Stage 0 — The Starting Point: Free Accounts

Every institution's AI journey started the same way — someone signed up for a free ChatGPT account.

• Zero cost, zero control — faculty and students used whatever model OpenAI happened to make available on the free tier
• No data protections — conversations, research queries, and student work flowed into a system with no institutional agreement
• No visibility — the institution had no analytics, no usage data, and no ability to govern what was happening

Stage 1 — Enterprise Agreements

Institutions recognized the problem. We are a university, not an ad farm for an AI lab.

• Enterprise licensing — sign an agreement with OpenAI (or another provider) to carve out data protections and a promise that institutional data won't be used to train models
• The promise model — you're relying on a vendor's contractual commitment, not a technical guarantee. Their promise is your only protection.
• Still single-vendor — you now have a formal relationship, but you're locked into one provider's models, one roadmap, one pricing schedule

Stage 2 — The Multi-Vendor Problem

OpenAI is not the only player. Anthropic, Google, Perplexity — each brings different strengths. But now the math gets ugly.

• $20/user/month × multiple providers — Claude for research, Gemini for search, ChatGPT for general use lands you at $60–80/user/month
• Fragmented experience — students and faculty get sent to different web interfaces with no centralized management, no shared analytics, no unified governance
• Cost comparison — your LMS itself costs ~$15/user/year. Four AI subscriptions cost ~$960/user/year. That's a 64× multiplier.

The Developer Pricing Revelation

The per-seat model is a consumer markup. The actual cost of intelligence underneath is dramatically lower.

• Developer API pricing — the same models that cost $20/user/month are available at ~$0.25 per million tokens (~750,000 words) via API
• 50,000-user example — at $20/user/month you'd spend ~$12M/year on LLM fees. At developer rates, even heavy usage typically costs tens of thousands per year.
• Eight figures vs. five — the gap between consumer pricing and developer pricing is not 2× or 5×. It can be 100× or more.

Stage 3 — LLM Agnosticism

The answer is clear — institutions need a platform layer that brokers access to all models at developer pricing.

• Model-agnostic access — Claude for deep reasoning, Gemini for multimodal, GPT for general Q&A — all through one interface
• Developer-rate economics — pay for actual token consumption, not per-seat licenses. Sam Altman himself has said API pricing converges toward the cost of electricity.
• Unified management — one dashboard for analytics, usage, governance, and cost controls across every model

The Integration Layer

Cost savings alone aren't enough. The platform needs to plug into the systems your institution already uses.

• LTI integration — embed AI tools directly into Brightspace, Canvas, Moodle, or Blackboard so students and faculty never leave the LMS
• Purpose-built features — tools designed for instructors (rubric-aligned feedback, curriculum assistance) and students (tutoring, study planning) — not a generic chat window
• SSO and provisioning — connect to your existing identity systems for seamless access and automatic role-based permissions

Stage 4 — But Who Is the Middle Layer?

You've solved cost and model access. But now there's a new question — who is this company sitting between your users and the language models?

• They hold all your data — the application layer aggregates every conversation, every query, every interaction. They see everything.
• Startup risk — maybe they raised a round. Maybe they're growing fast. But what happens if they pivot, get acquired, or shut down?
• The sad founder email — "We regret to inform you that we are winding down operations." Your entire AI infrastructure disappears overnight.

The Continuity Problem

This is not theoretical. AI startups fail, pivot, and get acquired regularly. Your institution needs AI tools that outlast any single vendor.

• The show must go on — if your AI vendor disappears, your students still need tutoring, your advisors still need analytics, your faculty still need tools
• Vendor lock-in by default — most AI platforms are closed-source SaaS. If the vendor goes away, you don't get to keep anything — not the code, not the data, not the configurations
• Continuity is infrastructure — just as you wouldn't let a single vendor hold the only copy of your LMS database, you shouldn't let a startup hold the only copy of your AI platform

Stage 5 — Code Ownership

The logical conclusion — you need to own the code that brokers your institutional AI interactions.

• Full source code, yours from day one — not open-core with paid features hidden, not a managed service you can't inspect. The complete production codebase running in your infrastructure.
• Deploy on your servers — the platform runs entirely within your data center or cloud tenancy. No external dependencies for core functionality.
• Survival guarantee — if the vendor changes direction, raises prices, or ceases to exist, your platform keeps running. You have the code. You have the data. You have control.

