Notes on making AI actually pay for itself.
Working notes on the parts of applied AI that decide whether it earns money or just costs it — what an agentic pipeline does with a Zendesk queue, what a voice agent has to get right beyond the voice, grounding retrieval in a live catalogue, the legal basis for B2B prospecting in Greece, what a consolidated software stack really costs, and which no-show interventions survive contact with evidence.
How a Swedish Shopify brand resolved 70% of its Zendesk tickets
A maker, a checker and a person with the final say: what an agentic support pipeline does with a Zendesk queue, why it costs about €0.30 a ticket, and which thirty per cent it deliberately leaves alone.
Read the article→PANDORAVOICE · VOICE AGENTSWhat an AI voice agent has to get right
Sounding human is the easy part and the part everyone demos. The three things that decide whether a phone agent is worth having are duller, and none of them is about the voice.
Read the article→SALESBOT · E-COMMERCEWhat AI product search has to get right
Semantic retrieval, image and voice input, and the one failure mode that destroys trust in a shopping assistant: confidently describing a product you do not sell.
Read the article→PANDORA CRM · COST ANALYSISWhat a six-tool SMB software stack actually costs
List prices for Mailchimp, Calendly, Intercom, Typeform, Hootsuite and Zapier, checked in August 2026 for a five-person business — and the finding that 72% of the bill scales with headcount rather than usage.
Read the article→LEADSBOT · COMPLIANCEGreek B2B data: what GEMI gives you, and what you may lawfully do with it
Greece's commercial registry publishes company email addresses in an open government API — and Greek law makes it unlawful to send marketing to them. The channels that remain legal, and the case that settled it.
Read the article→PLANARO · EVIDENCE REVIEWAppointment no-shows: what the evidence actually supports
The widely-quoted 30% salon no-show rate traces to reports that do not exist. Deposits have one randomised trial and it found nothing. What the research does support is narrower — and more useful.
Read the article→Written by the PandoraBot team. We publish list prices, name our sources, and label customer-reported results as customer-reported. Where we do not have evidence for a claim, we say so.