How Open Source AI Is Shifting Power Back to the Buyer

If you’re in IT procurement, pay attention: open source AI is more than a developer trend.  For many, it’s a hidden lever that could change AI solution negotiations. For those who aren’t familiar with the role open source is playing in enterprise AI negotiations, let’s run through some background.

Enterprise AI negotiations are starting to look a lot like cloud negotiations did 15 years ago. At first, buyers assumed they had no choice but to adopt a fully integrated stack from a handful of dominant vendors. Over time, that changed. Open ecosystems emerged, interoperability improved, and procurement teams gained leverage.

The same pattern is now emerging in AI agents.

Platforms like Salesforce Agentforce, Anthropic Claude-based enterprise agents, OpenAI proprietary agent frameworks, and Google Gemini tooling are powerful offerings. But they’re increasingly competing against a rapidly maturing open source ecosystem built around interchangeable models, orchestration frameworks, and “agent harnesses.”

That matters enormously for procurement professionals. Once enterprises understand they can build viable AI systems without locking themselves into a single proprietary stack, the negotiating dynamic changes.

The Shift Procurement Teams Should Pay Attention To

For the past two years, much of the AI conversation has focused which models are the smartest (GPT, Claude, Gemini, Llama, Mistral, DeepSeek, etc.). Now, the bigger strategic question is: who controls the orchestration layer?

In practical terms, this means: which system manages workflows? Which framework coordinates tools, memory, search, approvals, and reasoning? Which platform determines how AI agents interact with enterprise systems? This orchestration layer, sometimes called the “agent harness,” is becoming the new control point in enterprise AI and it’s increasingly open source.

What Is an “Agent Harness”?

To understand why this matters, it helps to understand what an agent harness actually does.

Think of a large language model (LLM) as a very powerful brain. It’s capable of reasoning, writing, and analysis, but on its own, it has no hands, no memory, and no guardrails. An agent harness is the complete infrastructure that wraps around that brain to make it a functional, reliable AI agent.

If the model is the brain, the harness provides the hands, memory, and safety boundaries. Specifically, the harness handles four critical jobs:

• Orchestration — managing the step-by-step loop of reasoning, acting, and checking results so the agent can complete complex multi-step tasks without going off the rails.
• Tool execution — connecting the model to the outside world, whether that means running a web search, querying a database, calling an API, or executing code.
• Memory — giving the agent context across sessions so it isn’t starting blind every time. Without memory, an agent forgets everything the moment a conversation ends.
• Error recovery and safety boundaries — ensuring the agent doesn’t call APIs out of sequence, spiral into infinite loops, or take actions it shouldn’t. In enterprise environments, this layer also manages human-in-the-loop approvals.

Raw language models are stateless by default. They have no inherent ability to persist information or manage multi-step workflows. The harness is what bridges that gap between “smart model” and “functional business agent.”

One of the most visible open source examples is LangChain and its LangGraph ecosystem. It started as a framework for connecting language models to enterprise data and tools and has evolved into its Deep Agents platform for orchestrating autonomous agents.

Capabilities that were once exclusive to proprietary enterprise AI platforms (persistent memory, multi-agent coordination, human approval checkpoints, long-running workflows) are increasingly becoming open source framework features available to everyone.

How Open Source Creates Leverage Against Consumption Pricing

The market is rapidly shifting toward token-based pricing, agent execution pricing, reasoning surcharges, workflow consumption charges, and API metering. These costs scale nonlinearly and it’s creating widespread pain for enterprise IT procurement leaders. 56% say their biggest AI spend challenge right now is unpredictable usage and scaling costs.

This is where open frameworks and open models become strategically important, even if an enterprise still uses frontier proprietary models. The key advantage is control.

Intelligent Workload Routing

Not every task requires the most expensive reasoning model. With open orchestration frameworks like LangChain or LangGraph, enterprises can dynamically route workloads: premium models for complex reasoning, cheaper models for summarization, open source models for internal automation, and local models for sensitive workflows. Instead of paying premium token costs for every interaction, companies can optimize workload placement. That becomes a major financial lever.

