Insights & Reporting

Insights reports on conversations your account has actually handled. It has three pages (Usage Reports, Call Logs, and Voice Trends) plus a read-only replay of any single workflow run on the flow canvas.

Reporting is derived from what your conversations emit as they run, so one appears once it has finished. It is a record of what happened, not a live view of what is happening now — for that, use Supervisor.

To compare your AI agents on containment, resolution, and transfers, see Agent Performance.

Channels

Insights covers more than phone calls. Every conversation records the channel it happened over, and each page has a Channel filter offering the media your account actually has data for in the selected range.

You only see a Channel filter when your account has more than one medium in range — with only phone calls, the control would have nothing to narrow.

What a duration means depends on the channel

Avg Duration, the Duration column, and the Minutes total are wall-clock length from start to end. That is a call's talk time, but on a web chat it includes the time a visitor spent reading before replying, and on a text message it means very little at all. Compare durations within a channel, not across them; the Channel filter is what makes that possible.

A conversation with no channel

A conversation whose medium was never recorded shows — rather than being assumed to be a call. Choosing any specific channel excludes it, since "not known" is not a medium.

Dates and times are UTC

Every date bucket, range preset, and timestamp in Insights is UTC, and the controls are labeled to say so. "Today" therefore begins at 00:00 UTC, which may not be your local midnight.

This matters when comparing against a report from somewhere else: if the numbers are close but not equal, check whether the other report uses local time before assuming one is wrong.

Usage Reports

Seven totals for the range you select, a volume chart, and a per-day table.

Calls and runs count different things

The two are counted differently.

• The Total Calls card counts conversations — one per conversation, however many workflows it passed through, and on every channel.
• The table's Runs column counts workflow runs. A conversation that hands off from one workflow to another contributes a run to each.

So on an account where workflows chain, summing the Runs column gives a larger number than the Total Calls card. Both are correct; they answer different questions ("how many people got in touch?" versus "how much work did this workflow do?").

There is a second reason the two can differ, and it is the more common one on a multi-channel account: the table only has rows for conversations that ran a workflow. A web chat handled entirely by an AI agent runs none, so it counts in the cards and in the chart but has no row in the table below. There is no workflow to attribute it to. Filtering by workflow narrows the whole page to workflow runs for the same reason.

The other cards

Averages and durations skip conversations that have no recorded length — a call rejected at a concurrency limit never connected, so counting it as zero seconds would pull the average down. For the same reason a missing duration shows as — rather than 0s.

Transfers are a voice and chat idea. A text message or an email is never transferred, so on an account with a lot of messaging traffic the Transfer Rate is diluted by conversations that could never have been transferred. Filter to a channel to read it fairly.

Failed under-reports slightly

A step that errored and never reported finishing leaves no error record, so its call is counted Abandoned rather than Failed. Treat Failed as a floor, not an exact count.

Large accounts see a top-N table

The detail table lists the busiest workflows, up to a limit. If your account has more workflows than that in the range, a notice above the table says how many are not listed. The chart and the seven cards always cover every workflow, so the totals stay complete even when the table is trimmed.

Call Logs

Every conversation in range, newest first, with search and filters for channel, status, type, workflow, agent, and date range. The list is paged; the page you are on is part of the address, so a link you share opens on the same page it did for you.

One row per conversation

A conversation is one row, however many workflows it went through. A caller who reaches an AI agent, has it invoke two workflows, and is then handed off is a single row with a single Conversation ID — the same id every part of the call reported under.

The columns are Channel, Conversation ID, Source, Destination, Handled by, Start Time (UTC), Duration, and Status. Channel is shown as an unlabeled icon; hover over it to see the name.

An inbound voice call's Destination is labeled Toll-free or Local by the kind of number it came in on. That's how its call minutes are billed on the Billing page's Usage tab, and hovering over the label says so. The label shows even when the workflow's Redact Data in Logging setting hides the numbers. A browser test call reads Browser test call under Source and has no Destination. A conversation's page shows the same under From and To. Calls made before the label was added have none.

Status is the conversation's outcome, not one leg's. A conversation whose workflow failed reads Failed even if the call itself completed, because a step that errored is the more useful fact.

Sub-tenant calls

If your organization has sub-tenants, Call Logs lists their conversations beside your own, with an Organization column and an Organization filter to narrow the list. This needs administrator access in your organization. Every role Tandem issues today (Owner, Admin and Member) carries it. Usage Reports and Voice Trends cover your own organization only.

Opening a sub-tenant's conversation shows its details log, step timeline and recording. Its call summary, transcript, wrap-up and collected data are not shown from the parent organization, and its workflow steps have no View in Flow. A sub-tenant row names the caller's contact without linking to it. Switch to the sub-tenant to see any of these.

