:
- • Traditional SaaS gross margins: 70-90%
- • AI-centric SaaS margins: 50-60% (structurally lower)
- • 84% of companies report AI cutting gross margins by more than 6%
- • Early-stage AI companies: Margins down nearly 10 percentage points YoY
- • Example: Replit's gross margin briefly went negative during 2024 usage surge
Source: Industry reports from Zylo, Drivetrain.ai, and SaaS CFO research (2025)
Translation: AI makes SaaS less profitable. Vendors can't afford to do it right, but they can't afford not to do it.
So they ship half-finished AI, bundle it into higher tiers, and hope customers don't notice it's mediocre.
4. Token Costs Are Killing Vendor Economics
Every time you use embedded AI in a SaaS tool, the vendor pays for those tokens.
And unlike traditional SaaS (where serving one user costs the same as serving 10,000), AI costs scale linearly with usage.
Real Token Economics:
- • Average AI spending increased from $62,964/month (2024) to $85,521/month (2025)—a 36% increase
- • Organizations spending $100K+/month on AI doubled from 20% to 45% in one year
- • Moderate deployments (5-10M tokens/month) cost $1,000-$5,000/month
- • Example: Mid-sized e-commerce brand enabled order-tracking AI workflow → token usage spiked 300% → monthly costs jumped from $1,200 to $4,800
Source: Zylo AI Cost Governance Report 2025
Now multiply that across every customer using your embedded AI. SaaS vendors are hemorrhaging money on token costs.
That's why they're implementing:
- • Token limits (HubSpot, Monday.com credit systems)
- • Usage caps (Notion's 20 AI responses for Free/Plus users)
- • Pay-per-use fees (ServiceNow, Zendesk)
- • Forced bundling (Slack, Atlassian, GitLab—spread costs across all users)
They're not trying to rip you off. They're trying not to go bankrupt.
5. The Market Is in Chaos (And Nobody Knows What's Next)
Remember the Cargo Cult AI problem? Companies copying visible rituals without understanding the underlying system?
That's the entire SaaS AI market right now.
The Pricing Chaos:
- • 73% of AI companies still experimenting with pricing models
- • Average company testing 3.2 different approaches in first 18 months
- • Hybrid pricing (subscription + usage/value-based) surged from 27% to 41% in 12 months
- • Seat-based pricing dropped from 21% to 15%
- • 60% of vendors deliberately mask rising prices by bundling AI features
Sources: McKinsey SaaS AI Study, Valueships AI Pricing Trends 2025
Translation: Nobody knows how to price this. Everyone's guessing.
Atlassian changed Rovo pricing twice in three months. Slack killed their AI add-on and bundled it. Notion did the same. ServiceNow went pay-as-you-go. Monday.com uses credits. Zendesk uses triple billing.
There is no strategy. Just panic.
The "SaaS Is Dead" Elephant in the Room
In December 2024, Microsoft CEO Satya Nadella said something that sent shockwaves through the industry:
"SaaS is dead."
Microsoft predicts AI business agents will replace traditional SaaS by 2030. Y Combinator says vertical AI agents could be 10x bigger than SaaS. Gartner forecasts AI agents will make 15% of daily business decisions by 2028 (up from 0% in 2024).
Sources: Microsoft BG2 podcast (Dec 2024), Y Combinator market analysis, Gartner AI predictions
Here's the theory: Why pay for Jira, Salesforce, GitLab, Slack, Notion, and seven embedded AIs that don't talk to each other...
...when you could have one AI agent that orchestrates all of them?
Instead of logging into eight different tools with eight different AI assistants, you'd say:
"Show me all P1 bugs from last sprint, the related GitLab merge requests, Slack conversations about the fix, and update the Jira tickets with a summary."
And an AI agent would do it. No switching tools. No eight different AI brains. Just one intelligent orchestration layer.
That's the future Microsoft, Y Combinator, and half of Silicon Valley are betting on.
And if they're right, every SaaS vendor embedding half-baked AI into their product is building the wrong thing entirely.
The COVID-Level Uncertainty Problem
Remember early 2020? Businesses had no idea what was coming. Do we go remote? Do we invest in digital tools? Do we cut costs? Do we double down on growth?
That's where SaaS buyers are right now with AI.
Option 1: Buy embedded AI from every SaaS vendor
Risk: Pay 30-110% more for mediocre AI that doesn't integrate. Token costs spiral. Vendor lock-in intensifies.
Option 2: Wait for AI agents to replace SaaS
Risk: Fall behind competitors who are using AI now. Miss productivity gains. Lose talent to AI-forward companies.
