Key takeaways
Evidence
Most AI vendor selection decisions are made on capability grounds. The economic risks — pricing model exposure, lock-in depth, and exit cost — are typically under-assessed at purchase time and fully understood only when it is expensive to act on them.
Evidence
Per-token and consumption-based pricing creates a fundamentally different budget risk profile than seat-based or platform licensing. Volume growth can produce non-linear cost increases that were not modelled at purchase.
Interpretation
Lock-in in enterprise AI is multi-dimensional: model lock-in, platform lock-in, data format lock-in, and workflow dependency lock-in do not all move together and are not all equally reversible.
Interpretation
Exit cost estimation is a legitimate procurement exercise, not a sign of distrust. Vendors that cannot support a clean exit should price their contracts accordingly.
The vendor selection problem
Evidence
The standard AI vendor evaluation process optimises for capability: which model performs best on the relevant tasks, which platform integrates most cleanly, which vendor has the best support model. These are legitimate questions. They are also incomplete ones.
Interpretation
The economic conditions of an AI vendor relationship are different from conventional enterprise software in ways that matter at scale. Pricing can be variable and exposure to cost growth is hard to predict. The market is consolidating rapidly, which means vendor risk profiles change faster than procurement cycles. And the technical lock-in created by AI deployments is subtler than conventional software lock-in — it does not always appear in a single integration layer, and it accumulates over time.
Interpretation
A mature AI vendor selection process evaluates economic and risk characteristics alongside capability. The questions are not complex, but they require deliberate effort to answer at purchase time.
Pricing model risk
Evidence
Enterprise software procurement is largely built around predictable cost: seat licences, platform fees, and annual true-ups. AI vendor pricing introduces a different risk structure.
Evidence
Consumption-based pricing (per token, per API call, per task) creates direct exposure to volume growth. If AI adoption expands — which is the intended outcome of most AI programmes — the cost grows with it. This is rational, but it means that a vendor contract that looks affordable at proof-of-concept scale can become a significant budget line at enterprise scale without any change in the agreed price per unit.
Pricing risk example: A European financial services company deployed an AI-powered customer service chatbot using a consumption-based API pricing model. Initial proof-of-concept costs were £2,000 monthly at 50,000 queries. After successful rollout to all customer channels, query volume reached 2 million monthly within six months, driving costs to £75,000 monthly — a 37x increase. The per-query price had not changed, but the absence of volume discounts or pricing caps meant the budget impact was far larger than procurement had modelled. The company renegotiated for tiered pricing with volume commitments.
Interpretation
The risk is not the price per token. It is the absence of a cost ceiling combined with the difficulty of predicting consumption at scale. Workload design choices, user behaviour, and pipeline architecture all affect token consumption in ways that are hard to model in advance. Organisations that approve an AI vendor based on a cost estimate at current usage and do not model volume growth scenarios are accepting more budget risk than they realise.
Evidence
Platform fees with consumption overlays are increasingly common in enterprise AI. A base platform licence provides access; usage above a threshold triggers consumption pricing. This model limits downside cost risk but can create pricing cliffs at volume thresholds that are difficult to anticipate.
Interpretation
Seat-based AI pricing (typically for AI-augmented SaaS products) is more predictable but creates a different risk: the cost is fixed whether or not the AI features are actively used. Organisations that commit to AI-augmented seat licences and then see adoption below projections pay for capability that is not being used. The reverse — adoption significantly exceeding projections — is typically welcomed but can still require a renegotiation.
Evidence
A particular challenge with SaaS-based AI pricing is that the underlying token consumption is often aggregated and obscured within subscription tiers, making it difficult to understand true unit economics or optimise usage patterns. For a detailed analysis of this visibility gap, see
SaaS Token Opacity: The Hidden Economics of AI Subscriptions .
Questions to ask at procurement :
What does total cost look like at 2x and 5x current usage?
Are there volume discounts at scale, and what triggers them?
What notice is required before pricing changes, and what is the cap on annual increases?
Is there a spend commitment that triggers a different pricing tier?
Lock-in dimensions
Interpretation
Lock-in in enterprise AI is not a single condition. It is a combination of factors that accrue over time and that do not all move together. Understanding which dimensions apply to a specific vendor relationship helps prioritise exit risk and mitigation strategies.
Evidence
Model lock-in occurs when outputs, evaluations, and fine-tuning are built around a specific model's characteristics. An enterprise that has fine-tuned a model, built evaluation frameworks around its specific output format, or accumulated significant prompt libraries optimised for its behaviour has invested in that model in a way that does not transfer. Switching models, even within the same provider, can require significant re-evaluation and re-optimisation.
Insight
Lock-in case study: Klarna's AI customer service implementation, which replaced 700 full-time agents with AI-powered support, was built on OpenAI's platform with extensive prompt engineering and workflow integration. The company's public statements indicate the system handles two-thirds of customer service chats. Switching to an alternative model provider would require not just API changes but re-optimisation of prompts, re-evaluation of quality metrics, and retraining of escalation workflows — representing months of engineering effort even with well-abstracted integration layers.
Evidence
Platform lock-in is the most visible dimension. It occurs when AI capabilities are embedded in a platform that also manages other enterprise functions — CRM, ERP, productivity suites — and separating the AI layer from the platform becomes costly or impractical. This is increasingly common as Microsoft, Salesforce, ServiceNow, and similar platforms embed AI deeply into their core product. The capability integration is a genuine benefit; the lock-in is a genuine cost. Both should be assessed.
