The Gap Between Capability and Adoption
For years, the narrative surrounding artificial intelligence has been dominated by capability. Industry headlines focus on which model scores higher on specific benchmarks, which architecture offers a larger context window, or how a new release improves human evaluation metrics by marginal points. While this information is not useless, it has very little impact on whether developers can actually adopt these models into their daily workflows.
Brendan O’Leary, Founding Developer Relations Engineer at Kilo Code, presents a different reality based on real-world usage data. Drawing from over 40 trillion tokens processed through his organization’s gateway, the evidence suggests that successful adoption depends on building trust rather than chasing technical superiority. Teams have been observed accessing the best available models with the highest benchmark scores but choosing not to use them. Conversely, other teams build deep dependencies on tools that may rank lower on public leaderboards because they work reliably for specific use cases.
Refer & earn: 30% of a client’s first month, then 10% of every month after that, for up to 24 months per client. See the partner programme →
This gap is not random. It is structural. For business leaders deciding how to implement AI agents, the lesson is clear: capability does not equal utility. If the foundation of the interaction is shaky, the upper rungs of adoption cannot be reached.
The Four Rungs of the AI Adoption Ladder
O’Leary describes a “trust ladder” that developers climb as they integrate AI into their work. Unlike traditional software tools, which are deterministic, AI systems are non-deterministic. This fundamental difference creates a unique barrier to entry. The ladder consists of four stages:
- Auto-complete: The initial stage where the system predicts single lines or snippets of code.
- Chat with Agents: Moving beyond prediction to interactive dialogue where the AI assists with reasoning or explanation.
- Code Agents: Tools that can autonomously execute tasks, such as editing files or running tests within a controlled environment.
- Orchestrated Multi-Agent Systems: Complex workflows where multiple specialized agents coordinate to achieve broader engineering goals.
The critical insight from the data is that these rungs are sequential. If the lower rungs are unstable—if suggestions are slow, inaccurate, or if the tool finds the wrong file and edits the wrong thing—trust is broken. When trust is broken, developers do not attempt to climb to the higher levels of autonomy. They revert to manual processes or abandon the tool entirely. Therefore, organizations cannot skip steps. You cannot expect your workforce to utilize complex multi-agent orchestration if they do not trust the basic autocomplete features.
Where Adoption Predictably Breaks
Understanding why trust fails requires looking at three specific technical failure points identified in the data. These are not theoretical risks; they are consistent patterns observed in production environments.
1. Context Construction
The first point of failure is how context is constructed for the AI. In agentic workflows, the model needs relevant information to make correct decisions. If the system provides too much irrelevant noise, the model may become confused or distracted. If it provides too little, it lacks the necessary background to act accurately. Poor context construction leads to hallucinations or incorrect actions, which immediately erodes user confidence. Engineering leaders must audit how their systems gather and present context to ensure it is precise and relevant to the task at hand.
2. Model Routing
The second failure point is multi-model routing. As enterprises scale, they often use different models for different tasks—using a smaller, faster model for simple queries and a larger, more capable model for complex reasoning. However, if the routing logic is flawed, a simple task might be sent to an expensive, slow model, or a complex task might be mishandled by a simpler one. This mismatch results in poor performance and unexpected costs. The data shows that developers notice when the system chooses the “wrong” model for the job, leading to frustration. Effective routing requires not just technical selection, but an understanding of the user’s expectation for speed versus accuracy.
3. Feedback Loops
The third and perhaps most critical area is the availability of feedback loops. In traditional software, errors are immediate and obvious. With AI, signals can be sparse and delayed. If a developer makes a request and receives a result that is subtly wrong, they may not realize it until hours later. Without immediate, clear feedback mechanisms, users cannot calibrate their trust. The higher up the adoption ladder a user goes, the sparser these signals become. Designing systems that provide clear, timely feedback on the AI’s reasoning and outcomes is essential for maintaining engagement.
Measuring Trust Beyond Vibe
Trust is often described as having a “vibe feeling,” but O’Leary argues that it can be measured. Instrumentation is possible beyond subjective feelings. By tracking metrics related to latency, error rates, correction frequencies, and task completion success, organizations can quantify trust. For example, if a team frequently overrides AI suggestions or reverts changes made by an agent, that is a measurable indicator of low trust.
This measurement capability has significant design implications. Systems should be built to capture these signals automatically. Instead of relying on annual surveys or anecdotal evidence, engineering leaders can monitor real-time data to identify where trust is breaking down. This allows for proactive adjustments to context, routing, or feedback mechanisms before adoption stalls.
Implications for Business Leaders
While the data comes from the domain of software development, the principles apply broadly to any knowledge worker leveraging AI. Whether your workforce involves marketing, legal, finance, or operations, the dynamics of trust remain similar. Non-deterministic tools require a foundation of reliability before autonomy can be granted.
Businesses looking to scale AI agents should consider the following actions:
- Assess Your Foundation: Before deploying complex agents, ensure that basic AI features (like autocomplete or chat) are reliable and fast. If the base layer is weak, higher-level automation will fail.
- Audit Context Strategies: Review how your AI systems gather information. Are they providing the right amount of context? Is it relevant?
- Optimize Routing: Ensure that your model selection logic aligns with user expectations for speed and accuracy. Mismatches here lead to perceived incompetence.
- Instrument for Trust: Implement monitoring that tracks user interactions with AI. Look for signs of friction, such as high override rates or delayed corrections.
The industry narrative often pushes for rapid adoption of the most advanced models. However, the data from 40 trillion tokens tells a different story. Adoption is a gradual process built on trust. By focusing on the technical pillars that support trust—context, routing, and feedback—organizations can create sustainable AI workflows that deliver real value.
Faida Workforce helps organizations navigate these transitions by aligning workforce capabilities with technological realities. Our approach focuses on practical integration rather than hype, ensuring that teams can effectively leverage AI tools without compromising on trust or reliability.
Sources
What 25 Trillion Tokens Reveal About How Developers Actually Adopt AI Agents
30% of their first month, then 10% every month after that
Refer a business to Faida Workforce and you earn 30% of the client’s first paid month, then 10% of every month for up to 24 months per client — rising to 15% from the accrual after you pass 10 active clients. Paid by PayPal, batched on the 1st.
10 Pro clients = $2,598.70 over 24 months — $3,748.13 at the 15% tier.
Figures are potential earnings worked from the published rates, assuming every referred client pays for the whole cap — not a promise, not an average. The $1 trial is never commissionable.
