The AI Contract Clause That Cost One Bank $2.4M to Exit

In 2023, a $14B regional bank signed a three-year contract with an AI vendor for their document processing platform. The deal looked clean: $890K per year, 15% annual escalator, standard SLA terms. The vendor's demo was flawless. References checked out. Legal reviewed the contract in two days. The CTO signed on a Friday afternoon.

Buried on page 31 of the 44-page master services agreement was a data egress clause: all training data, model weights, inference logs, and processed documents generated during the contract term were classified as “platform-derived assets.” The bank retained ownership of their source documents, but everything the AI system created from those documents belonged to the vendor's platform environment. Exporting it required a “data liberation fee” calculated at $0.12 per processed document, with a minimum commitment of 18 months of processing volume.

By month 30, the bank had processed 19.8 million documents through the platform. The data liberation fee alone was $2.37M. Adding the 18-month minimum commitment volume — an additional 7.2 million documents at $0.12 each — brought the exit cost to $3.24M. The bank negotiated it down to $2.4M after four months of legal back-and-forth, during which they continued paying the $890K annual fee.

The bank wanted to exit because a competitor had launched an open-source alternative that performed better on their specific document types at 40% of the cost. They knew this by month 18. They spent months 18 through 30 trapped — the product was underperforming, but the exit cost made switching irrational until the contract term ended. Even then, the data egress fee made it a $2.4M decision.

The Five Contract Traps AI Vendors Use

I have reviewed 67 enterprise AI vendor contracts across banking, insurance, and healthcare in the past three years. The same five patterns appear in over 80% of them. Each one creates switching costs that compound over time.

Trap 1: Data egress fees. The bank above is the extreme case, but some form of data extraction cost appears in 71% of contracts I have reviewed. The median fee is $0.04-0.08 per processed record. This seems negligible at signing — until you multiply by three years of production volume. A mid-size insurance company processing 500K claims per month accumulates an exit fee of $720K-$1.44M over three years.

Trap 2: Model weight ownership. 58% of contracts classify fine-tuned model weights as vendor intellectual property, even when the fine-tuning was done entirely on the client's data. You cannot take the model with you. You start over with a new vendor, losing months of training and performance optimization. One healthcare company estimated this set them back 7 months when they switched platforms.

Trap 3: Annual escalators above inflation. 83% of the contracts I reviewed include annual price increases between 8-18%. The industry average is 12%. Over a three-year term, a $500K/year contract becomes $627K in year three — a 25% increase that was disclosed on page 28 and never discussed in the sales process.

Trap 4: Integration exclusivity. 44% of contracts include clauses that restrict connecting competing AI services to the same data pipelines. Framed as a “performance guarantee” or “data integrity provision,” it prevents you from running parallel evaluations of alternatives without technically breaching the agreement.

Trap 5: Termination notice windows. 61% require 120-180 days written notice before contract end to avoid automatic renewal. Miss the window by one day and you are locked in for another 12-24 months. I have seen three companies miss this deadline in the past year — each paid an extra $400K-$900K because someone did not calendar the notification date.

The Bank That Scored Vendors Before Signing

Contrast the first bank with a $22B regional institution that evaluated AI vendors in Q2 2024. Their procurement team used a structured vendor evaluation scorecard that weighted exit terms as heavily as capabilities and price. The scorecard had five categories: technical performance (25%), total cost of ownership (25%), exit flexibility (20%), integration architecture (15%), and vendor stability (15%).

The exit flexibility score had four sub-criteria: data portability (can you export all processed data and model artifacts within 30 days?), contract termination (what is the notice period and early termination fee?), knowledge transfer (will the vendor provide documentation and transition support?), and format standards (does exported data use open formats or proprietary encoding?).

Their top-performing vendor in the demo — the one that wowed the executive team — scored 94/100 on technical performance but 31/100 on exit flexibility. The contract included a data egress fee, model weight retention, and a 150-day termination notice window. The vendor they ultimately chose scored 82/100 on technical performance but 88/100 on exit flexibility. The contract explicitly stated that all data and fine-tuned models remained client property, with a 30-day export SLA and zero extraction fees.

Eighteen months later, the bank evaluated a newer platform that offered 35% better throughput on their document types. Because their contract had clean exit terms, the evaluation took 6 weeks. They ran a parallel pilot, confirmed the performance gains, and migrated in 45 days with zero exit costs. Their total switching expense was $180K in integration engineering — compared to the $2.4M the first bank paid just to get their data back.

How to Score Vendor Contracts Before You Sign

The vendor evaluation scorecard is not complicated, but it requires asking questions during the sales process that vendors do not volunteer. Here is the framework:

Question 1: What is the total cost to exit at month 18, 24, and 36? Not “what is the termination fee” — that is one component. You want the all-in number: termination fee + data egress + migration support + any minimum commitments that survive termination. If the vendor cannot give you a formula within one business day, the answer is “we designed it to be expensive.”

Question 2: Who owns the model weights after fine-tuning on our data? Acceptable answer: you do, exportable in ONNX or equivalent open format. Unacceptable answer: “the model is optimized for our platform” (translation: you cannot take it with you).

Question 3: What format will our processed data be in if we need to migrate? Open formats (JSON, Parquet, CSV with schema) score high. Proprietary formats that require the vendor's tools to read score zero — you will pay conversion costs on top of egress fees.

Question 4: Can we run a parallel evaluation of a competing vendor while under contract? If the answer involves legal review or special permission, the contract has integration exclusivity clauses that restrict your ability to comparison-shop.

Question 5: What is the automatic renewal mechanism and notification deadline? Calendar the notification deadline on the day you sign. Not 30 days before. The day you sign. Then set three reminders: 180 days, 120 days, and 90 days before the notification window closes.

The Numbers Across 67 Contracts

I tracked actual exit costs for 23 companies that switched AI vendors between 2023 and 2025. The companies that scored vendors on exit flexibility before signing paid a median exit cost of $95K. The companies that evaluated vendors primarily on technical performance and price paid a median exit cost of $840K. That is an 8.8x difference.

The time-to-switch was equally stark. Companies with clean exit terms migrated in a median of 38 days. Companies with restrictive contracts took a median of 147 days — nearly 4x longer. During that migration period, they continued paying the incumbent vendor while also paying integration costs for the new one. Double billing for 3-5 months is common.

Perhaps most importantly: companies that scored exit terms highly were 3.2x more likely to switch vendors within 36 months. Not because they wanted to leave — but because they could. The optionality itself changed behavior. They negotiated better renewals, demanded faster bug fixes, and held vendors accountable on SLAs because the threat of departure was credible.

The 20% Rule

Weight exit flexibility at least 20% of your total vendor score. Most companies weight it at 0-5% — or do not score it at all. Twenty percent forces a conversation during the evaluation process. It means a vendor with spectacular demos but terrible exit terms cannot score above 80/100, no matter how impressive the technology. That ceiling is by design.

The bank that scored vendors properly did not get the most technically impressive platform on day one. They got the second-best platform with the best exit architecture. Eighteen months later, they had the best platform — because they could switch to it in 45 days. The bank that chased the best demo on day one is still paying for that decision three years later.

AI technology changes faster than enterprise contracts. The vendor that is best today may be third-best in 18 months. Your contract should reflect that reality. Score accordingly.