Avoiding Vendor-Controlled Metering

In proprietary ecosystems, the vendor often controls orchestration, memory, workflow logic, and billing visibility. That means enterprises may not fully understand which actions consume tokens, which prompts drive cost spikes, or how inefficient workflows become over time.

Open harnesses provide far more observability and tuning flexibility. Companies can reduce unnecessary context, cache responses, compress prompts, limit reasoning depth, reuse embeddings, and selectively downgrade workloads. That optimization can materially reduce enterprise AI spend.

Open Models Are Improving Rapidly

Models from Meta Llama, Mistral, DeepSeek, Qwen, and others are improving at an extraordinary pace. For many enterprise tasks, open models are “good enough.” This includes activities like document extraction, internal search, classification, summarization, workflow routing, contract analysis, and structured data extraction. Note that “good enough” at a fraction of the token cost becomes extremely compelling at scale.

Local Inference Changes the Economics Entirely

Some organizations are beginning to run smaller models locally on private infrastructure, inside cloud VPCs, or even on high-end enterprise workstations. That dramatically changes the cost structure: no per-token API charges, no external inference billing, and potentially lower data governance concerns. It does not eliminate infrastructure costs, but it converts variable AI spend into more predictable infrastructure spend.

Why This Changes Negotiations

This does not mean enterprises will stop buying from OpenAI, Anthropic, Salesforce, or Google. Many companies will still use those models heavily. But it changes the balance of power, because enterprises are realizing: “We may not need your entire proprietary stack, and we may not want all of our AI workloads billed at premium consumption rates.”

That is a major distinction. A company could now use open orchestration, mix multiple models, swap providers dynamically, route workloads based on price/performance, keep enterprise memory and workflows portable, and selectively minimize expensive token consumption.

Reduced Switching Costs

If workflows are built in an open framework, changing models becomes easier. Today, Claude may outperform on reasoning, GPT may lead in coding, Gemini may excel in multimodal workflows, and open models may be sufficient for many internal tasks. A portable architecture allows enterprises to renegotiate aggressively because suppliers know workloads can move.

Better Pricing Pressure

When procurement can credibly say “we can route 40% of these workloads to lower-cost or open models,” pricing conversations change immediately. This is especially important because AI pricing remains highly fluid. The market is still discovering pricing equilibrium. Competition matters.

Avoiding “AI Platform Lock-In”

Many vendors are attempting to become the system of intelligence layer for the enterprise. Once deeply embedded, these platforms become difficult to replace. Open harnesses reduce that dependency by separating the workflow layer, the reasoning layer, the model layer, and the infrastructure layer. That modularity is strategically valuable.

What Procurement Leaders Should Do Now

One important consideration: open source AI does not eliminate enterprise costs. Companies still need security controls, governance, observability, identity integration, evaluation frameworks, skilled engineering talent, infrastructure, monitoring, and compliance review.

The IT procurement conversation is less about “cheap AI” and more about strategic optionality, pricing leverage, and consumption control.

With that in mind, here are four specific actions IT procurement leaders should take now to leverage open source AI:

Build AI Benchmarking Discipline Early

Most organizations still lack normalized benchmarks for cost per workflow, cost per agent task, model accuracy by use case, token consumption, human review rates, and latency. That benchmarking capability will become essential in negotiations.

Push Vendors on Consumption Transparency

Ask vendors directly: What actions trigger token charges? How are reasoning loops billed? What causes context expansion? How is memory stored and priced? Are agent retries billed separately? Can workloads be routed externally? Are there token caps or guardrails? These questions matter more than feature demos.

Encourage Multi-Model Strategies

The future is unlikely to be single model. Most enterprises will eventually run a combination of premium frontier models, smaller specialized models, open source local models, vendor-specific copilots, and internal orchestration layers. Procurement should negotiate accordingly.

Treat AI Like Cloud Procurement in 2012

The organizations that gained leverage in cloud negotiations were the ones that avoided early lock-in, maintained portability, benchmarked aggressively, and preserved architectural flexibility. The same playbook is emerging again in AI.

Interested in learning more about AI procurement and cost management best practices. Contact us.