How AI Vendor Lock-In Starts Before the Contract

At a Glance

  • Lock-in starts in engineering. SDKs, tuned prompts, free credits, and click-through licenses create dependency before any contract.
  • Look past the software inventory. Find AI usage in cloud billing, marketplace, identity, expense, and API data, not only ITAM records.
  • Assign an owner. Every material AI usage pattern needs someone who can explain what it is for.
  • Test discounts against commitments. A lower rate can raise total spend if it displaces committed or reserved spend.
  • Ask three questions first. What dependency are we creating, what value do we get, and what would it cost to change direction?

AI dependency usually forms through daily engineering and product choices rather than a single strategic decision. By the time the recurring cost is visible, the alternatives may no longer be practical. No single function sees the whole decision, so ITAM practitioners collaborating with FinOps, Engineering, product, and others can bring their insights into an existing forum while technology choices can still change.

Many practitioners will recognize this story. An engineering team receives free credits to test a new AI platform. The model performs well, the provider’s software development kit (SDK) speeds integration, and native guardrails help the experiment pass security review. Within weeks, the team has tuned prompts, tool connections, and evaluation thresholds around the platform.

The pilot becomes a production service, and consumption and cost grow.

Procurement is then asked to negotiate a better deal, but the organization is no longer comparing providers from a neutral position. Changing direction would mean redeveloping workflows and integrations, reevaluating alternatives, repeating security and risk assurance, and retraining engineers in the skills they built during the pilot. Before a contract is signed, much of the negotiating leverage may already be gone.

This is AI vendor lock-in by stealth: a series of reasonable delivery decisions that add up to a commercial commitment nobody approved as a whole.

Lock-In Forms One Choice at a Time

Vendor control becomes visible at purchase and renewal, but the dependency behind it forms continuously in between.

Research published by the IBM Institute for Business Value in June 2026 found that 91% of surveyed senior executives did not fully understand their organizations’ AI dependencies, and 71% said switching their primary AI vendor or model would be difficult.

The choices that create that dependency rarely look commercial:

  • Adopting model-specific interfaces, proprietary agent frameworks, or provider-native controls because they speed delivery.
  • Tuning prompts, tool schemas, and evaluation thresholds to one provider’s or model’s behavior.
  • Accepting credits or free access that remove the need for an early funding decision.
  • Accepting a click-through license on first use, which binds the organization to supplier terms before any negotiation.
  • Accepting public marketplace pricing when a negotiated private offer already exists.
  • Routing inference to the cheapest available model without regard to where commitments are held.

These choices can collectively narrow the organization’s options. An evaluation suite tuned to the incumbent makes alternatives look weaker than they are. Provider-specific metering makes comparison and migration harder. An abstraction layer reduces integration effort, but it does not retune prompts, reproduce model behavior, or re-baseline an evaluation suite.

Detect Growing AI Usage Before It Becomes a Procurement Decision

An organization cannot evaluate a dependency it cannot see. The problem resembles shadow IT, but AI usage is harder for conventional IT asset management (ITAM) and SaaS management to detect. Model APIs, embedded features, marketplace services, and free credits do not appear as applications or seat-based entitlements.

Detection means combining ITAM and procurement records with cloud billing, marketplace transactions, identity data, expense records, and API telemetry. Discovery alone is not enough: each material usage pattern needs an accountable owner who can explain what it is for.

A Lower AI Price Can Increase Total Technology Spend

Judge AI sourcing on more than the quoted price. A supplier may offer an attractive direct agreement that pulls demand away from a marketplace or another committed platform. The unit price falls, but the organization is left with unused enterprise commitments, reduced marketplace drawdown, or reserved capacity the displaced workload was expected to consume.

A Discount That Would Have Doubled the Cost

A FinOps practitioner was asked to support moving a contract out of one cloud marketplace to secure a substantial discount from a competing provider. Rather than block the move, they modeled it. Every dollar moved would stop counting toward an enterprise commitment the organization had already signed, and the shortfall would still be payable at term end.

On a $1 million move, the organization would pay the new provider once and pay again to cover the missed commitment. The practitioner explained to the business that the move would roughly double the cost of the workload rather than discount it.

Principle: The discount is real, and so is the commitment it displaces. Only one of them appears on the quote.

The same effect shows up in two other places.

Model Routing

When an AI gateway sends each request to the lowest-priced model, practitioners report that routing logic optimizes unit price with no view of the commercial position. Some have had to ask engineering to force traffic back toward a provider whose commitment was at risk, after routing concentrated consumption on another supplier and left underspend accumulating. Routing decisions are commercial decisions as well as architectural ones.

Purchasing Channel

The same model can cost different amounts depending on how it is bought. Practitioners report higher premiums on older models consumed through a hyperscaler’s managed service in regional capacity than when buying direct, and they see the pattern across more than one provider.

Many organizations buy through several channels at once. In the Tokenomics Foundation’s September 2026 State of Tokenomics survey, 87% of respondents consumed models through a cloud token provider, and half spread their AI consumption across four of the six procurement and hosting channels the survey tracked.

Practitioners also describe that third-party model consumption through a hyperscaler may not draw down that hyperscaler’s enterprise discount agreement at all, so any discount requires a separate private pricing agreement.

