Lock-in is not a feeling, it is a measurement. By the time a procurement team says “we are locked into this AI vendor,” the conversation is already off the rails; half the room is talking about contractual lock-in, the other half is talking about integration depth, and nobody is naming the eight distinct axes along which lock-in accumulates. The result is a debate that resolves to whichever axis the loudest voice picks, not to a decision the organization can defend at a board level. The fix is to stop treating lock-in as a single variable and start treating it as a portfolio of eight scored axes that produce a single composite score the procurement, engineering, and finance functions can argue from. This piece names the eight axes, gives a 0-to-3 scoring guide for each, and shows the composite that turns the abstract worry into a deliberate audit you can run most quarter. Run it once and you will be surprised by which vendor is locking you in. Run it most quarter and you will start making sourcing decisions that compound rather than corner you.
It sits inside the framework laid out in the AI build-vs-buy-vs-hire decision matrix for 2026. The matrix’s seventh principle holds that most sourcing decisions get re-litigated quarterly; the lock-in audit is the artifact that makes that re-litigation honest.
Why an audit, not a feeling
Most AI procurement teams in 2026 talk about lock-in the same way they talked about cloud lock-in in 2018; as an anxiety, not a measurement. An anxiety produces a vague preference for “open” alternatives that often costs more than any actual lock-in would have. A measurement produces a defensible decision: this vendor scores 17 out of 24, that one scores 9, and the gap is worth $X in switching insurance.
Lock-in resists single-variable thinking because it accumulates along several independent axes that interact non-linearly. A vendor with airtight contract terms but trivial data egress is far less locked-in than a vendor with permissive contracts but proprietary embedding models. A vendor whose weights you cannot inspect but whose API you can swap behind a thin adapter is far less locked-in than a vendor whose weights are public but whose orchestration would require rewriting half your application code.
The audit is the discipline of looking at most axis separately, scoring each on a consistent scale, and only then collapsing to a composite. The composite is for the board; the per-axis scores are for the engineering team.
The eight axes
The eight axes that constitute meaningful AI vendor lock-in in 2026 are:
- Data egress; can you get your data out and at what cost
- Model weights; do you have access to the actual model artifacts
- Prompts; are your prompts in your repo or the vendor’s UI
- Eval set; does your eval set live in your infrastructure
- Contract terms; what do the legal documents say about exit
- Integration depth; how deep does the vendor sit in your code path
- On-call ownership; when it breaks, who carries the pager
- Rollback feasibility; can you undo the deployment
Each axis is scored 0 to 3, where 0 is no lock-in (full portability) and 3 is severe lock-in (effectively trapped). The composite ranges from 0 to 24. Anything above 16 is a board-level escalation; anything below 8 is healthy; the middle range is where most engagements sit and where the per-axis detail is most useful.
Axis 1: data egress
Data egress measures how easily and cheaply you can extract most byte the vendor holds: raw inputs, outputs, embeddings, fine-tuning datasets, logs.
Score 0: full self-serve export in an open format (Parquet, JSONL, CSV), no fees, covers many categories.
Score 1: full export available but requires a support ticket, takes 1-7 days, no additional fees.
Score 2: partial export; some categories (typically logs or embeddings) are missing, or the format is proprietary. Fees may apply for large volumes.
Score 3: no meaningful export. Data is stranded; switching means reconstructing your data history from upstream systems most organizations do not have for AI workloads.
The most under-tested aspect is embeddings. Many vendors export inputs and outputs but not embeddings, on the grounds that embeddings are derivative work product. If your retrieval system depends on those embeddings, you are at score 2 or 3 even when the headline export looks complete.
Axis 2: model weights
Model weights measure whether you have access to the actual model artifacts.
Score 0: weights are open, downloadable, and you have a working deployment in your own environment.
Score 1: weights are open or open-weight but you do not currently run them. The option exists; you have not exercised it.
Score 2: weights are proprietary but the vendor offers weight escrow or licensing that triggers on specified conditions.
Score 3: weights are proprietary, no escrow exists, contract forbids reverse engineering. Your only path off the vendor is a different model entirely.
The detail on weights escrow is in the AI escrow strategy; escrow is the cheapest mechanism to drop axis 2 from a 3 to a 2.
