I'm looking at Perplexity for maybe generating summary reports for our clients. We use Salesforce and Zendesk, so the data's there, but building custom dashboards is slow.
Has anyone actually put Perplexity into a production workflow for external, customer-facing stuff? I'm skeptical about accuracy. What happens when it hallucinates a number in a client report? How are you handling citations or grounding the data?
Your skepticism about hallucinations in customer-facing reports is absolutely the right place to start. I've seen a few teams try this, and the ones who make it work treat the LLM purely as a drafting assistant for a human analyst. They'll have it generate narrative summaries from structured data points, but every single figure and claim is manually verified against the source system before the report goes out the door.
That process can be faster than building a full dashboard, but it's not a fully automated pipeline. For grounding, one approach is to only feed it very clean, pre-aggregated data summaries and instruct it to only comment on what you provided. Even then, you need a rigorous human sign-off step. The moment a client spots one invented number, trust in the entire service evaporates.
Stay curious.
Yep, "drafting assistant" is the only viable mode right now. The human verification step you mention is critical, but it also becomes the bottleneck. You're just shifting the labor from building the dashboard to fact-checking the AI's output.
I've seen teams get lazy with that verification, because the draft looks so polished. Then you get a beautifully written paragraph confidently stating that Q3 revenue was "approximately eleventy billion dollars." Good luck explaining that to a client.
If the process requires a human to manually verify every figure against the source system anyway, you have to ask if having the AI write the English sentences around those numbers is really saving any meaningful time.
CRM is a necessary evil