Before approving a rate, FinOps, IT financial management (ITFM), and Procurement should test whether the decision:

  • Reduces spend counted toward a cloud enterprise commitment or marketplace arrangement.
  • Displaces demand covered by reserved capacity or savings commitments.
  • Changes negotiated discount tiers or creates overlapping commitments during migration.
  • Introduces a new minimum spend, credit expiry, or auto-renewal obligation.
  • Changes the usage, and therefore the cost, of other existing tools, including cloud.

Commitment Is Not the Problem

The answer is not to avoid AI dependency. Organizations choose managed services and provider-specific capabilities because they improve speed, reliability, security, and access to scarce capacity. UK Government guidance on technical lock-in accepts that the value of a complex managed machine learning service may outweigh limited portability. A twelve-month commitment can be economically sound where demand is forecastable and reserved throughput protects a valuable business service.

Optionality has a price too: One procurement team facing a cloud renewal chose a one-year term over three because it lacked a reliable view of usage. The shorter term cost more per unit, but bought time to establish demand and enter the next negotiation with evidence. That premium was a deliberate payment for flexibility, not waste.

Portability is not free either. Routing, abstraction, and provider-neutral controls add cost, and may block the provider-specific capabilities that created the value in the first place. Maximum portability is a position to argue for, not a default requirement.

Why the Calculation Is Harder Now

AI-enabled software is moving between seats, outcomes, requests, tokens, credits, and hybrid pricing models, each with different pooling, expiry, and rollover rules. As the billing unit shifts, so does the effective price, and AI value becomes hard to measure without the right people, metrics, and observability. A 2026 survey of more than 230 software and AI companies found that hybrid pricing had become the most common model, used by 37% of respondents, up from 25% a year earlier.

The definitions behind a unit are not always documented. Practitioners describe a provider that applied a weekly per-user budget without stating whether the week was calendar or rolling, and could not confirm the basis when asked.

A Billing Model That Changed Without a Renewal

GitHub moved Copilot billing from a seat-based subscription to metered consumption. A FinOps practitioner modeled the new pricing against existing usage and estimated that the organization’s bill would rise by around 350%. The team then had to build credit visibility, model access controls, and chargeback from scratch, after the commercial terms had already changed.

Principle: The billing model is a commercial term, and it can change without a renewal event.

Supplier contracting practice is also immature, which works against the buyer. Practitioners negotiating with frontier model providers report resistance to termination for convenience, no clear answer on whether a commitment can be increased mid-term, discounts that lapse once committed volume is exceeded, and new product commitments bundled into the renewal of an existing one.

Options can also be removed from outside of the organization. Sovereignty rules, regulatory restrictions on named models, and supplier failure, acquisition, or discontinuation can each make an economically attractive option unusable. Weigh these risks in the commercial decision, before the architecture is set, rather than in a later compliance review.

Standardized billing data will reduce some of this measurement dependency over time. Release planning for the FinOps Open Cost and Usage Specification (FOCUS) includes AI model identity and token consumption in version 1.5. Until adoption is broad, be cautious about reconciliation processes that work with only one provider’s format.

No Single Function Can See the Whole Decision

AI lock-in is a commercial shift-left problem. It needs commercial input while the technology choice can still change, not only at purchase and renewal, and that means bringing the relevant personas into the process much earlier. Where an AI Investment Council or equivalent forum already exists, it can provide that cross-functional view. This does not require a new approval layer or governance body. The aim is to stop one team from optimizing a supplier price while the organization’s overall technology economics get worse.

Persona Contribution to the decision
FinOps Tests whether forecast usage will turn the discount into realized value and whether the economics hold once the workload is running.
Procurement Structures the commitment and negotiates substitution, export, and transition rights before technical dependency removes the leverage to do so.
ITAM Discovers the supplier, entitlement, and embedded capability, and finds the owner who can explain what the usage is for.
ITFM Tests affordability under alternative demand scenarios and models stranded commitment exposure across the portfolio.
Engineering Explains how deeply the provider is embedded, what migration would require, and where routing and deployment choices carry commercial consequences.
Security, Legal, and Risk Identify constraints that make an attractive option unusable, and own the risk register entry for supplier continuity and failure.

Three Questions Before Making an AI Commitment

1. What dependency are we creating?

Identify the provider-specific elements across application, model, data, controls, operations, skills, and contract. Record who owns each one, whether the supplier appears in the relevant ITAM, procurement, and risk inventories, and the rough effort required to replace it.

2. What value are we receiving in return?

Quantify the benefits, such as faster delivery, better model quality, lower unit cost, guaranteed capacity, and reduced operational effort. Then test whether the forecast volume is credible and whether the quoted saving survives its effect on existing commitments. A lower supplier rate creates no value if it raises total committed spend.

3. What would it cost to change direction?

Look beyond data migration. Include redevelopment, prompt and evaluation work, assurance, retraining, contractual exposure, commitments stranded after the workload moves, and the cost of running two providers in parallel during the transition. Then define the conditions that would trigger a review, so the decision rests on a review date rather than an assumption.

Make the Dependency a Decision

Lock-in forms through the daily interaction of architecture, delivery, consumption, and commercial terms, so the commercial view has to arrive just as early. The aim is an informed decision, made while the organization can still choose. That decision can include commitment, and it stays with the teams that own it.