The owner of a thirty-person professional services firm signed a two-year Microsoft 365 Copilot Enterprise deal in mid-2025 for twenty seats, added a ChatGPT Teams subscription for the leadership group, and bought an AI customer-service tool on a three-year contract. Eighteen months later, the Copilot deal has delivered productively for five of those twenty seats and is sitting dormant for the rest. The chatbot vendor has been acquired. Pricing has changed. The contract auto-renews in six months. She is asking a fair question. If I were buying these tools fresh today, what would I do differently.
The tooling, on the whole, has delivered. The buying model has not. That is the honest shape of the twelve-month review for most owner-operators who bought their first serious AI tools in 2024 or 2025. The technology improved faster than the contracts allowed for, prices moved, vendors consolidated, and regulators sharpened expectations. This post is the answer to her question, written for owners who are looking at a renewal date in the next six to twelve months and want a clear-eyed view of what the next round should look like.
What did the 2024 to 2025 buying model get wrong?
Firms treated AI like traditional SaaS. Per seat. Multi-year commitments. Headline price as the lead metric. Limited attention to usage economics, vendor lock-in, data rights, or the operational cost of integrating tools into core workflows. Moneypenny’s 2025 survey of 750 UK decision-makers found roughly seven in ten firms were either using or actively considering AI, and most of that buying happened on standard SaaS terms before the buying model had settled.
The mistake was not the tooling choice. Copilot, ChatGPT Teams, the customer-service chatbots and the AI-augmented CRMs largely did what they said on the tin. The mistake was the contract shape sitting underneath them. McKinsey’s 2025 State of AI report found that only around 6 percent of organisations qualify as AI high performers, and the differentiator is not which model they chose, it is how they treated foundations: data, governance, workflow redesign, and the discipline to redesign rather than sprinkle.
Mavvrik’s 2026 cost data tells the harder version of the same story. Between 80 and 85 percent of enterprises miss their AI cost forecasts by more than a quarter. Around 30 percent of generative AI projects are abandoned after proof of concept. The 2024 to 2025 contract did not contain the controls that would have caught those drifts early. The 2026 contract has to.
How has AI pricing actually moved in twelve to eighteen months?
Every major vendor has added a usage component to its pricing, and the value has moved with it. OpenAI’s ChatGPT business pricing shifted twice in eighteen months. Microsoft 365 Copilot stayed at thirty US dollars per seat but added Copilot Studio at separate capacity-pack pricing, with packs of 25,000 Copilot Credits at 200 US dollars per pack per month, alongside a pay-as-you-go meter where firms pay only for credits consumed. Anthropic moved its enterprise offering toward usage-based with seat-based admin. Google bundled Gemini into Workspace then split into Business and Enterprise tiers.
The Register’s April 2026 analysis put the pattern bluntly. AI is not like SaaS, where costs shrink with scale, because every query carries a real compute cost tied to GPU capacity and energy. Vendors are now raising base prices and steering toward token-based or credit-based billing for the workloads that actually drive value. Buyers who locked in long per-seat deals discovered six months later that the same product was effectively more expensive because the workloads that mattered had moved to the usage tier they did not have access to.
The practical consequence: per-seat pricing alone is unstable. It is still useful as a comparison anchor and an entry tier, but the 2026 contract needs to model usage scenarios, set caps with overage protection, and ask for usage dashboards before signing, not after the first surprise bill.
What changed with context windows, RAG and your data?
Context windows expanded from 32K to 200K to over a million tokens between mid-2024 and mid-2026. Tools sold as retrieval-augmented generation, priced on the basis that the model needed help finding documents, became thinner where long-context models could ingest the documents directly. The buying calculus shifted. A RAG-based vendor signed in 2024 may now be doing a job that the long-context model native to a tool you already pay for could handle in a single session.
The nuance matters. RAG is not dead, and for document collections beyond a few hundred items, well-structured retrieval is still cheaper, faster and more auditable than stuffing everything into context. But the cost difference between the two architectures has narrowed for the document sets most owner-operated firms actually work with: hundreds, not millions, of contracts, emails and policies. The buying question moved from which model has the biggest context window to what information the system needs to use, how it is stored, and what your rights are over it.
That last question is where 2024 contracts often sat thin. Debevoise’s late-2025 analysis on contractual use limitations argued that the biggest enterprise AI problem in 2026 is NDAs and client contracts signed in 2023 to 2024 that restrict how that data can be used in AI systems. The 2026 contract carves out training-data use, names data-processing roles under UK GDPR, and pins model versions so behaviour does not silently change beneath the workflow.
Why is vendor volatility now a procurement issue, not a footnote?
The list of AI products that died, pivoted or were absorbed during 2024 to 2026 is short but instructive. Inflection AI sold to Microsoft in a 650 million dollar deal that effectively transferred staff and a technology licence. Adept sold to Amazon. Stability AI restructured after the founder departed. Humane shut down its AI Pin and was sold to HP for around half what investors had put in. Rabbit’s R1 device sold around 100,000 units against heavy returns. OpenAI retired Sora in 2026 after burning compute against thin revenue, according to Digital Applied’s 2026 review of AI product failures.
For an owner-operator who signed a three-year contract with a smaller AI vendor in 2024, the relevant question is not whether their specific vendor will fail. Most will not. The question is whether the contract carries useful clauses if it does. Change of control. Sustained service degradation. Material pricing moves above an agreed threshold. The right to export data and conversation histories in a usable format. None of those clauses were standard in 2024 mid-market SaaS deals. They are now part of the 2026 baseline.
Vendor lock-in compounds the problem. Switching AI vendors is no longer just an API migration. It is context, workflows and the institutional memory your team has built inside the vendor’s tool: agents, prompt libraries, fine-tuned preferences, conversation history. The 2026 contract treats data export as a first-class right, not a footnote in the appendix.
What does the 2026 buying playbook look like?
Twelve-month pilots, not three-year commitments. Usage caps with overage protection. Model-version pinning with named upgrade paths, in line with the FINOS AI governance framework’s guidance on explicit versioning and advance deprecation notices. Exit clauses on vendor change of control and on price moves above an agreed threshold. Training-data carve-outs and named data-processing roles under UK GDPR. A right to audit usage analytics. Named acceptance criteria and named decision points baked into the contract, so the renewal becomes a real review rather than a default.
For the contract sitting on the owner’s desk right now, the practical sort is short. Pull the contract. List the exit options. Sense-check per-seat utilisation against cost, the dormant fifteen seats on a twenty-seat Copilot deal are the conversation, not the active five. Ask the vendor for a 2026 pricing comparison and a current usage breakdown. Then decide whether to renew at the same shape, switch shape from per-seat to usage or vice versa, or move vendor entirely.
The point is not to second-guess the original decision. The 2024 deal was reasonable for the market that existed then. The point is that the 2026 deal is being signed into a different market, and the contract shape that fitted the first round will not fit the second. Asymmetric risk applies here. A wrong year-one deal cost some money. A wrong year-two deal locks in another two years of the wrong shape.