Evidence
Data format and pipeline lock-in occurs when AI workflows are built around a provider's data infrastructure: vector databases, embedding models, knowledge base formats, or proprietary retrieval systems. If migrating AI capability requires rebuilding the data layer, the migration cost is substantially higher than it appears when only the model API is considered.
Interpretation
Workflow dependency lock-in is the least visible but often the most durable. It occurs when human workflows — the way people actually do their jobs — adapt around an AI tool's capabilities and limitations. When users build their work patterns around specific AI behaviours, replacing the AI creates disruption that is not measured in integration costs. It appears in productivity dip, user resistance, and retraining expense.
Consolidation and acquisition risk
Evidence
The AI vendor market is consolidating rapidly. Point solutions that exist independently today may be acquired, discontinued, or meaningfully changed within a two to three year investment horizon.
Interpretation
The consolidation risk is asymmetric. Consolidation into a stronger acquirer can be a positive outcome: better support, deeper integration, more durable investment. Consolidation into a distracted or strategically misaligned acquirer can mean reduced investment in the product, changed pricing, or eventual discontinuation.
Interpretation
The categories at highest consolidation risk are typically small pure-play vendors addressing specific AI cost visibility, optimisation, or governance functions. These products are acquiring customers and demonstrating value in areas where larger platforms have identified gaps — which makes them attractive acquisition targets. Buyers should assess whether they are comfortable with the range of plausible acquirers, not just the current vendor.
Questions to ask:
Has the vendor raised venture capital, and at what valuation relative to its current revenue? (High multiples create acquisition pressure.)
Is the vendor's core function something a larger platform would absorb?
What happens to your data, your integrations, and your contract terms in a change of control?
Estimating exit costs
Interpretation
Exit cost estimation is a legitimate procurement exercise. It produces a clearer understanding of the real cost of the vendor relationship and creates a negotiating basis for contract terms that reduce exit risk.
Evidence
Integration rebuild cost : the engineering effort required to rebuild the integration between your workflows and an alternative vendor. This is the most tractable exit cost to estimate — it is roughly proportional to the initial build cost, subject to the quality of the abstraction layer.
Evidence
Data migration cost : the effort required to export, reformat, and re-ingest data from the vendor's storage and retrieval systems. Proprietary formats significantly increase this cost. Open, documented formats significantly reduce it. Ask at procurement how data is exported, in what format, and what the migration tooling looks like.
Interpretation
Re-evaluation and re-optimisation : if the vendor relationship has produced fine-tuned models, specialised prompts, or evaluation frameworks, these may not transfer. The cost of rebuilding this intellectual work is real but hard to estimate in advance. It can be partially mitigated by maintaining provider-agnostic abstractions.
Interpretation
Transition productivity cost : the period during which teams are operating on a reduced AI capability or adapting to a different tool. This is often the largest real cost of AI vendor transition, and the one most consistently excluded from exit cost estimates.
Exit cost benchmark: A US technology company evaluated switching from one major cloud AI provider to another for its document processing pipeline. The analysis revealed: integration rebuild (3 months engineering, £180k), data migration (reformatting 2.5 million documents, £45k), prompt re-optimisation (6 weeks, £60k), and estimated productivity impact during transition (15% throughput reduction for 2 months, £120k opportunity cost). Total exit cost: £405k against annual vendor spend of £240k — a 1.7x multiple. The company negotiated improved pricing with the incumbent vendor instead of switching.
Speculation
A rough rule of thumb: for a well-embedded production AI deployment, total exit cost typically runs between one and three times the annual vendor spend, depending on integration depth and workflow dependency. This is a planning assumption, not a binding estimate — but it is useful for framing the economic case for transition decisions.
Contract terms that reduce economic risk
Interpretation
Several contract provisions materially reduce the economic risk of an AI vendor relationship. Many are negotiable, particularly for significant spend levels.
Evidence
Pricing caps and commitment tiers : a cap on annual price increases and a defined discount structure at volume thresholds converts variable pricing risk into a managed cost envelope.
Evidence
Data portability commitments : a contractual obligation to provide data in documented, open formats on request. This is increasingly standard for credible enterprise AI vendors; its absence is a risk signal.
Evidence
Change of control provisions : a right to exit the contract or renegotiate terms in the event of a material change of ownership. This provision is common in enterprise software and increasingly available in AI vendor contracts.
Evidence
Model version stability : a commitment to support a specific model version for a defined period, with advance notice before deprecation. This reduces the risk of a silent capability change disrupting a production deployment.
Evidence
Audit rights : the right to inspect the vendor's use of your data, model behaviour logs, and billing calculations. Particularly important in regulated industries where data governance obligations extend to third-party processors.
A framework for economic risk assessment
Interpretation
Pricing risk : how variable is total cost as usage grows? What are the worst-case scenarios at 2x and 5x current volume? Is there a spend ceiling or commitment structure available?
Interpretation
Lock-in depth : which lock-in dimensions apply — model, platform, data, workflow — and how reversible is each? What is the estimated exit cost at a planning horizon of two and four years?
Interpretation
Consolidation exposure : what are the plausible acquirer scenarios? Are any of those scenarios unacceptable? What contract provisions address change of control?
Interpretation
Governance alignment : does the vendor's data handling, audit capability, and SLA structure align with your compliance obligations? Can they support the governance requirements that will apply to production AI at scale?
Interpretation
Capability is necessary. Economics and risk are also necessary. The organisations that evaluate AI vendors on all four dimensions consistently make fewer expensive mistakes.
References and further reading
