Open Versus Closed AI Is the Wrong Question: Why Businesses Will Use Both

Businesses do not need to join an AI religion. They need to choose the right model for each workload. Open models can offer control, customisation and lower operating costs. Closed services can offer frontier capability, managed infrastructure and faster deployment. A serious enterprise will often use both.

Current reporting shows AT&T increasing its use of open or open-weight models while continuing to use proprietary services. Different reports describe open models handling roughly 40 per cent of employee requests, with the company aiming to increase that share and reporting large savings in selected workloads. The precise percentages change as routing evolves; the strategic signal is more important than one snapshot.

Open source and open weight are not identical

The vocabulary is frequently blurred. An open-source system normally provides code and licensing rights that allow inspection and modification. An open-weight model may make trained parameters available while withholding training data, full code or other components. Licence restrictions can still limit commercial use.

A business should therefore read the actual licence and technical documentation rather than treating “open” as a complete risk assessment.

Where open models can win

• High-volume routine tasks: classification, extraction, summarisation and internal support can become expensive at premium token prices.
• Data control: self-hosting can reduce the need to send sensitive inputs to an external service, although it creates new security duties.
• Customisation: teams can tune, constrain or optimise a model for specialist language and processes.
• Continuity: model weights can reduce dependence on one provider’s pricing, availability or policy.
• Edge operation: smaller models can run on devices, local servers or disconnected environments.

Where closed systems can win

• Frontier performance: complex reasoning, multimodal work and sophisticated tool use may arrive first in managed models.
• Speed of adoption: an API can be easier than building and maintaining inference infrastructure.
• Integrated tools: search, code execution, agents and enterprise controls may come as one service.
• Operational support: contracts, monitoring and service guarantees can simplify responsibility.
• Fast-changing capability: customers receive improvements without scheduling their own model upgrades.

The model router becomes the control centre

Instead of asking employees to choose manually every time, businesses can create a routing layer. The router examines task sensitivity, required capability, cost, latency and output risk, then sends the request to an approved model.

1. Public and low-risk tasks can use a lower-cost model.
2. Confidential routine work can run on an approved private deployment.
3. Complex expert tasks can move to a stronger managed model with additional review.
4. Prohibited data should not enter any model until governance changes.
5. High-impact decisions require human authority regardless of which model produced the analysis.

Cost is more than token price

A downloadable model is not free to operate. Businesses pay for accelerators, energy, engineers, monitoring, security, evaluation and updates. A proprietary API can also create hidden costs through repeated prompts, poor routing and vendor lock-in.

Total cost should include failure. A cheaper model that produces more errors may create expensive verification work. A premium model used for simple extraction may waste money without adding value.

Sovereignty without fantasy

Running a model internally can improve control, but self-hosting does not automatically create safety. The organisation becomes responsible for patches, access, logging, data retention and incident response. Sovereignty means accepting responsibility, not merely owning files.

This aligns with the growing infrastructure nature of AI. Model choice connects directly to chips, memory, energy and operational skill.

A practical portfolio decision

Start with a workload inventory. Group tasks by value and risk. Test multiple models using real examples and consistent scoring. Record accuracy, cost, speed and human correction time. Then design routing rules with a review date, because model capability changes quickly.

The King Flow distinction remains important: a business workflow can recursively evaluate and route model outputs without claiming access to the engineering substrate inside those models.

The winning AI stack will not worship one model. It will govern many capabilities around one business purpose.

Sources and date note

This article reflects information available on 6 September 2026. See current reporting on corporate adoption of open models and AT&T’s open-weight strategy. Reported percentages and savings should be treated as company-specific snapshots.