Building an exit strategy for every AI vendor

In late 2022, a large U.S. property-and-casualty insurer signed a three-year enterprise agreement with a major AI platform vendor, embedding that vendor's APIs across twelve internal underwriting and claims-triage workflows. By mid-2024, a competitor had released a model with measurably better performance on insurance-domain benchmarks, and the insurer's head of AI infrastructure wanted to migrate. The migration did not happen on schedule. Data-egress clauses, proprietary fine-tuning formats, and a vector database schema tied to the vendor's own retrieval layer created an exit cost the procurement team had never modeled. Internal estimates — shared with Xither under anonymity — put the unplanned cost of migration at $4.2 million, roughly 60% of the original contract value, consumed before a single production workflow had switched vendors.

That insurer's situation is not an outlier. A 2024 Gartner survey found that 58% of enterprises reported unexpected switching costs when attempting to change AI platform vendors within 24 months of initial deployment. What the insurer discovered — and what two other enterprise teams subsequently turned into structured practice — is that AI vendor exit strategy is a procurement engineering problem, not a legal one. The legal team can negotiate better termination clauses; only the platform engineering team can architect portability.

How the insurer audited its own lock-in

The first concrete action the insurer's AI infrastructure lead took was a lock-in surface audit — a structured inventory of every technical dependency on the incumbent vendor. The audit ran for six weeks across three work streams: data residency and egress, model artifact portability, and integration coupling. What it found was a layered dependency stack that had accumulated unintentionally. Fine-tuned model weights were stored in the vendor's proprietary checkpoint format, incompatible with Hugging Face's safetensors or ONNX without a conversion pipeline that did not yet exist internally. Embeddings powering twelve RAG pipelines were generated by the vendor's proprietary embedding model, meaning the vector store would have to be rebuilt from scratch on any new platform.

The audit produced a dependency score for each workflow: a simple 0–10 integer counting how many of ten identified lock-in vectors applied. Workflows scoring 7 or above were classified as migration-critical and placed in a remediation queue. Of the twelve workflows audited, nine scored 6 or above. The insurer had, without intending to, built a system that was near-maximally coupled to a single vendor across every technical dimension that mattered for portability.

Three teams that built portability before they needed it

The insurer's audit circulated — anonymized — through a peer CIO network. Two other enterprises, a global logistics firm and a mid-market financial services company, used it as a template to run their own audits before committing to new vendor contracts. A third team, the AI platform group at a large regional health system, had independently arrived at similar practices through a HIPAA-driven data-residency review. Across all three, a common architecture emerged: abstraction-first deployment, where every vendor integration is mediated by an internal interface layer that the team owns.

1. 1

   Q1 2023 — Logistics firm lock-in audit

   4 weeks

   Platform engineering team ran a pre-contract lock-in surface audit modeled on the insurer's methodology. Identified 6 of 10 lock-in vectors in the proposed vendor agreement before signing.

   ›
2. 2

   Q2 2023 — Abstraction layer design

   6 weeks

   Logistics team built an internal LLM gateway (OpenAI-compatible REST interface over LiteLLM) routing traffic to the primary vendor while supporting a secondary vendor as a hot standby. All fine-tuning artifacts were stored in safetensors format in a vendor-neutral S3-compatible store.

   ›
3. 3

   Q3 2023 — Financial services firm replicates pattern

   3 weeks

   FinServ team adopted the logistics firm's gateway pattern, adding a contract clause requiring the vendor to provide a 90-day data-export window with no egress fees upon contract termination.

   ›
4. 4

   Q4 2023 — Health system portability review

   8 weeks

   Health system's HIPAA compliance review independently surfaced 7 of 10 lock-in vectors. Team standardized on open embedding models (BGE-M3 via Hugging Face) for all RAG pipelines, eliminating vendor-native embedding dependency.