Option 3: Build your own AI layer on top of existing SaaS
Risk: Data security concerns. API limitations. Vendor ToS violations. Integration maintenance hell.
Option 4: Do nothing and wait for clarity
Risk: Competitors gain 2-3 year AI maturity advantage. Your team falls behind on AI literacy. Harder to catch up later.
There is no clear answer. That's the problem.
Organizations are paralyzed by uncertainty. SaaS vendors are experimenting with pricing. Microsoft is declaring the death of the industry. Y Combinator is funding AI agent startups to kill SaaS.
And everyone's flying blind.
What Actually Works Right Now
Alright, enough doom. Let's talk solutions.
Because while the SaaS AI landscape is a mess, there are patterns that work.
1. Use Embedded AI for Narrow, High-Volume Tasks
Embedded SaaS AI is terrible at understanding your business. But it's pretty good at repetitive grunt work within that tool.
Good Use Cases:
- • Slack AI: Summarizing long threads you missed (not strategic decisions)
- • Notion AI: First-draft blog posts or meeting notes (heavy editing required)
- • Zendesk AI: Triaging simple support tickets (humans handle complex issues)
- • Salesforce Einstein: Lead scoring based on historical data (not customer conversations)
Rule: If the AI only needs to know about one tool's data and the task is repetitive, embedded AI can work.
2. Don't Pay for AI You're Not Using
Sounds obvious, but here's the trap: Vendors are bundling AI into higher-tier plans and raising prices.
Audit your SaaS AI spend quarterly:
- • Which AI features are actually being used?
- • What's the measured ROI? (Time saved, tickets resolved, errors reduced)
- • Are you hitting token/credit limits? (If not, you're overpaying)
- • Could you downgrade to a lower tier and lose nothing?
If your team isn't using Slack AI, Notion AI, or Rovo, don't pay for the bundle.
3. Invest in One External AI Tool You Control
Here's the dirty secret: ChatGPT Pro ($20/month) or Claude Pro ($20/month) is better than every embedded SaaS AI combined.
Why?
- • Unlimited usage (no token limits, credit systems, or overage fees)
- • Works across all your tools (just copy/paste context)
- • You control the prompts, context, and workflows
- • Latest models (not 6-month-old versions vendors license cheaply)
Instead of paying $50/month for Salesforce Einstein that only knows CRM data, pay $20/month for Claude that can analyze CRM exports, Jira tickets, GitLab issues, and Slack threads.
One powerful external AI > Seven mediocre embedded AIs.
4. Prepare for the Agentic Future (But Don't Wait for It)
AI agents that orchestrate multiple SaaS tools are coming. But they're not here yet (despite the hype).
What you can do now:
- • Build AI literacy in your team (they need to know how to work with AI, not just around it)
- • Document your workflows (AI agents will need clean, documented processes to automate)
- • Reduce tool sprawl (fewer tools = easier orchestration later)
- • Invest in APIs and integrations (agentic systems need accessible data)
Think of 2025-2026 as the "AI literacy" phase. 2027-2028 will be the "AI orchestration" phase.
Don't bet the farm on agents replacing SaaS tomorrow. But start preparing for the transition.
The Bottom Line: Frankenstein's Monster Is Alive
SaaS vendors stitched AI onto their products like Dr. Frankenstein assembling body parts. And just like the novel, the creation is alive, but it's not what anyone hoped for.
The SaaS AI Reality Check:
- ✗ Embedded AI is mediocre because vendors can't afford to train custom models
- ✗ Data security concerns prevent you from using external LLMs
- ✗ Token economics are destroying SaaS margins
- ✗ Pricing is chaos (bundling, credits, pay-per-use, add-ons—nobody knows what works)
- ✗ The market is paralyzed by uncertainty (is SaaS dying? Are agents the future? Should we wait?)
- ✗ Every tool has its own AI brain, and none of them integrate
But here's the thing: You don't have to accept this Frankenstein's monster.
Use embedded AI for narrow, high-volume tasks where it actually works. Don't pay for bundled AI you're not using. Invest in one external AI you control. Build AI literacy in your team. Prepare for orchestration, but don't wait for perfection.
The SaaS AI revolution isn't here yet. We're in the awkward middle phase where everything is broken but nothing is finished.
And just like Frankenstein's monster, the creation may be ugly, but it's not going back in the lab.
Time to figure out how to live with it.
Drowning in SaaS AI Chaos?
We help businesses cut through the hype, audit AI spend, and build strategies that work in the real world—not the demo. No cargo cult rituals, no Frankenstein solutions. Just honest assessment and measurable results.
Let's Fix Your AI Stack