Know What You're Buying - The Four Types of Enterprise AI, and Why Each One Is a Different Negotiation

Type 1: Direct AI infrastructure

This is buying raw model access directly from a provider: the large language models themselves, priced by the token. It looks like a simple usage-based software deal on the surface, and underneath it behaves like a cloud-compute contract, which is a very different animal.

The trap is that token pricing is unfamiliar terrain for most procurement teams, and unfamiliar pricing is where margin hides. Two features of the pricing catch buyers out. First, output tokens cost far more than input tokens, commonly three to ten times more, because generating text takes more compute than reading it, so a workflow that produces long outputs has a completely different cost profile than one that produces short ones. Second, reasoning-capable models bill for hidden "thinking" tokens that never appear in the visible answer, and on a hard task those can outnumber the visible output several times over. Consumption also compounds at scale in ways that stay invisible during a pilot. A workflow that costs a rounding error when three people test it can generate a serious monthly bill once it runs in production across the organization.

The scale of the difference between models is the part most budgets miss. Across major providers, the spread between a budget model and a frontier model runs from roughly 10 times to more than 100 times on a per-token basis. Frontier reasoning models reach $15 to $75 per million tokens and beyond, while efficient models sit under $1. Paying frontier rates for work a cheaper model would handle identically is the single most common source of overspend, and the contract structure either exposes that lever or hides it.

What changes in the negotiation: you are negotiating a rate structure, not a license. That means itemized pricing that lists input, output, and reasoning tokens separately, volume commitment tiers with defined discount thresholds, prompt-caching discounts written into the contract rather than assumed, and usage caps with the right to renegotiate if you exceed them. Treat the committed volume itself with suspicion. The most common 2026 failure is committing to a token budget that looks generous, then blowing through it months before renewal as adoption spreads, so the commitment has to come with the right to rebalance, not just a bigger number. It also means converting a price-per-million-tokens figure, which is operationally meaningless to a CFO, into a cost per unit of work the business actually does. Paying a rate per token tells you nothing. Paying a defined amount per contract reviewed, or per document processed, is a number finance can track.

Type 2: AI-powered applications

This is a finished product with AI inside it: a contract-review tool, a spend-analytics platform, a sourcing-automation solution. You are not buying raw compute. You are buying a solution built on top of someone else's model, and the underlying AI cost is buried inside the vendor's margin where you cannot see it.

These deals feel familiar, which is exactly the risk. They look like ordinary software, so they get negotiated like ordinary software, and the AI-specific questions never get asked. Two of those questions matter more than the price. Which model actually powers this, and who controls that relationship? And what is the real value of the AI feature, separate from the software it is wrapped in? If the vendor's own model provider changes terms or pricing, your product changes with it, and you have no seat at that table unless the contract gives you one.

What changes in the negotiation: you treat it as software, and you add the AI due diligence a normal software deal skips. Ask the vendor to disclose the underlying model and who owns that dependency. Establish a baseline of the current-state process before go-live, so you can prove the tool actually improved something rather than taking the vendor's word for it. Tie payment to performance at go-live, not just to software delivery. And require a return-on-investment review with finance at 90 days, before anyone declares victory. This last point is not academic: by MIT’s 2025 estimate, roughly 95 percent of enterprise generative AI pilots never produce a measurable profit-and-loss return, and the ones that do almost always started with a baseline nobody could argue with.

Type 3: Agentic AI systems

This is the fastest-growing and least understood of the four. Agentic systems do not wait for a human instruction at each step. They take a goal, break it into steps, execute those steps across multiple systems, and complete a workflow on their own. An agent that runs a sourcing event end to end, or onboards a supplier, or processes invoices from receipt to payment, is agentic AI. Adoption has already left the pilot stage: one major vendor reported more than 29,000 agentic deals closed since launch in its most recent fiscal-year results, other large software vendors report hundreds of thousands of custom agents in use, and several have restructured their commercial models around autonomous tiers.

The commercial model is genuinely new, and that is what makes it dangerous to negotiate. The pricing model is fragmenting into per-seat, per-token, per-task, and per-outcome structures at once, with outcome-based pricing, charging only when the agent completes a defined result, as the notable new entrant. Real examples are already in market: about $0.99 per resolved support conversation from one vendor, $1.50 per automated resolution from another. Outcome pricing sounds buyer-friendly, and it can be, but it shifts the whole negotiation onto the definition of the outcome. If "resolved" is defined loosely, you pay for work that was not really done.

Two traps sit underneath the pricing. The first is what practitioners call agent-washing: existing chatbots and scripted workflows rebranded as agents, priced like autonomy but delivering none. The second is cost behavior under failure. A single failed step with automatic retries can multiply that step's cost several times over, and in a workflow with five or six system integrations, those retries compound. Cheaper per token rarely means cheaper per outcome once retries and growing context are counted.

What changes in the negotiation: you negotiate on outcomes, and you pin the outcome down precisely. Require a measurement framework agreed before signing, with a documented baseline, defined go-live criteria, and finance validation at 90 days. Define the successful outcome in the contract, including what does not count and does not get billed. Insist the scope of what the agent may and may not do lives in the contract, not just in a configuration screen someone can quietly change. Require human checkpoints for any action that commits money, selects a supplier, or touches a regulatory obligation. And build a pilot-to-production gate, so full funding releases only after the system proves itself on real work, not on a staged demonstration.