Why Ownership Enables Better AI

Owning the platform isn't only risk mitigation. It unlocks capabilities no external SaaS can match.

• Institutional memory — your AI knows the student's major, their year, their academic history, their goals — because it's connected to your SIS and LMS. The student never re-explains themselves.
• Connected agents — AI that pulls from advising records, financial aid data, course catalogs, and career services — not just answers from training data
• Contextual, not generic — the difference between "Explain photosynthesis" and "Explain photosynthesis in the context of your BIO 201 syllabus, given you struggled with the cellular respiration unit"

Generic AI vs. Institution-Connected AI

A side-by-side comparison of what students and faculty actually experience.

Making a student re-enter context the institution already has signals that the AI isn't truly integrated. It's a toy, not a tool.

Stage 6 — The Full Ownership Stack

From a free ChatGPT account to a fully owned, institutionally integrated AI platform.

• Your infrastructure — the platform runs on servers you control, in a data center or cloud tenancy you manage
• Your code — the complete source code is yours, with a perpetual license. Inspect, modify, audit, extend.
• Your data — every conversation, every interaction, every analytic stays within the institutional boundary. No vendor even has access.
• Any model — OpenAI, Anthropic, Google, Mistral, open-source, or your own fine-tunes. Switch anytime.

But the Story Doesn't End There

Full ownership at API pricing is powerful. But there's still a cost problem — and it's getting bigger.

• High-end models are expensive — Claude Opus, GPT-4.5-class, Gemini Ultra cost $15–75 per million output tokens. Heavy institutional use adds up fast.
• LLM spend can exceed salaries — at scale, some power users generate more in monthly LLM costs than their own compensation. This is a real phenomenon happening right now.
• The paradox — you solved the per-seat markup, but raw API consumption of frontier models at scale is still a significant budget line

The Rise of Private Models

Something has changed in the last 12 months. The small, local models that used to be toys are now genuinely capable.

• Rapid capability gains — Llama 3, Mistral, Phi-3, Qwen, and Gemma can now handle tasks that required GPT-4 just a year ago
• Run on your hardware — these models deploy on standard GPU infrastructure in your data center. No API calls, no per-token charges, no data leaving your premises.
• True zero-trust privacy — when the model runs on your servers, there is no external party to trust or distrust. The data never leaves your network. Period.

Stage 7 — Toward Institution-Owned AI

The complete picture — a hybrid architecture where institutions own everything and choose what goes where.

• Local models for privacy-sensitive tasks — advising, HR queries, protected research, mental health support — all run on-premise with no external calls
• Frontier models for peak performance — when you need the absolute best reasoning (complex research, advanced code generation), route to Claude or GPT via API — still at developer pricing, still through your owned platform
• Intelligent routing — the platform automatically selects the right model for each task based on complexity, privacy requirements, and cost targets

The Journey, Summarized

From zero control to full sovereignty in seven stages.

Each step increases ownership and decreases risk. Most institutions are between stages 1 and 3 today.

What This Means for SUNY

SUNY's scale — 64 campuses, 1.3M+ students — makes this journey uniquely impactful.

• Shared infrastructure, campus autonomy — a SUNY-wide platform with campus-level isolation. Each institution customizes agents, connects their own data, and governs their own AI — while sharing the underlying infrastructure investment.
• Cost at scale — at SUNY's volume, the difference between per-seat licensing and API-rate consumption isn't millions — it's tens of millions per year
• Leading by example — SUNY has an opportunity to define what institution-owned AI looks like for public higher education nationally. Not following a vendor's roadmap — setting the standard.

Live Demo

See an institution-owned AI agent in action.

• Joseph — University AI Assistant — a production agent connected to institutional data, running through an institution-owned platform, at developer-rate API pricing
• What to notice — contextual awareness (knows the student), connected data (pulls from institutional systems), model flexibility (switch models mid-conversation), full analytics (the institution sees everything)
• Try it yourself — agent.ibl.ai

Let's Talk

The path to institution-owned AI is not a five-year plan. The technology exists today.

• Full source code ownership — production-grade platform, yours from day one
• Any model, any infrastructure — cloud, on-premise, hybrid, or fully air-gapped
• Built for higher education — LTI, SIS integration, multi-campus isolation, agent catalogs for every department

Miguel Amigot, CTO · miguel@ibl.ai

ibl.ai — your institution's AI, under your control.