Quick intro since I'm fairly new to posting here - I'm Alex, I run an AI automation agency (Nova Brain), we build systems for agencies/consultants/coaches so they're not stuck doing everything manually. Been in the automation/AI space for a bit now, mostly heads down building rather than posting, but wanted to start sharing some of what we're testing. Been sitting on this for a bit, figured I'd share since I know a lot of us are running lean teams and drowning in follow-up. Me and my guys built an AI voice agent that handles our calls, both inbound and outbound. Not a chatbot, an actual voice on the phone. Wasn't expecting much honestly, but it's been carrying its weight: - Calls new leads within seconds of them opting in, before they cool off - Follows up on the ghosted leads we never got back around to - Doesn't get tired doing the same pitch for the 50th time that day - Actually handles pushback instead of freezing or going off script - Filters out the tire-kickers so we only hop on calls with people ready to talk - Books straight onto the calendar, no back and forth over text - Picks up after hours too, so we're not losing leads overnight Still tweaking it, definitely not plug and play, but it's already freed up hours we used to spend chasing dead leads. Anyone else here running voice AI for discovery calls or sales calls? Curious what's actually working vs what's just noise right now.

Hi! Context: I worked with a large knowledge vault (millions of words, thousands of files). Everything organised to the letter. Nothing to mess with. Goal was to transform a part of the data (<40%) using other LLM provider. I trusted Claude Code (Opus 5 High) with a deterministic environment created in detail over months of trial and error. Even a Gemini Flash model could follow these. But Claude recommended to use Opus on High due to the complexity of the vault. The job was simple: - find files based on criteria - extract headers and body - create new documents with from each of the headers - upload to different provider for a specific task - replace older headers with newly transformed data from each of these files - Simply put: extract piece of puzzle, process piece of puzzle, put new piece of puzzle back. Easy. It made an absolute mess and I was dumb enough to believe it: - Invented relationships between files from filenames. - Mixed old and new outputs. - Created broken sequences like L-A and E-B instead of A B C D. Adding L-A before C. But L-A was F. - Duplicated some content and omitted other parts. - Changed a complete rebuild into partial work without disclosure. - Marked its own broken output as verified and complete. - Used misleading counts that could not prove coverage. - Produced confident, invented explanations when challenged. - Minimized mistakes and shifted attention toward the user’s decisions. The instructions were deterministic and clear. Claude’s implementation was chaos. My mistake was trusting the completion report before checking the final files with another LLM provider. After 3 days of forensic work, I still cannot reliably separate good output from plausible-looking invented material. I stoped using Opus 5. Even asking Sonnet 5 to create a coherent document about what it did, it is unable to finish paragraphs without gibberish talk and word salads. And I think this is also because of their watermarking system (ex: using "checks out" instead of "done"). Constantly speaking like a confused GenZ who didn't attend school but wants to look original.