Chapter 17: Support in the AI Era
I want to close this section with some observations about where support is heading as AI continues to reshape the landscape. The honest answer is that AI is compressing the support lifecycle just as aggressively as it’s compressing the build lifecycle.
Monitoring and alerting systems are getting smarter. We’re approaching a point where AI can not only detect that something is wrong, but diagnose why and propose a fix — sometimes implement it autonomously. The Operational Runbook, which we currently structure as “if this, then that” scenarios for human consumption, could become an instruction set that an AI agent executes directly. Server goes down? The agent reads the runbook, follows the recovery procedure, validates the system is back up, and notifies the team. Dependency gets deprecated? The agent identifies the affected code paths, proposes the migration, runs the test suite, and opens a pull request for review.
We’re not fully there yet, but we’re close enough that our support processes need to be built with this trajectory in mind. That’s actually one of the reasons we invest so heavily in documentation and the Spec Tree during delivery. These assets aren’t just for human developers — they’re for the AI systems that will increasingly be doing the maintenance work. A well-structured Spec Tree with clear acceptance criteria is as valuable to an AI support agent as it is to a human one. Arguably more so, because the AI can consume and cross-reference the entire tree instantaneously.
What this means practically is that the economics of support are shifting. The 25% model may compress over time as AI handles more of the reactive maintenance autonomously. The value of the support contract shifts from “we have people who can fix things” to “we have systems — powered by AI and backed by people — that keep your software healthy.” The human element doesn’t disappear, but it moves further upstream into oversight, judgment calls, and the kinds of strategic decisions that AI isn’t ready to make independently.
For the FTE model, the implications are similar. A placed developer with AI tools can manage a significantly larger and more complex system than they could have pre-AI. The leverage of a single developer has increased dramatically. This is good for clients because they get more value per dollar. It’s good for the placed developer because the work is more interesting — less time debugging CSS, more time making architectural decisions and guiding AI agents. And it’s good for us because we can serve more clients with a lean, highly capable team.
The Future of AI-Driven Support
TODAY NEAR FUTURE ┌───────────────────────────┐ ┌───────────────────────────┐ │ │ │ │ │ Issue detected │ │ Issue detected │ │ (monitoring alert) │ │ (AI monitoring) │ │ │ │ │ └─────────────┬─────────────┘ └─────────────┬─────────────┘ │ │ ▼ ▼ ┌───────────────────────────┐ ┌───────────────────────────┐ │ │ │ │ │ Human investigates │ │ AI reads Runbook │ │ & diagnoses │ │ & diagnoses │ │ │ │ │ └─────────────┬─────────────┘ └─────────────┬─────────────┘ │ │ ▼ ▼ ┌───────────────────────────┐ ┌───────────────────────────┐ │ │ │ │ │ Human implements fix │ │ AI implements fix │ │ with AI tools │ │ autonomously │ │ │ │ │ └─────────────┬─────────────┘ └─────────────┬─────────────┘ │ │ ▼ ▼ ┌───────────────────────────┐ ┌───────────────────────────┐ │ │ │ │ │ Human validates │ │ AI validates via │ │ & deploys │ │ Spec Tree + tests │ │ │ │ │ └───────────────────────────┘ └─────────────┬─────────────┘ │ ▼ ┌───────────────────────────┐ │ │ │ Human notified. │ │ Reviews if needed. │ │ │ └───────────────────────────┘ The Spec Tree becomes the CONSTITUTION that AI agents govern by. The Runbook becomes their PLAYBOOK.
The end state, if I’m being speculative, is that support becomes largely autonomous — with human oversight reserved for escalation, strategic decisions, and novel problems that AI hasn’t encountered before. The Spec Tree becomes the constitution that AI agents govern by. The Operational Runbook becomes their playbook. And the FDE who built the system originally might get a call once a quarter when something genuinely unusual happens. Until then, the ship runs itself.