Observe · Conversation intelligence
Conversation intelligence for your AI phone and chat agents
Observe is the conversation intelligence and call analytics layer for your AI agents. It reads every conversation and shows you what they actually handle — containment, top topics, caller sentiment, and call scoring — with a forensic receipt for every call. So you’re never guessing whether it’s working, and you always know what to fix next.
Free credits. No credit card. Zero balance pauses — never bills.
Scripted demo — a replay of the real product, not a live session.
The one call analytics metric that matters
Containment: how much it handled without a human
Most call analytics drowns you in charts. Observe opens with the answer instead. The first number you see is containment — the share of conversations resolved without a person — beside your conversation trend and warm transfers. It is the fastest way to tell whether your AI receptionist is earning its keep — no vanity metrics, no dashboard maze to decode.
Containment
The share of conversations resolved end to end by the agent — the one call analytics number that says whether it is earning its keep.
Conversations
Volume with a trendline, plus outcome and sentiment breakdowns — so a busy week and a rough week look different at a glance.
Warm transfers
How often the agent handed off to a human — each one with full context passed along, and each one traceable to its Receipt.
Watch the walkthrough
Watch: call analytics on real demo data — 7 minutes
A full tour on live HVAC demo data — every number on screen is real.
The metrics
Call analytics across 24 metrics, grouped the way you read them
Conversation intelligence is only useful if you can take it in. Observe scores every conversation across 24 metrics today, sorted into the categories they belong to. Every metric is one registry row — so new ones appear in the picker automatically, no dashboard rebuild required.
AI & handover
How much your agent resolves on its own — and why it hands off when it does.
- AI containment
- The share of conversations resolved end to end without a human — the headline number.
- Escalation reasons
- Why the agent transferred, ranked — so you can shrink the top reasons and lift containment.
AI agent
The core view of what the agent handled and what callers came for.
- Conversations
- Total conversations handled in the window, with a trend against the previous period.
- Transfers
- How many conversations the agent warm-transferred to a person, with full context passed along.
- Outcomes
- Every conversation sorted into completed, transferred, abandoned, error, or other — the reconciliation view.
- Top topics
- What callers contact you about, ranked by volume, each drillable to the exact conversations.
- Goal completion
- How often the agent reached the goal it was there for — a booked visit, a captured lead, an answer.
- Knowledge gaps
- Clustered questions the agent could not answer, ranked so the most valuable fix sits on top.
- Voice latency p50 by language
- The typical response time before the agent speaks, broken out by the caller’s language.
- Voice latency p95 by language
- The slow-turn response time by language — the tail experience, not the average.
- Voice latency by stage
- Where the response time is spent across the pipeline, so you can see what to tune.
Quality
How the conversation felt, and how well the agent carried it.
- Caller sentiment
- Average caller sentiment, read from the transcript instead of a survey.
- Sentiment mix
- The spread of positive, neutral, negative, and frustrated conversations across the window.
- AI-assessed satisfaction
- A transcript-derived CSAT across every conversation, not just the few who answer a survey.
- AI talk ratio
- How much of the talking the agent did versus the caller — a check on monologuing or terseness.
- Dead air per call
- Average silence before the agent responds — the clearest “does it feel laggy” signal.
Resolution
How quickly, and how completely, issues actually get solved.
- Resolution time (median)
- The typical time a resolved conversation takes from start to finish.
- Resolution time (p90)
- The slow-tail resolution time — how long the hard cases run.
- First-contact resolution
- The share of callers whose issue was solved on the first contact, with no follow-up.
- Repeat-caller rate
- The share of callers who came back within the month — a loyalty or a failure signal, by line.
Cost & ROI
What each conversation costs, and what the agent saves.
- Dollars saved by AI
- The estimated value of conversations the agent handled that would otherwise need a person.
- Cost per conversation
- What an average conversation cost to handle, as a plain unit you can read.
- AI spend
- Total agent spend over the window, so cost lives next to the outcomes it bought.
Volume & answer
When the calls arrive — and the always-on payoff, quantified.
- After-hours share
- The share of conversations that arrived outside business hours — coverage a 9-to-5 line can’t give.
Dollars saved · the math
The logic behind the “Dollars saved” metric
Observe’s dollars-saved metric asks a simple question: what did the agent handle that would otherwise have needed a person on the clock? For an always-on AI answering service, that adds up fast. Here is the labor math it’s reasoning from. The wage below is an illustrative assumption — your numbers will vary — but the shape of it holds.
A person on the clock vs. the agent
At roughly 200 calls a month, that’s about $66 of talk time for the agent — against a receptionist you’d pay for the hours, not the calls. Observe tallies the version of this for your actual volume and reports it as dollars saved, call by call.
Illustrative math based on a $20/hour wage for business-hours coverage; round-the-clock figures assume a rotating team at the same wage. Labor costs are assumptions for comparison, not a guaranteed saving. Flowyte voice is billed at $0.11/minute. Your numbers will vary with wages, hours, and call volume.
