We are noticing a phenomenon worth taking seriously: more and more enterprise AI projects are stalling at deployment—not because models aren't powerful enough, but because engineering fundamentals are missing. A tech team's failure post-mortem this week made this crystal clear—LLM output formats went haywire, critical constraints got lost, and the root cause wasn't the model itself, but unstructured prompts, unsanitized user input, and unvalidated outputs.
What This Is
This is a tech team's failure post-mortem. The scenario is textbook: a business team uses an LLM to auto-generate documents with strict word count, fixed format, and a required closing line. The output routinely failed to meet spec—format broken, word count exceeded, closing line missing.
Four root causes surfaced after investigation:
First, the prompt mixed system constraints, user input, and format requirements together, so the model couldn't tell which took priority. Second, no delimiters (placeholders) were used to separate user input from system instructions, allowing user input to directly contaminate system rules. Third, the code layer didn't validate length, so long text triggered the model's context window (the maximum amount of text a model can read at once) truncation, eating the constraints. Fourth, when model output didn't comply, the program accepted it as-is without checks—no retry mechanism.
The fix was unglamorous: split prompts into four sections—"role, task, constraints, user input"—isolate with placeholders, run keyword validation on output, retry if it fails.
Industry View
This case reflects a judgment Andrew Ng has been making repeatedly—90% of Agent (AI assistants capable of autonomously executing multi-step tasks) projects stall at deployment, and the bottleneck usually isn't model intelligence but engineering capability. In other words, today's "AI project failure" blame shouldn't fall on model vendors.
But pushback is worth hearing: enterprise teams will say that model APIs (Application Programming Interfaces) have sloppy error feedback and messy context management, and vendors aren't blameless. One representative view: "prompt engineering" (i.e., designing instructions to make models output reliably) has been overhyped—it's essentially a patch covering up API design flaws, and the real solution is stronger structured-output capabilities from the model side.
Another easily overlooked angle: these fundamentals aren't sexy, so traditional enterprise CTOs (Chief Technology Officers) don't want to do them and won't budget for them. Result: models upgrade generation after generation, but business-side stability never improves.
Impact on Regular People
For enterprise IT leaders: if internal AI projects keep failing, don't upgrade model versions yet—go back and check three lines: prompt structure, input isolation, output validation.
For individual careers: people who can write prompts as "four-part structured templates" will hold a hidden advantage over the next two years—not just product managers and engineers, but operations, marketing, and finance roles too.
For the consumer market: stability issues in C-end (consumer-facing) AI products won't see significant improvement in the short term, dragging down both willingness to pay and retention rates.