TechAmerica
SearchSubscribe
AI decision model customization
Image credit Sqwish

AI Decision Model Customization Gains Momentum as Usage Surges

AI decision models are gaining adoption as developers seek faster, cheaper alternatives to conventional generative AI for structured decisions. More than 25% of requests through Vercel’s AI model router went to decision models in October. UK startup Sqwish has raised $3 million to develop customizable models, while Microsoft has expanded access to its Decision-1.1 system. Growing competition is shifting attention toward accuracy, confidence calibration, and customer-specific training.

AI decision model customization is emerging as a priority for developers building faster, more reliable software automation. Decision models are gaining adoption because they can select predefined actions without generating lengthy responses.

In October 2026, more than 25% of requests through Vercel’s AI model router went to decision models. Meanwhile, Microsoft has expanded its decision-model offering, and UK startup Sqwish has raised $3 million to develop models that improve using customer data.

AI Decision Model Customization Gains Importance

Decision models are specialized AI systems designed to choose between predefined options and estimate confidence in their selections.

Unlike conventional generative AI, which produces text responses, these models focus on structured decisions such as classification, request routing and workflow control.

The approach can reduce processing costs and response times when software must evaluate large numbers of routine decisions.

However, developers increasingly face another challenge: adapting general-purpose decision models to individual application requirements.

That involves training models on relevant operational data, improving confidence calibration and testing how reliably they perform in production.

Decision Model Usage Surges Across AI Platforms

Decision models accounted for more than one-quarter of requests routed through Vercel’s AI model infrastructure during October.

The rapid growth follows the September introduction of TypeSafe AI’s Jev models, which helped establish the category.

According to OpenRouter’s decision model rankings, Jev 1.13 and Jev 1.14 together processed approximately 1.56 billion requests in the reporting period ending November 17.

Other models from Perplexity, Liquid AI and OpenAI also appeared among the most-used systems.

These figures indicate substantial developer activity, although platform-specific request volumes do not measure the entire AI market.

TechAmerica previously reported on Jev’s rapid adoption through Vercel AI Gateway, including its early use among paying developer teams.

Sqwish Raises $3 Million to Develop Custom Decision Models

UK-based AI startup Sqwish announced a $3 million pre-seed funding round on November 17.

Co-founders Ushnish Sengupta and Federica Freddi previously worked on smaller language models before adapting their research to decision-focused systems.

The company’s kettei family of models is designed for classification, routing and other structured decisions.

According to the Sqwish website, its platform allows customers to fine-tune models using their own data.

Rather than relying exclusively on general training, the approach incorporates examples relevant to specific applications.

For instance, customer support teams could use historical escalation decisions to improve how a model routes new requests.

Sqwish also describes continual learning, in which outcomes from production systems contribute to subsequent model training.

The company reports improvements in accuracy, cost and response times across selected evaluations. Those figures remain company-reported results rather than universal performance guarantees.

Microsoft Expands Decision-1.1 Availability

Microsoft announced the general availability of Microsoft-Decision-1.1 on November 17, expanding competition in specialized AI models.

According to Microsoft’s technical announcement, the model is available through Microsoft Foundry, OpenRouter and Vercel.

It supports structured decision-making for classification, prioritization, routing and workflow management.

Microsoft says Decision-1.1 is more than twice as fast as its first decision model. The updated system also supports image inputs, including photographs and screenshots.

The company reports that more than 4,000 customers used its decision-model technology following the earlier public preview.

Microsoft emphasizes confidence calibration, allowing applications to defer decisions when uncertainty exceeds acceptable limits.

TechAmerica has also covered the broader competition among decision models, including their use in automation and agent monitoring.

Why Confidence Calibration Matters More Than Rankings

Decision models must do more than select the correct answer frequently. Their confidence estimates also need to reflect how likely they are to be correct.

For example, a model that reports 90% confidence should ideally be correct approximately 90% of the time across comparable decisions.

When confidence estimates are poorly calibrated, software may act on incorrect predictions without recognizing the uncertainty involved.

This becomes particularly important when AI systems control financial processes, customer communications or automated workflows.

Developers can address these risks by evaluating predictions against representative data and establishing thresholds for human review.

Custom training can improve performance on specialized tasks, but it may also introduce overfitting or reduce reliability when conditions change.

Consequently, production evaluations are essential even when a model performs well on public benchmarks.

Smaller Decision Models Could Work Alongside Larger AI Systems

Specialized decision models are increasingly positioned as companions to larger reasoning models rather than direct replacements.

A larger model might interpret a complicated problem or develop a detailed plan. A smaller decision model could then select routine actions during execution.

This division resembles the distinction between fast, intuitive decisions and slower, more deliberate reasoning.

Developers sometimes describe these approaches as System One and System Two intelligence.

For software agents, combining both approaches could reduce the cost of repeated decisions while preserving access to more capable reasoning when necessary.

Customization Could Define the Next Stage of Decision AI

Decision models have gained attention through their speed, low operating costs and ability to return structured answers.

As more developers introduce competing systems, those characteristics may become less distinctive.

The next challenge is demonstrating whether models can adapt to specialized workflows without sacrificing reliability.

Training on customer-specific data, maintaining calibrated confidence estimates and evaluating production outcomes will be increasingly important.

The growing adoption of decision models suggests an expanding role for specialized AI within software. Their long-term value will depend on how accurately they handle real decisions, not simply how quickly they respond.

Shivangi Yadav

Shivangi Yadav is a technology writer at TechAmerica.ai, covering artificial intelligence, startups, digital platforms, consumer technology, mobility, and emerging technologies. Her reporting follows major developments across the global technology industry… All stories by Shivangi

Related articles

More on artificial intelligence, AI Decision Models, AI Model Customization.

See More - All AI stories

Newsletters

The day's tech news, in five minutes

Written by the reporters who cover each beat. No sponsored stories, and you can unsubscribe in one click.

Send me

By subscribing you agree to our Terms and Privacy policy.

Thank you

Your newsletter subscription has been received. Thank you for subscribing to TechAmerica.ai.