Design systems that can be operated by humans, AI agents, or a combination. This prevents projects from failing due to over-automation or requiring a complete refactor when human intervention is needed, ensuring flexibility and saving future development costs.
Reframe hallucinations as signals of poor data quality or retrieval, not model failures. The AI is improvising because you failed to provide the correct script—the authoritative information, or 'canon.' This shifts focus from blaming the model to fixing your data pipeline.
Counter the hype by following a clear progression: Skills -> Workflows -> Agents. If you cannot create a reliable, deterministic workflow with a predefined path, an autonomous agent attempting to improvise will almost certainly fail. This structured approach mitigates risk and ensures reliability.
Historically, the 'build' phase was the primary bottleneck in software development. With AI making building nearly instantaneous, the critical path to success has shifted. Mastery of the 'define' (scoping) and 'feedback' (learning) stages is now what separates winning teams from the rest.
When all competitors can access the same frontier models, the technology itself is not a sustainable advantage. The only durable moat is velocity: how fast your organization can iterate through the build-measure-learn-improve loop. The fastest team to learn and adapt wins.
After struggling to improve an AI system, a team replaced it with simple canned statements and a decision tree. User complaints vanished overnight. This reveals a critical, counterintuitive skill for AI engineers: recognizing when a non-AI solution is superior and having the courage to implement it.
Critical AI context shouldn't be buried in a GitHub repo managed by engineers. Instead, create a dedicated 'Canon Manager' role. This subject-matter expert is responsible for maintaining the authoritative knowledge base ('canon') that AI systems rely on, ensuring accuracy and proper governance.
Unlike traditional software, AI prototypes can be built almost instantly. This requires a mindset shift: if a project doesn't demonstrate tangible value on its very first day, it should be abandoned immediately. Sticking with a weak AI concept leads to costly slow failure.
Bloomberg spent eight figures on BloombergGPT, only for GPT-4 to make it obsolete weeks later. This is a cautionary tale: the high cost, maintenance, and opportunity cost of fine-tuning often outweigh marginal performance gains, especially as foundation models advance relentlessly. Most teams should avoid it.