Axis 3: prompts
Prompts are the second-most valuable AI asset (after evals); their portability matters.
Score 0: prompts live in version-controlled repos in vendor-neutral format. Retargeting takes a thin adapter.
Score 1: prompts live in your repo but use vendor-specific syntax (function-calling formats, role tokens) requiring a non-trivial port.
Score 2: prompts live in the vendor’s UI with limited export. Re-creation in your repo is achievable but lossy.
Score 3: prompts live exclusively in the vendor’s UI with no export. Switching means rewriting most prompt from scratch.
The 2024-era pattern of “many our prompts are in the vendor playground because the eval tools are good there” produces a lot of axis 3 = 3 today. Repatriating prompts to version-controlled repositories is one of the highest-ROI lock-in reductions in 2026.
Axis 4: eval set
Evals encode what good means for your use case; if they are stranded, you have lost the ability to compare alternatives.
Score 0: eval set in your repo, vendor-neutral format. Eval runner targets any vendor; results in your warehouse.
Score 1: eval set in your repo but the runner is vendor-specific. Porting requires rewriting the runner; test cases survive.
Score 2: eval set in the vendor’s product. Test cases exportable; scoring functions and historical results are not.
Score 3: eval set exclusively in the vendor’s product. Switching means rebuilding from scratch; a 3-to-6 month effort.
The matrix’s fifth principle is that eval infrastructure should rarely be bought; an axis 4 score of 2 or 3 violates that principle and should be remediated regardless of the rest of the audit.
Axis 5: contract terms
Contract terms measure what the legal documents say about exit. Legal is the primary scorer here.
Score 0: no-cause termination with reasonable notice (30-90 days), data return obligations specified, pricing transparency, and unambiguous IP ownership of contributed materials (prompts, fine-tuning data, evals).
Score 1: reasonable termination rights but silent on data return mechanics or IP of contributed materials.
Score 2: restrictive termination (long notice, termination-for-convenience fees) or ambiguous IP language.
Score 3: multi-year minimum-commits, auto-renewal with short opt-out windows, exclusivity, or IP assignment of contributed materials to the vendor.
The under-noticed detail in 2026 is auto-renewal. Many contracts written in 2023-2024 have 12-month auto-renewals with 60-day opt-outs; miss the window and you are committed for another year regardless of usage changes.
Axis 6: integration depth
Integration depth measures how deeply the vendor sits in your code path; how much application value depends on vendor-specific behavior vs. A generic capability the vendor happens to provide.
Score 0: vendor sits behind a thin adapter (under 200 lines) with a vendor-neutral interface. Swapping is a config change and adapter rewrite.
Score 1: adapter exists but several application-layer features depend on vendor-specific behavior (output formats, latency profiles, function-calling).
Score 2: vendor SDK conventions leak through the adapter and influence application structure. Swapping is a multi-module refactor.
Score 3: no adapter. Vendor API calls are inline; application design assumes vendor-specific behavior. Swapping is a rewrite.
Axis 6 is the most often underestimated. Teams who say “we have an adapter” frequently have one that handles the request/response shape but not the behavior; latency, parsed format, failure modes. A behavior-passing-through adapter is a partial solution.
Axis 7: on-call ownership
On-call ownership measures who carries the pager at 3am. A vendor whose product breaks the application but whose support is unreachable creates operational lock-in to whatever workarounds the team builds.
Score 0: 24/7 support, guaranteed SLA, status page, post-incident reports, escalation to a human at the vendor.
Score 1: business-hours support, reasonable response times, status page, post-incident reports. After-hours your team works around issues.
Score 2: asynchronous support (email/ticket) with multi-day response. No reliable status page. Your team carries most operational burden.
Score 3: support non-existent or community-forum only. Your team is the de facto support team. Switching is the only path to operational sanity.
Axis 8: rollback feasibility
AI features have a peculiar rollback problem: model behavior is part of the deployment, and may not be reproducible if the vendor changes the model under you.
Score 0: explicit model versioning, long deprecation windows (12+ months), full reproducibility, pinning supported. Rollback is a config change.
Score 1: versioning with shorter deprecation (6 months) or in-version behavior changes (silent updates).