Who handled a conversation

Handled by shows the first workflow or AI agent that took the conversation, marked with the Workflows or Agents icon, and +N when more followed. Hover over it to see every one, in the order they took the call.

The Type filter sorts conversations by what handled them:

• Workflow only: one or more workflows ran, and no AI agent took part.
• Agent only: an AI agent handled it, and no workflow ran.
• Workflow + agent: both, in either order. A workflow that hands the call to an agent and an agent that hands it to a workflow both count.

A conversation nothing answered, such as one the caller abandoned before a workflow or agent picked up, matches none of the three.

The Agent filter lists the agents with conversations in the selected range. It matches a conversation when the agent answered it or was switched to during it. Picking a workflow and an agent shows only conversations both of them handled.

Switches between agents are read from the call details log, so on a conversation older than that log only the agent that answered counts. See Conversations older than the action list.

Filters from Agent Performance

A drill-down sheet on Agent Performance ends with Open these conversations in Call Logs. The link opens Call Logs on that agent, the sheet's date range and channels, and the sheet's own conditions. Each condition shows as a chip above the list, for example Not contained, Ended: Caller hung up, Outcome: Unresolved, Sentiment: Negative, Taken over, Version: v2.1, Transfer to: Billing · Failed, Tool: lookup · failed, Repeat contact, or Day: Oct 3, 2026 (UTC). These conditions have no control in the Filters menu, so the chips are where you see and change them.

To drop one, click the x on its chip. The list refreshes without that condition and keeps the rest. A chip that carries two conditions, such as a transfer destination and its result, drops both. When the last chip is gone, the list is the agent's calls in the range.

Each conversation is listed once, even if the agent handled it more than once. Agent Performance counts every time the agent answered, so its number can be higher than the number of rows here.

These conditions describe one agent's calls, so they apply only while exactly one agent is chosen in the Agent filter. With none or several, the chips stay and a note reads "These filters describe one agent's calls, so they are not applied until exactly one agent is chosen."

If the list is not filtered by the chips, a warning reads "These filters aren't supported yet, so this list isn't filtered by them." Treat the list as unfiltered. It appears when the server did not confirm that it applied the conditions.

The call details log

Expanding a row lists everything that happened on the conversation, in order, with a time and the actor for each: the agent starting, a workflow being invoked, a hand-off out and back, a switch from one agent to another, a transfer, the conversation ending.

A workflow entry carries that run's step timeline underneath it — every step in order, how long each took, and which exit it left by — which is what the earlier list showed for the row as a whole.

View in Flow sits on the workflow entry rather than on the row. With more than one workflow on a call there is no single flow to open, so the button belongs to the run you picked. It replays that run on the canvas step by step, highlighting the path it took, and always loads the workflow version that ran — a call from before an edit still replays against the flow as it was.

If a hop shows a name you do not recognise, it is a kind of hop this page is older than. The log shows it under its raw name rather than hiding it.

Four things the log deliberately does not contain:

• Variable values. Step inputs and outputs are not retained in reporting.
• Tool arguments and tool results. The log records that a tool ran, not what it was passed.
• What was said. Use the row's transcript button, described below.
• A status, on rows where nothing recorded one. Those show — rather than a guess.

Recorded calls

A call that was recorded, or that Tandem tried to record, carries a small icon beside its Status. Hover over it to see which of three states the recording is in:

• Recorded — the recording is stored.
• Recording pending — the recording has ended and is being stored.
• Recording failed — no recording of this call was stored.

A pending recording moves to one of the other two once it is stored or fails. A call with no icon was not recorded, or is still in progress. Calls recorded before this indicator was added show no icon either.

Expanding the row repeats the state as a badge next to Call Details:

• Recording: 4m 12s — the length of the stored audio. It can be shorter than the call's Duration, because a stretch where recording was paused is missing from the audio.
• Recording: Ready — stored, with no length available.
• Recording: Processing — still being stored.
• Recording: Failed — followed by the reason when Tandem has one, such as No audio captured or Deleted to protect card details. The second means Tandem deleted the recording because it could not confirm that recording had paused during card entry.

There is no way to play a recording from Call Logs yet. See Call Recording for which calls are recorded.

What a conversation cost

A conversation's page shows a Cost section under its summary: each charge on the conversation with its quantity and cost, and the total, at your contract's rates. It is the same figure the Billing page's By Conversation view lists; see Usage. Call, IVR, and AI Voice time is rounded up to the next half minute, and the section says so; see How a call is billed. On a browser test call, it also says why the call shows a Local call, in charge. It is not shown inside a sub-tenant, or on a sub-tenant's conversation opened from its parent, because a sub-tenant's usage is billed to its parent.