   ›
5. 5

   Q1 2024 — Logistics firm executes partial migration

   11 days

   When a preferred vendor released a model with superior routing performance, the logistics team migrated two of five workflows to the new vendor in 11 days. The gateway layer required no application-layer changes.

   ›
6. 6

   Q2 2024 — Insurer completes remediation

   5 months

   After the $4.2M exit-cost event, the insurer rebuilt all twelve workflows behind an abstraction layer. Estimated cost of remediation: $1.1M. Projected future migration cost: under $200K per workflow swap.

What portability architecture actually looks like

The logistics firm's implementation is the most documented of the three cases and serves as the clearest reference. Their platform engineering lead, drawing on the LiteLLM proxy project (v1.28 at time of implementation), built an internal gateway that presented a single OpenAI-compatible `/v1/chat/completions` endpoint to all internal consumers. Routing rules, rate limits, cost allocation tags, and fallback logic lived in the gateway's configuration layer — not in application code. When the team wanted to test a new model, they changed a routing rule. No application team wrote new integration code.

Embedding portability was the most technically consequential decision the three teams made. The health system's choice of BGE-M3 — an open-weights model available on Hugging Face under the MIT license — meant their vector stores could be rebuilt on any infrastructure that could run inference, without licensing negotiation or vendor approval. The logistics firm made the same choice for new pipelines while accepting that their legacy vendor-native embeddings would remain a stranded cost in the short term. The financial services firm, starting from scratch, committed to open embedding models for all net-new RAG pipelines as a procurement policy.

Fine-tuning artifact portability required a different intervention: format standardization at the point of creation. The insurer's proprietary checkpoint format was the direct cause of roughly $900K of its exit cost — a conversion pipeline had to be built, validated, and tested against production behavior before migration could proceed. All three forward-planning teams adopted safetensors as their canonical weight format and used Hugging Face Hub (either the public service or a self-hosted instance via Hugging Face Enterprise Hub) as their artifact registry. Neither vendor approval nor vendor tooling was required for export.

Contract mechanics that enabled the architecture

Architecture alone was insufficient. The financial services firm learned that even with a portable technical stack, contract terms could still impede migration. Their legal team, working from a checklist surfaced by the platform engineering group, negotiated three specific clauses into their enterprise agreement: a zero-egress-fee data export window of 90 days post-termination; a vendor obligation to maintain a current data export API throughout the contract term; and a prohibition on format changes to exported artifacts without 180 days' notice. None of these clauses required the vendor to offer lower pricing — they were framed as data-stewardship terms, not adversarial exit provisions, and all three were accepted in the first negotiation round.

The insurer, in its post-migration retrospective, identified that a single contract clause — a data-egress fee of $0.12 per GB with no cap — had generated $610K of its $4.2M exit cost. That clause had passed legal review without flag because no one on the legal team had modeled the data volumes the AI system would accumulate over 18 months of production operation. The fix the insurer implemented for future contracts was a volume-projected egress cost estimate, produced by the platform engineering team and attached to every enterprise AI vendor agreement as an exhibit before signature.

Outcomes and what they cost to achieve

Across the three forward-planning teams, the total investment in portability architecture — gateway build, embedding model selection and validation, artifact registry setup, and contract negotiation support — ranged from $180K to $340K per team, depending on organizational size and existing infrastructure. Against the insurer's $4.2M exit cost, the return on that investment is straightforward to model. But the more operationally significant outcome was speed: the logistics firm's 11-day workflow migration was not a cost story, it was a competitive agility story. The team could respond to model-quality improvements in the market without a multi-quarter migration program.

The health system's outcome was different in character. Its portability work was driven initially by HIPAA compliance, and the migration agility benefit was secondary. What it gained most directly was the ability to demonstrate to its compliance team that no patient data was permanently resident in a vendor-controlled system — a requirement that, as of mid-2024, several state-level health data regulators were beginning to codify explicitly. Portability architecture, in that context, is also a regulatory risk management tool.