Top topics
What your callers actually call about
Every call analytics tool counts calls. The useful question is what they were about. Observe groups every conversation into topics and ranks them, so the shape of your phone line stops being a guess.
Say you run an HVAC business. Observe can tell you how many callers this week reported the AC is out — then you drill from that topic straight into the exact conversations behind the number, and read the transcripts. The insight and the evidence are one click apart, which is what turns a topic chart into a decision.
Top topics · this week
Illustrative example — values from the scripted demo, not live data.
Sentiment mix
Illustrative example — values from the scripted demo, not live data.
Call scoring & caller sentiment
Every call scored, so you open the rough ones first
Nobody has time to listen to every recording — the old way to do call monitoring. So Observe scores every conversation for you on consistent signals — its outcome (resolved, transferred, or abandoned), whether the agent completed its goal, and how the caller felt. That’s call scoring without a QA team.
Caller sentiment then becomes triage. Observe sorts conversations so frustration surfaces instead of hiding inside an average, and rolls the whole week up into a sentiment mix. You open the calls that went sideways first — and the Receipt is right there to tell you why.
The improvement loop
Blind spots come to you, ranked
Most platforms show observability as QA plumbing you have to dig through. Observe keeps a running list of the questions your agent couldn’t answer, ordered by how often they come up — which means the most valuable fix is always at the top, and fixing it takes a sentence, not a ticket.
- 1See the gap — the question, how often it came up, and why it missed.
- 2Tag the call into Assist — it attaches as a chip, and the copilot reads that exact conversation.
- 3Fix it in plain English — "add our heat pump services to the knowledge base."
- 4Deploy — the change lands on your draft first, then goes live when you say so.
Questions the agent couldn’t answer
Illustrative example — values from the scripted demo, not live data.
Receipt
Call · voice · 3m 12s- Call startedvoice
- Caller"Do you have any openings tomorrow morning?"
- Knowledge lookupservice area & hoursgrounded 0.86
- Tool — Book appointmentGoogle Calendarsuccess · 412ms
- Guardrail — pricing disclosureallowed
- Agent"You’re booked for the 8-10 AM window tomorrow."
- Call endedcompleted
Illustrative example — values from the scripted demo, not live data.
The Receipt
A forensic receipt for every conversation
Open any conversation and the Receipt shows exactly what happened, turn by turn: the full AI call transcription alongside every knowledge lookup with a grounded score, every tool call with its status and duration, every guardrail decision, every verification event, every handoff — which means “why did it say that?” is a question with an answer, not a mystery. It’s call-by-call transparency most call analytics never gives you.
It’s the same transparency you get before publishing - the guardrails you set show up here as decisions you can audit, call by call.
Custom dashboards
Build custom call reporting dashboards
Observe ships with a dashboard that answers the first questions, then gets out of your way. Build and edit your own: add and arrange the widgets you care about from the metric catalog, apply filter lenses by agent, channel, or time, and start from a template instead of a blank page. Every widget still drills down to the exact conversations behind the number — then schedule any dashboard as a PDF report, emailed on the cadence you choose.
New to this? Start with the call center metrics that matter and build a dashboard around them.
Build & edit
Add, arrange, and resize widgets on a dashboard that is yours to shape.
Filter lenses
Slice by agent, channel, and time — the whole dashboard follows the filter.
Templates
Start from a prebuilt dashboard instead of a blank canvas.
Assist by chat
Describe the dashboard you want and Assist assembles it from the metric catalog.
Scheduled PDF reports
Email any dashboard to your team daily, weekly, or monthly — the real charts, as they render on screen.
Drill-down everywhere
Every widget clicks through to the exact conversations behind it.
Scripted demo — a replay of the real product, not a live session.
Describe the report you want
The dashboard you want, without building it by hand
Assist operates Observe by chat: ask for “a dashboard for after-hours AC emergencies and how many turned into bookings” and it picks the metrics, adds the widgets, and filters them for you — which means the person who needs the numbers can get the dashboard, even if they’d never build one.
Assist drives the same dashboard + metric APIs you do — nothing bespoke. So anything it builds, you can open, edit, filter, and schedule like any other dashboard, and every widget still drills down to the exact conversations behind the number.
Scripted demo — a replay of the real product, not a live session.
See it in the Studio
Watch the gap list turn into a fix
The Observe act below replays the real flow: the metrics strip, the ranked gap list, a call tagged into Assist, and the Receipt scrolling through one conversation.
ACT 5 — OBSERVE: See what it handled, what it missed, and the receipt for every call.
Scripted demo — a replay of the real product, not a live session.
FAQ
Conversation intelligence, answered plainly
What is conversation intelligence?
What is call scoring?
How do I know if my AI receptionist is working?
What is containment for an AI agent?
Can I get scheduled call reports by email?
How do I find out what the agent couldn’t answer?
What is a Receipt?
How do I fix a bad call?
Can I build my own dashboards?
Can the AI build reports for me?
Build an agent that shows its work
Draft it in minutes, put it on a line, and read its first receipts tonight.
Free credits. No credit card. Zero balance pauses — never bills.