INSIGHTS
How to Choose an AI Model: Executive Framework
The wrong AI model choice costs more than money—it costs adoption, security, and competitive edge. Here's how to match the right LLM to your actual business needs.
The wrong AI model choice costs more than money. It costs adoption. It costs security. It costs competitive edge.
We sat down with one of our Co-Founders to cut through the marketing and get to the actual decision framework. After 1,000+ hours in executive workshops, they start the conversation somewhere most vendors avoid.
1. The Question That Reframes Everything
Ask most consultants how to choose an AI model and you get feature comparisons. Token limits. Context windows. Benchmark scores.
Our Co-Founder asks something else:
"Who's your favorite child? Which person you know is the best?"
The question lands because it's obviously absurd. Different people have different strengths. Different communication styles. Different contexts where they shine. Getting the best from any relationship requires understanding what makes that person tick.
Large Language Models work the same way.
GPT-4 excels at broad knowledge tasks. Claude Opus handles complex reasoning and longer contexts. Gemini integrates deeply with Google Workspace. Kimi K2.5 delivers strong performance at a fraction of the cost. NotebookLM specializes in document analysis and synthesis.
The "best" model depends entirely on what you're trying to accomplish, how your team works, and what constraints you face around data, cost, and technical infrastructure.
The selection process starts with understanding your use case—not comparing feature tables.
2. The M365 Trap (And the Shadow AI Risk)
Here is the mistake our Co-Founder sees most often:
"The most common mistake is they think because Copilot and GPT are in M365, it's the smart or only choice to use. Couldn't be more wrong and risky for their data and teams."
Microsoft's integration is convenient. It's also a default that bypasses critical evaluation. Three problems emerge:
Underutilization. You buy 500 Copilot licenses. Maybe 50 people use them consistently. The rest ignore the tool or use it sporadically. The per-seat model assumes adoption that rarely materializes.
Wrong fit. Copilot works well within Microsoft's ecosystem. If your team lives in Slack, Notion, Figma, or specialized industry tools, the integration value drops significantly.
Shadow AI proliferation. This is the security risk nobody talks about:
"If you don't give people any education or tools, they are putting their employment issues and your company info into a free version of Gemini or GPT at home."
When the sanctioned tool doesn't meet needs, employees find alternatives. They paste proprietary data into free versions of ChatGPT, Claude, or Gemini—tools with data retention policies they haven't read, hosted on infrastructure they don't control. Research from Cyberhaven (2024) found that 74% of organizations experienced potential data exposure from employee AI use, with sensitive information flowing to unmanaged AI tools.
The M365 default feels safe because it's familiar. The reality is more complex.
3. Same Problem, Different Solutions
Our Co-Founder shared a comparison that illustrates the context-dependence of model selection:
"The smaller creative business looking to build a chart or infographic based on materials quickly vs an enterprise exec doing the same thing. Smaller business would use NotebookLM and the bigger company would use Claude Opus 4.6 or Claude in PowerPoint."
Same output. Completely different infrastructure, cost structure, security posture, and workflow integration.
The creative business needs:
- Quick synthesis of source materials
- Visual output without enterprise IT involvement
- Minimal cost for occasional use
- No security review process
NotebookLM fits perfectly. Google hosts it. The business uploads documents, gets visual summaries, moves on. No API keys. No infrastructure. No procurement.
The enterprise executive needs:
- Integration with existing PowerPoint workflows
- Security review and compliance alignment
- Consistent output formatting for executive consumption
- IT support and troubleshooting
Claude Opus through PowerPoint integration fits here. The infrastructure exists. The security team has evaluated it. The workflow matches how work actually happens.
Neither choice is wrong. Neither choice is universally right.
The professional services firm building client presentations might need yet another approach—API-based access to multiple models depending on the specific task, with custom guardrails and output formatting.
4. The Cost Conversation (And Why It's Wrong)
Most executives don't understand AI pricing because it's abstract:
"They don't understand the currency as it's just so abstract."
Per-token pricing. Per-million-tokens pricing. Input vs. output costs. Context window premiums. The unit economics feel foreign compared to familiar software licensing.
Our Co-Founder's framework reframes the discussion entirely:
"We suggest and use APIs for our agents and wrappers in internal chatbots and solutions as it's MUCH cheaper to make specific wrappers with guardrails and specific prompts than it is to just get a workforce a Copilot package few will use, and nobody will use consistently."
