When I was learning to code, and new errors were popping up on the screen repeatedly, my instructor used to joke, “Computers do what we tell them to do.” Several years later, while integrating AI into various business workflows and working closely with a tech investor who has hired thousands of people, they explained why many projects fail in similarly simple terms: “It’s not the tech, it’s the execution.”

As we advance AI development now in 2026, the frequency of mistakes made by the models themselves is becoming much less of an obstacle for AI projects. If you want your own personal AI, modern off‑the‑shelf models now run locally on $3,000 consumer grade hardware and reach around 85-90% of the quality of models like ChatGPT or Gemini. In another six months it’s projected to be another 5‑10% better because the open source community redesigns the large models into smaller, local versions as new updates come out. So for business right now the primary source of AI failures is our own misunderstanding, and organizations often misapply powerful AI tools, creating organizational, and execution related obstacles that can’t be solved by more compute power or bigger models.

So let’s examine what works and what doesn’t in terms of moving past proof of concept and seeing real value from the science of AI.

Why AI Projects Lose Momentum

Misaligned Scope & Over‑Engineering - The intuition behind deciding to use generative AI is almost always right. But unfortunately we can’t just toss AI at the whole organization and see huge ROI. Successful projects start with a clear business problem and apply AI as a modern, adaptable solution where it makes sense. Clear communication and deliberate alignment between those who understand the business problems and those who understand AI mechanics is essential.

The Promise of Agentic AI – Agentic AI is popular, but agents should be created and applied through a deliberate process. The landscape is filled with businesses who offer to create agents because agentic AI is trending. An entry‑level AI engineer can make agents all day long because the technology is that good, especially with cloud solutions like Azure and GCP. But you should be considering: Can your team clearly explain why you need those agents, and do your agents fit your use case? Agents follow instructions and behind the scenes they work based on data, language, code, and math. Turn‑key, on‑demand agents can be effective when the problem is well defined and the data pipeline is solid. But if your underlying problems, or your agents themselves, aren’t well-defined then the promise of agentic AI is just talk. The trick to future proofing your business for this type of tech advancement is to have problem solvers on hand capable of adapting your specific business problems to new tech as it arrives.

Replacement vs. Automation - Beyond the ethical concerns of fully replacing people with AI, full human replacement projects have low success rates. Successful AI implementations augment rather than replace, automating repetitive work while people focus on judgment, creativity, and complex problem solving. For example, our AI assistant for preschool teachers handles the administrative burden—observation tracking, parent updates, and planning—while providing developmental analysis, giving teachers hours of extra time while they maintain full control over classroom decisions and student interactions.

The Wrapper – Recent years saw businesses sometimes paying consultants to basically build quick wrappers for large language models like ChatGPT. This can often end up being less private, more expensive, and even less accurate than using ChatGPT directly, for example. Some companies also may use large, expensive models (Claude Opus, GPT‑4) for tasks that could run on cheaper models or even smaller, local models running completely offline.

Data Architecture Bottlenecks - Unfortunately, traditional data architecture techniques don’t always support modern ETL/ELT and MLOps workflows. This creates situations where data infrastructure updates potentially slow progress. Recently, companies have been adding DevOps responsibilities to Data Scientist roles; sometimes this works, sometimes it doesn’t. A lot of projects get stuck in the pilot phase because:

• The DevOps and Data Engineering teams have not yet fully updated the architecture for data science and AI, which might involve learning new processes and tools.

• The Data Scientists and Data Engineers are expected to handle infrastructure management that wasn’t part of their training.

• Teams are sometimes working separately instead of collaborating, and some people may resist potentially adapting their role to a new and different workflow.

Data Quality - AI is built on machine learning, and machine learning learns from data. The data is the entire picture of the world the model knows, so AI struggles depending on how data is presented and how fragmented or chaotic it is.

Data drift: Real‑world data changes so a business might deploy AI that performs well initially, but three months later performance has seriously degraded because the model doesn’t account for data distribution shifts over time. You can’t just use whatever data is lying around. Human monitoring and retraining of these models is necessary and can be expensive as well.

Pre-Launch Checklist

You can achieve the value and capability of responsible, trustworthy, and useful AI. Before committing resources to an AI project, consider how confidently you can check these boxes:

• Problem definition: Can you clearly state the specific business problems? Can you define measurable success criteria (e.g., time saved, clients served, % lift in conversion, churn reduction)?

• Data readiness: Is your data accessible (how is it stored, what format is it in, how is it processed, how are permissions structured, does it use APIs)? Is it representative of the intended group?

• Skill alignment: Are cross‑functional teams communicating effectively (regular syncs, shared docs, collaboration)? Does the team possess the required skills for your use case whether it’s machine learning, natural language processing, data architecture engineering, domain expertise, or something else?

• Cost‑benefit analysis: Have you calculated total cost of ownership (infrastructure, training, monitoring, retraining)? Does the projected ROI justify the investment?

• Timeline & risk overview: Do you have realistic milestones?

  
  

Moving Forward

The main difference between AI projects not meeting expectations and those that succeed is not as much the sophistication of the technology anymore. Plus, over the next several years AI will become even smarter and easier to use for all people. To succeed now and in the future, leaders and teams should focus on:

• Clarity of strategy

• Alignment and communication (of domain and AI expertise in your organization)

• Quality of data and infrastructure

• Skillful, strategic execution of tools

The science of AI can deliver transformative value when applied carefully to well‑defined problems, with proper infrastructure and realistic expectations. The question isn’t whether to adopt AI, but how to do it strategically. Start small, measure precisely, and scale what works. When reaching outside for assistance, prioritize partners who can give you clear problem definition and realistic milestones over buzzwords and hype.

Contact us at hello@codachord.com to discuss your next steps and advance with the science of AI.

- Ben Lasko, Founder & CEO of Codachord