Perplexity
Each of these is a real decision Perplexity is working through, written so you can make the call yourself — incomplete information, someone who pushes back, no answer key.
Perplexity develops an AI answer engine that combines real-time web search with generative AI to deliver source-cited responses, plus research, file analysis, image generation, and computer-use capabilities. In 2026 it launched Computer, a product for agentic AI. It serves consumers, professionals, enterprises, and developers through freemium and subscription plans, per-seat enterprise licensing, and usage-based APIs. Its highest-priority verticals are Finance, Healthcare, Legal, and Consulting. Users answer more than 780 million queries every month, and Perplexity maintains a 200B+ URL index. It controls its full stack: the models, the agent harnesses, and the search infrastructure underneath. It builds and runs the inference engine behind every Perplexity query, deploying dozens of model architectures at scale with tight latency and cost budgets. Its products run on deployment models ranging from multi-tenant public cloud to single-tenant and on-premises solutions for enterprise customers. It builds controls including SSO, SCIM, role-based access, customer-managed encryption keys (BYOK), admin APIs, and enforceable policies governing access to data, models, and tools.
Your first one is free. No card.
32 live right now.
The rest
Trace index failures before fresh sources fall behind
Failed indexing runs are delaying fresh pages from reaching source-cited answers. You must decide what the system can rank automatically and what engineers must verify before changing production.
Run the runbook to the approval gate while engineers verify each step
Search answers can stall after a failed deployment while engineers reconstruct which runbook applies. You must weigh faster incident triage against evidence an engineer can trust before production changes.
Open a research path without taxing everyday search
Pro researchers are hitting limits during complex investigations, while shared inference still serves everyday searches. You must weigh a differentiated research path against latency and capacity for everyone else.
Choose the node cutover while answers fail during drains
Some answer requests stop halfway through whenever inference nodes drain during peak research traffic. You have to weigh removing the exposed node image against preserving capacity for live searches.
Restore hits in Redis without rolling back the backend release
Cache misses are pushing answer assembly past 900 milliseconds and loading PostgreSQL repeatedly. You must weigh a fast configuration rollback against a targeted release-path change that protects future backend scale.
Guide the next research step after a long answer
Repeat users with long questions open several citations, then often leave without a follow-up query. You must weigh helpful action ranking against relevance, trust, and inference cost.
Choose a page policy for original reporting before the 48-hour rebuild
Original-source selection fell 21 points after a parsing-model update, and weak claim support now reaches 18% in the affected set. You must weigh recall against a durable policy change that cannot be cheaply reversed.
Fix stalls in KV-cache allocation so p99 drops to 2.4 seconds
p99 time to first token rose from 2.4 to 3.3 seconds after a model rollout. You must weigh a fast scheduler change against preserving completion and answer behavior.
Trace connector failures without trusting an AI story blindly
Connector failures are taking about 90 minutes to diagnose because engineers reconstruct agent traces by hand. You must weigh faster investigation against evidence review and organization-boundary protection.
Bound research spend without losing cited answers
Deep research calls can produce useful answers before developers discover the cost was outside their budget. You must weigh a firm spend boundary against the retrieval depth that makes the answer useful.
Seal connector boundaries before enterprise access is suspended
Two developers received project names from another organization after credential refreshes. You must weigh a hard credential cutover against disruption to active agent workflows.
Prevent duplicate writes when connector timeouts trigger retries
Connector timeouts are producing 18 duplicate downstream writes each day. You must weigh reliable retries against the risk of repeating an action that already landed.
Speed connector reviews without weakening agent-loop evidence
Connector onboarding takes 10 business days, and repeated review defects delay enterprise workflows. You must weigh faster draft assembly against the engineer judgment needed to protect permissioned tool behavior.
Show workflow spend before approvals stall recurring research
Scheduled research workflows stop after spending credits, leaving coordinators with incomplete briefings. You must weigh useful visibility against protected connector details and uncertain agent plans.
Decide on stale access before credential migration closes
Removed access sometimes still reaches workspace records through resumed sessions. You must weigh rollout timing against proof that every connector fails closed.
Stop retries from draining credits before scheduled work finishes
Paginated connector retries are charging some scheduled workflows twice and stopping them early. You must weigh recovery from partial failures against exact credit settlement.
Ship a brief for one researcher in Projects instead of waiting
Repeat users are reopening old threads and copying citations into personal notes to recover context. You have to weigh a focused personal brief against a broader collaboration vision.
Settle the limits of health answers before expansion in 18 days
Human review finds unsupported or overconfident health claims in a meaningful share of sampled answers. You have to weigh stronger safeguards against answer completion and the scheduled expansion.
Restore follow-up questions after the mobile answer layout changes
First-session follow-up rate fell from 31% to 24% after the layout release. You have to weigh a faster interaction fix against a broader citation-layout change.
Ship a flagged view of source conflicts without slowing answers
Repeat researchers are opening multiple citations manually after longer questions, signaling a comparison gap. You must weigh a fast answer flow against a clearer path through conflicting and incomplete source evidence.
Commit the segment rebuild while live reads still carry the load
A damaged index segment is producing citations that do not support some generated claims. You must weigh source-grounding recovery against the risk of degrading live answer availability during an irreversible cutover.
Cut duplicate writes before fresh pages miss the morning index
Duplicate documents are up 19%, and affected domains are arriving 14 minutes late. You must weigh crawler-side prevention against a Search-side filter when only one maintenance window remains.
Surface conflicting evidence before research becomes a confident answer
Review samples show conflicting internal sources becoming confident research outputs. You must weigh fewer wrong answers against the interruption cost of asking people to clarify context.
Stop drafts from citing files that left the matter before rollout
A halted filing draft cited a document removed 19 minutes earlier. You must weigh uninterrupted legal research against a control that must be trusted before sensitive drafts proceed.
Fix retries for Slack posting so finished work is not sent twice
Duplicate Slack posts now affect 3.8% of completed posting tasks. You must weigh fast recovery from real failures against the credibility lost when completed work repeats.
Ship a provenance label for citations without overclaiming verification
Readers lack a clear signal for how a displayed citation was retrieved and verified. You must weigh customer-visible trust context against latency, coverage, and the risk of overstating uncertain evidence.
Contain agent egress without shutting down useful source retrieval
Verified logs show suspected metadata requests through browser-linked agent workflows. You must weigh irreversible containment against keeping legitimate source retrieval available for consumer research.
Stabilize retrieval before retries exhaust the evening capacity
p99 retrieval latency has reached 910 ms and partial answers are at 2.1%. You must weigh a targeted connection-reuse change against the safer rollback path in one deployment window.
Ship a review timeline so users stop replaying tasks before accepting
Users replay completed tasks because the current result does not show enough of the agent’s path. You have to decide what evidence earns trust without turning review into another full task.
Ship context partitioning while keeping 85% of signed-in workflows
Adversarial tests reproduce cross-origin access while authenticated workflows depend on existing session continuity. You have to choose an irreversible isolation path without losing the workflows enterprise users rely on.
Fix the sheet hidden by a permission prompt before the next release
File-upload tasks now complete at 74%, down from 91% after a Chromium update. You have to restore completion without weakening the browser permission boundary.