Score 2: versioning in name only; rapid deprecation, silent behavior changes, best-effort reproducibility.
Score 3: no versioning. Vendor changes the model whenever. Rollback impossible; last quarter’s evals cannot be reproduced.
See why AI inference cost is the new database cost line; an axis 8 score of 3 means you have no defense when the vendor’s pricing changes correlate with model changes that invalidate existing evals.
Composite scoring
The composite score is the sum of the eight axis scores: 0-24.
0-7: healthy. Maintain quarterly audits to catch drift.
8-15: middling. Most organizations live here. The work is to drive the highest-scoring axes down through targeted contractual or architectural changes.
16-20: concerning. Findings are board-relevant. Plan a 12-to-18-month portability program with named owners, axes, and milestones.
21-24: severe. The vendor is trapping the organization. The output is a switching plan, not a remediation plan.
Run the audit at engagement start, at most QBR, and at most contract renewal. The cadence matters more than the precision of any single scoring.
Frequently asked questions
Why eight axes, not five or twelve?
Eight is the smallest set that covers the empirically distinct lock-in modes we have seen produce surprises in audits. Fewer axes consistently miss surprises; more axes produce diminishing returns and audit fatigue.
Which axes are typically worst in 2026?
Axes 4 (eval set) and 8 (rollback feasibility) are the most often at 3. Eval-in-vendor-product was a defensible 2023 choice that has aged badly; rollback feasibility is structurally hard because most vendors have not committed to long-term model versioning.
How long does an audit take?
A first audit takes 4-8 hours per major vendor with engineering, legal, and procurement participation. Subsequent quarterly audits take 1-2 hours per vendor.
Who runs the audit?
Procurement owns the calendar and the composite. Engineering owns axes 1, 3, 4, 6, and 8. Legal owns axes 2 and 5. Operations owns axis 7.
Does the audit apply to internal AI tooling vendors too?
Yes. Internal tooling vendors (eval platforms, observability platforms, prompt management platforms) frequently score worse on axes 3 and 4 than the foundation model vendors do. Audit them too.
What about open-source AI?
Open-source AI scores 0 on axis 2 by definition but can still score 2 or 3 on axes 6 and 8 if the integration is brittle or if the version you use is not actively maintained. The audit applies regardless of license.
How does this relate to the consolidation play?
The AI vendor consolidation play trades best-in-class capability for fewer vendor relationships. The lock-in audit is the artifact that determines whether the consolidation has produced a lock-in problem worse than the fragmentation it solved.
What is the cheapest way to reduce a 3 to a 2?
Axis-dependent. For axis 2, weights escrow; for axis 3, prompt repatriation to repo; for axis 4, eval set repatriation; for axis 5, contract amendment; for axis 6, adapter discipline; for axis 8, model version pinning where the vendor offers it. The remediation cost is small per axis; the cumulative effect is large.
Should the audit be public to the vendor?
Selectively. Sharing axes 5 and 6 with the vendor is a reasonable negotiation lever. Sharing axes 4 and 8 may signal weakness. Use judgment.
How does this connect to the build-vs-buy-vs-hire matrix?
The matrix decides what to source from where; the audit measures the consequences of each sourcing choice on portability. Together they keep the organization from making sourcing decisions whose consequences only surface at exit.
Key takeaways
Lock-in is not one variable, it is eight: data egress, model weights, prompts, eval set, contract terms, integration depth, on-call ownership, and rollback feasibility. Each is scored 0-3; the composite ranges 0-24. Below 8 is healthy; 8-15 is middling; 16-20 is board-relevant; 21-24 is severe.
The audit’s value is the cadence. Run it at engagement start, at most QBR, and at most renewal. The per-axis scores produce remediation actions; the composite produces escalation discipline. Both audiences; the board and the engineering team; need both views.
The most under-treated axes in 2026 are eval set (axis 4) and rollback feasibility (axis 8). Eval-in-vendor-product was a 2023 default that has aged badly; model versioning has not matured industry-wide. Both axes deserve remediation attention disproportionate to their composite contribution.
The audit is the cheapest exposure of mispriced sourcing decisions. The price of running it quarterly is small; the price of discovering at renewal that you scored 22 on a vendor you committed to for another year is not.
Arthur Wandzel
