We've been integrating AI coding tools into our Salesforce Industries and Vlocity work for a while now. Here's what we actually found — the wins, the surprises, and the limits.
Unfiltered observations from using AI coding assistants on real Salesforce Industries projects.
DataRaptor definitions, OmniScript JSON skeletons, Integration Procedure templates — the kind of repetitive structure that eats hours. AI generates a solid first draft in seconds. We still review and adjust every line, but the blank-page problem is gone. On a recent B2B CPQ project, we estimated a full day saved per developer per week on this alone.
Paste a Vlocity error log and a description of the flow into an AI — you get a targeted explanation and candidate fixes faster than searching community forums. Not always right, but the hit rate is high enough to make it the first stop.
Nobody wants to document a 200-step Integration Procedure. AI turns a pasted JSON config into readable prose in under a minute. Documentation coverage on our projects has improved significantly since we made this a standard step.
General Apex, SOQL, and JavaScript — AI is strong. Vlocity-specific objects, naming conventions, and version-specific quirks — hit and miss. It works best when a senior dev frames the question with enough context. Don't expect it to know the difference between VLO12 and VLO15 governor limit behaviours without being told.
Before opening a PR, paste the diff and ask the AI to review for logic errors, security issues, and performance problems. It catches things — not everything, but enough to reduce the burden on human reviewers and improve what lands in the PR.
AI can describe patterns. It cannot weigh the political, operational, and historical context of a troubled implementation. In our project recovery work especially, the judgment calls that matter — what to rewrite vs. rescue, what tech debt to accept — remain firmly human decisions. AI is a tool, not a Triarius.
Describing a slow pricing flow to an AI and asking for a root cause analysis surfaces hypotheses quickly. On one engagement, an AI suggested checking a DataRaptor Extract filter that was doing a full-table scan — verified within minutes, resolved within an hour. We would have found it, but it took a fraction of the time.
AI occasionally invents method signatures or Apex classes that don't exist. On standard platforms this is manageable — you compile and find out quickly. In declarative Vlocity config it can be harder to spot. Every AI suggestion that references a specific API or object field needs verification. This is not optional.
The biggest mistake teams make is treating AI as a replacement for experience. In our work, the correct model is a multiplier: a senior Vlocity architect using AI tools produces more, faster, with higher quality than either the architect alone or the AI alone. The AI handles volume and speed; the human handles judgment and verification.
This is exactly the TriarIT principle applied to tooling. You still need the veterans. You just arm them better.
The tools we actually use day-to-day, and what we use them for.
Codex runs as a cloud agent that can read repositories, write and execute code, browse documentation, and open pull requests — all without leaving a browser tab. We use it for exploratory tasks and rapid prototyping where spinning up a local environment would slow us down.
Claude Code lives in the terminal alongside the codebase, giving it full context of the project structure, git history, and running processes. We use it for deep, multi-file work — analysing large Vlocity metadata exports, performing codebase-wide refactors, and reviewing complex Integration Procedure chains end-to-end.
Where AI fits — and where it doesn't — in a TriarIT engagement.
Human-led. AI helps summarise large volumes of existing documentation, metadata exports, and log files — turning a week of reading into a day of review.
AI role: supportHuman-led. AI is used to stress-test designs by asking it to find weaknesses, or to quickly prototype an alternative approach in prose before committing to implementation.
AI role: sounding boardShared. AI generates scaffolding, completes boilerplate, and suggests fixes during debugging. All output is reviewed by a senior developer before it enters version control.
AI role: active co-developerAI performs a first-pass review of every PR — checking for logic errors, missing null checks, and performance anti-patterns. Human reviewers focus on architecture and intent.
AI role: first reviewerAI generates test class scaffolding and edge-case scenarios from spec documents. Execution, interpretation, and sign-off remain human.
AI role: test generationAI drafts technical documentation from code and configuration. The result is reviewed and corrected by the developer who built it — a much faster loop than writing from scratch.
AI role: primary drafterWe're happy to share more of what we've learned — or to bring these practices to your project.