Answer detection on outbound calls

An outbound call placed by a Convert to Voice node with Detect Human or Answering Machine on shows what detection decided. The verdict is a badge in the Status cell. Hover over it to see the reason. Open the call and the Details tab lists it as Answer detection, with the reason in parentheses.

• Human — a person answered.
• Machine — an answering machine or voicemail answered.
• No answer — the call was not answered.
• Detection failed — the call was answered, but no human or machine verdict could be produced. The workflow takes the node's detectionFailed exit.

The reason is the rule that produced the verdict, such as Brief greeting, Voicemail phrase heard, Continuous speech, or No clear speech before the time limit. A verdict that carries no reason, such as No answer, shows the verdict alone. A label Tandem does not recognise appears under its raw name.

A call with no badge did not use detection, or was placed before this was recorded. An outbound call whose workflow was started through the API has no channel until detection reports or a voice agent takes the call. It is then listed with the Voice channel and the Outbound direction.

Conversations older than the action list

The action list is recorded from the change forward and nothing was filled in behind it. An older conversation shows its workflow runs and their step timelines, with no list of hops above them. It also keeps whatever ids it reported under at the time, so a call that produced two rows then still produces two rows now.

Reading a web chat

The transcript button at the end of a row (the speech-to-text icon, View Transcript on hover) opens the conversation and shows everything that was said, in order. It appears on Voice and Web Chat rows only — those are the channels a transcript is kept for. If a web chat's full transcript has aged out, the panel shows the agent's own retained history instead and says so above it: that history is trimmed from the front, so a long conversation may not start where the visitor did.

The download button beside the panel's close saves the whole conversation as a text file — the header, then one line per message. Times in the file are UTC, not your local time, so the file still reads correctly for whoever you send it to.

A call a supervisor took over

When a supervisor took a call over from its agent, the conversation's details say who did and when (Taken over (UTC)), how they closed it out (Supervisor disposition, with their note), and why, when they said (Why they stepped in). The supervisor's disposition sits beside the agent's own Disposition rather than replacing it, and also appears on a call a supervisor joined and closed out without taking over.

The transcript carries on past the agent's last turn. A line reads [name] took over the call. The agent left., and what the caller and the supervisor said after it follows. A downloaded transcript includes both parts.

One row per workflow run

The earlier list, still in place for organizations the change has not reached.

A conversation that ran a workflow gets one row per run, so one that chained two workflows appears twice, and the rows carry Workflow and Workflow ID columns instead of Handled by. A conversation that ran no workflow, such as a web chat handled entirely by an AI agent, gets a single row with — in those columns and an empty execution path. Nothing is being withheld there: without a workflow run there are no steps to list.

Expanding a row shows that run's execution path and its View in Flow button. There is no action list and no transcript panel.

Voice Trends

A flow diagram of how conversations actually moved through your steps, with a Top Paths table beneath it.

Despite the name, this page covers every channel that runs a workflow — an SMS-driven flow is drawn here alongside a phone menu. What it cannot show is a conversation that ran no workflow, because there are no steps to draw. Use the Channel filter to look at one medium's flow shape on its own.

Global and per-workflow views

• Global groups by step type — "Play Message", "Condition", "Transfer to Queue". Useful as a shape-of-traffic overview across everything.
• Per Workflow groups by step name, giving the real picture of one flow. Select one or more workflows to switch to it.

Global is intentionally coarser. Grouping by name across unrelated workflows would merge steps that merely share a name, so the global view trades detail for a view that spans everything.

Repeated steps are listed, not drawn

A flow diagram cannot draw a loop. When callers go back to an earlier step — a menu re-prompt, a retry after invalid input — that movement is removed from the diagram and reported beneath it as Repeated steps, with a count per step.

So a step showing "×146" means callers returned to it 146 times in the range. Those calls are not missing from the totals; they are just not drawable as an arrow.

Top Paths

The most common whole-call routes, as a share of all runs in the range. Two caveats:

• These are the top few paths, not every path, so the percentages do not add up to 100%.
• A call running longer than an hour may have its earliest steps missing from its recorded path, so very long calls are under-represented here.

When a page looks empty

An empty Insights page is usually accurate rather than broken:

• Reporting history starts when reporting was switched on for your account. There is no backfill, so calls from before then do not appear.
• The range may genuinely have no conversations — try a wider one.
• A filter may exclude everything. The workflow and channel filters only offer values that have data in the selected range, so widening the range can reveal more options.
• A channel filter hides conversations with no recorded medium, and a workflow filter hides conversations that ran no workflow. Clearing both is the quickest way to tell an empty range from an over-narrow one.

Ranges are limited to 90 days. A longer range is refused rather than silently truncated, so a number you read is always the whole number for the window shown.