The math works like this:
Option A: Workforce Copilot licenses
- 500 seats at $30/month = $15,000/month = $180,000/year
- Adoption rate: 20% (generous)
- Effective cost per active user: $150/month
- Output quality: Variable (depends on prompt skill)
- Consistency: Low (everyone prompts differently)
Option B: API-based solution with guardrails
- Development: $40,000 one-time
- Monthly API costs: $2,000-$5,000 (depending on volume)
- First-year total: $64,000-$100,000
- Adoption rate: 100% (embedded in workflow)
- Output quality: Controlled (standardized prompts)
- Consistency: High (same guardrails for everyone)
The API approach costs less and delivers more consistent results. It requires upfront development investment and ongoing technical support—trade-offs that don't fit every organization.
The broader principle:
"In AI, costs shouldn't be considered as absolutes, they are value-priced against time savings and output improvement."
A $500/month API bill that saves 40 hours of analyst time delivers $8,000+ in value (at $200/hour fully loaded cost). The same $500 spent on unused Copilot licenses delivers zero.
Cost per token matters less than cost per outcome.
5. The Professional Services Starting Point
For the $25M+ service company with limited technical infrastructure, our Co-Founder recommends a phased approach:
"I'd start with the M365 or Gemini in their workspace and take that as far as they can, with little upfront costs. From there, build what you need using the best model for the use case on an API / on-demand basis (if you have somebody around to do that)."
Phase 1: Use what's already available.
- Microsoft Copilot (if already in M365)
- Google Gemini (if in Google Workspace)
- Native AI features in existing tools
This phase builds organizational AI literacy without new procurement. Teams learn what works and what doesn't. Use cases emerge organically. Infrastructure gaps become visible.
Phase 2: API-based expansion.
- Custom wrappers for high-value use cases
- Integration with existing workflows
- Guardrails and output standardization
- Multi-model approach (different models for different tasks)
This phase requires technical capability—internal or contracted. The investment targets specific, validated use cases rather than hoping workforce-wide licenses get adopted.
Phase 3: Strategic infrastructure decisions.
For organizations with data sovereignty requirements or scale needs, the open source vs. closed model decision becomes relevant:
"I've seen banks and major companies do very well with open source models housed behind a firewall, and I've seen poor results from closed models. The closed ones like Claude Opus are outstanding, but Kimi K2.5 is a personal favourite and is a fraction of the cost with a different skill set."
Closed models (Claude, GPT-4, Gemini):
- Easier to implement
- Better documentation and support
- Higher per-token costs
- Data leaves your infrastructure
Open source models (Llama, Mistral, hosted internally):
- Full data control
- Lower per-token costs at scale
- Require technical infrastructure
- Support burden falls on your team
Kimi K2.5 (closed but cost-effective):
- Strong performance on reasoning tasks
- Significantly lower cost than Claude Opus
- Good balance of capability and price
- Less ecosystem support than major players
The right choice depends on your technical sophistication, data sensitivity, scale, and budget.
The Decision Framework
Choosing the right AI model requires answering five questions:
1. What's the actual use case? Not "we need AI." What specific task? What output format? What quality standard? What volume?
2. Where does work happen? Microsoft shop? Google Workspace? Mix of specialized tools? The answer determines integration priorities.
3. What's your data sensitivity? Can it leave your infrastructure? Does it need to? Regulatory requirements?
4. What technical capability do you have? API integration requires different skills than buying licenses. Be honest about what your team can support.
5. What are you actually trying to improve? Time savings? Output quality? Consistency? Scale? Different models excel at different outcomes.
The framework isn't about finding the "best" model. It's about finding the right model for your specific context—and being willing to use different models for different problems.
Aspiro AI Studio helps professional services firms select and implement the right AI models for their specific use cases—from initial evaluation through API integration and custom deployment.
Not sure which model fits your needs? Try our Use Case Prioritizer to identify your highest-leverage AI opportunities.
Related Reading:
- How Can AI Benefit My Business? — The foundational benefits of AI implementation
- What Challenges Might I Face Building an AI Innovation Lab? — Organizational considerations for AI adoption
- Azure OpenAI vs Copilot Studio — Microsoft's AI options compared
- The Real Cost of AI Implementation — Understanding true AI investment requirements
- How to Implement AI: A CEO's Playbook — From use case audit through pilot
About the Author: Issy is the AI Integrator at Aspiro AI Studio. She translates visionary strategy into executable reality—and writes about what actually works.
References:
- Cyberhaven (2024). "74% of Companies Experience Data Exposure from AI Use." Retrieved from https://www.cyberhaven.com/blog/74-percent-of-companies-experience-data-exposure-from-ai-use
- Anthropic. "Claude Models: Capabilities and Context Windows." Retrieved from https://www.anthropic.com/claude
- OpenAI. "GPT-4: Model Overview and API Documentation." Retrieved from https://openai.com/gpt-4
- Google. "Gemini for Workspace: Enterprise AI Integration." Retrieved from https://workspace.google.com/products/gemini/