Engageably

05 — Journal

← Back to Journal

Platform commitments in AI are a category error

May 18, 2026 · 7 min read · Studio · Benjamin Rodriguez

Executive TL;DR

  • Companies buy AI platforms like enterprise software. They behave like volatile infrastructure with a half-life in months.
  • Multi-year, vendor-locked, deprecation-blind terms are wrong for a capability layer that rewrites itself faster than any procurement cycle can track.
  • The CEO and CFO who signed these deals own a structural exposure, not a transformation asset. The P&L will surface it at the next migration.
  • Stop the wrong fix. The answer is not to stop investing. It is to change the unit of commitment from platform to outcome.

The Shift

Enterprise software used to follow a clean arc. Evaluate, select, commit, amortize. What you signed for in March was what you ran in October, and what you ran three years later. Procurement assumed that stability.

AI does not hold that shape. Stanford’s 2026 AI Index reports that performance on a standard coding benchmark, SWE-bench Verified, rose from 60% to near 100% in a single year. Industry produced over 90% of notable frontier models in 2025 (Stanford HAI 2026 AI Index, 2026). The platform the buyer picked at signing is often not the best capability available by go-live. The product keeps moving while the contract is still being drafted.

Why It Matters Now

Deprecation is no longer theoretical. OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT on a single date in February 2026 (OpenAI Help Center, 2026). Microsoft Foundry now sets every model it releases a retirement date at launch, eighteen months out (Microsoft Learn, 2026). The vendor writes the half-life into the documentation. When the vendor retires a model, the buyer absorbs the friction. Integration rewrites. Model retraining. Renegotiation leverage, gone. The vendor absorbs none of it. And the substitute is rarely a like-for-like swap. Moving from GPT-4.1 to the next generation of reasoning models costs more per unit of output. GPT-5 runs $10.00 per million output tokens, a 25% increase over GPT-4.1 (TensorOps, 2026).

A platform commitment in AI is not a hedge against uncertainty. It is a contractual obligation to absorb volatility on the vendor’s behalf.

What Most Companies Are Still Doing

Running an RFP. Scoring vendors on current benchmark performance, negotiating volume tiers, signing multi-year terms, and treating the result as a closed decision. That motion is right when what you are buying is stable. It is the wrong motion here.

The result is a portfolio that looks decisive on the org chart and is brittle in the field. The contract sets the capability ceiling, not what the market can now deliver. Every quarter, the gap between the committed platform and frontier capability widens. And new commitments keep landing on the old terms. Gartner’s 2026 CIO survey found 84% of respondents expect to increase generative AI funding in 2026 (Gartner, 2026). At the same time, Gartner projected that 30% of generative AI projects would be abandoned after proof-of-concept by the end of 2025 (Gartner, via Infomineo, 2025). More money, signed on old terms, into a capability layer still proving out.

What the Best Operators Are Doing Instead

The operators gaining durable advantage have separated the commitment question from the capability question. They commit deeply to outcomes, workflows, and data infrastructure. They hold model and platform selection deliberately loose. The four moves below share one principle: the contract anchors what the business needs, not what the vendor currently sells.

1. Outcome anchoring over vendor anchoring. Contracts specify the business result and the performance threshold, not the model or the vendor delivering it. You secure substitution as a right at signing, instead of asking for it as a favor at renewal. The unit of commitment is the outcome.

2. An integration layer as a strategic asset. An integration layer sits between the workflow and the model. A model swap becomes a settings change, not a workflow rebuild. This is the only thing that holds migration cost down when the vendor’s roadmap moves without you.

3. Quarterly capability reviews. Quarterly reviews treat model selection as an ongoing operational question. The review asks one thing. Is the committed capability still the best available capability for the contracted outcome. If the answer is no, you exercise the substitution right.

4. Deprecation clauses as standard terms. Before signing, the best operators negotiate explicit remedies for model deprecation: migration credits, timeline guarantees, exit rights. The market is starting to price this directly. Gartner notes emerging AI warranty products that refund license fees if a model fails to meet specified accuracy or fairness metrics (Gartner, 2026). Performance risk is now a contractable liability. Absent these terms, the liability sits entirely on the buyer.

Implications for the Next 12 Months

The Write-Down Problem Becomes Visible

As model generations turn over, organizations holding commitments to deprecated or underperforming capabilities will face a binary choice. Absorb the underperformance, or recognize the migration cost. Neither outcome is invisible to the P&L. Migration is not free. A move from a retired proprietary model to its successor can mean materially higher usage costs and cloud spend (TensorOps, 2026).

The CFO who has not pressure-tested the deprecation scenario in the AI line items is carrying an unpriced liability. The vendor will price it for them, on its own timing.

Most enterprise procurement and legal teams learned to negotiate on price, uptime, and data residency. They never learned to negotiate on model substitution rights, deprecation timelines, or capability benchmarks. The gap between what these teams know how to protect and what AI commitments actually require is widening with every enterprise deal signed under old terms. Closing that gap is a 2026 priority, not a 2027 one.

The Vendor Market Will Consolidate Around This Tension

Vendors who build in substitution-friendly terms and make model swaps easy by design will attract the operators who have already learned this lesson. Vendors who resist will keep customers only until the switching cost falls below the performance gap. Hold that against the broader picture. McKinsey’s 2025 State of AI survey of 1,993 participants found 88% report regular AI use in at least one business function. Only about a third report scaling AI programs. Just 6% qualify as high performers, with more than 5% of EBIT attributable to AI (McKinsey, 2025). The high performers are the ones who will move first on terms. The CEO who understands this dynamic now has negotiating leverage they will not have at renewal.

Executive Next Step

Before the next AI renewal or new commitment reaches signature, map every current AI commitment against three things. Its deprecation clause. Its substitution rights. The performance threshold that triggers either. What is uncontracted is unprotected. Run the diagnostic before the vendor runs it for you.

Sources

← Back to Journal