SightLift.ai · Team Activity Dashboard

Topics

Each conversation is assigned exactly one topic via rules-based pattern matching on title + first message. Non-work is filtered out of all summary views by default.

Sub-topics

Within each topic, a finer split into recurring patterns (e.g., marketing → linkedin-social, brand-voice, email-outreach). Click any sub-topic to see its conversations.

Output types

What the conversation actually produces (the artifact).

Complexity

Weights agentic patterns (tool calls, tool variety, extended thinking, attachments). Captures how powerful was the workflow. Independent from effort below.

Simple

score < 5 · quick Q&A, single ask + response

Medium

score 5–19 · typical assisted work, a few iterations

Complex

score 20–49 · multi-step builds, several tool calls

Deep

score ≥ 50 · long, tool-heavy, often with extended thinking

Effort

Weights human input (turns + chars + attachments). Captures how much the user put into it. Two conversations can have very different complexity but similar effort — or vice versa.

Quick (one-shot)

score < 6 · 1–2 turn ask, minimal context

Light (refinement)

score 6–14 · 3–5 turn polish, light back-and-forth

Substantial (workshop)

score 15–34 · 6–15 turn iteration with real revisions

Deep (collaboration)

score ≥ 35 · 16+ turns, heavy back-and-forth, often with attachments

Tokens

Used to estimate cost. Counted per conversation by content type — input tokens (human messages + tool results) vs. output tokens (assistant text + tool calls + thinking).

Input tokens

What the model reads on each turn: system prompt + prior context + new user message + tool results. These accumulate as conversations get longer.

Output tokens

What the model generates: assistant text + tool_use blocks + extended thinking. Always cheaper to produce more but priced higher per token.

Why character-calibrated

This sandbox can't reach the tiktoken BPE blob to run real tokenization. Calibrated ratios per content type get within ~10% of true counts on representative samples — close enough for cost planning.

Cost model

Published per-1M-token rates × the team's provider/model mix × tokens consumed. Input and output are priced separately because they differ by 3–5× per model.

Per-1M-token pricing (May 2026)

Provider mix (tied to data source)

Anthropic Claude

100% — the conversations.json export came from claude.ai, so we know every conversation was Claude. No assumption here.

OpenAI

0% — not represented in this export. If the team uses OpenAI for other work, that usage would need a separate export from platform.openai.com.

Google Gemini

0% — same caveat as OpenAI.

Within-Anthropic model mix (assumed)

Claude Sonnet 4.6

78% — recommended default for most work. The export doesn't tell us which model handled each conversation, so this is a heuristic.

Claude Haiku 4.5

12% — assumed for quick lookups and high-volume work.

Claude Opus 4.6

10% — assumed for the hardest reasoning tasks (long workshops, complex strategy).

Token budget review

A focused, per-person read for one decision: should this employee get more tokens? Opened from the Cost tab's by-user table when someone hits a cap. SightLift measures the spend and renders an advisory assessment — it never approves; that happens in your budget tool, which is why you enter the “since” date yourself (SightLift can't see the approval).

The read

Efficiency-based, not an ROI verdict: it says whether the spend looks like real work done efficiently — not what it produced (measuring outcomes needs a session→artifact link that doesn't exist yet). Thresholds are calibrated to your tenant's own distribution, not fixed numbers.

Supports request

Aligned, efficient, converging work.

Supports, with notes

Efficient, but some orphan spend or right-sizing worth noting.

Needs your call

Signals conflict — telemetry can't classify it; you decide.

Worth a look

Spend looks like spinning, not progress.

Alignment — is it real work?

The share of a person's tokens spent on topics that match a company.md objective. A conversation on a specifically-named priority topic counts full; one covered only by an org-wide (scope: all) objective counts partial (0.4) — otherwise a single org-wide goal would make every conversation read 100% aligned. Non-work conversations are never aligned, even under an org-wide objective. The rest is orphan spend. Topics are classified from content, so this can't be gamed by self-labeling. No company.md → alignment is omitted. This is a view of where the money went, not a ranking input — it never moves a play up your action plan. The Cost tab's Aligned column and “Share going to priority work” KPI apply this same per-conversation crediting; they differ only in denominator (the Cost tab shows the share of work spend, the review the share of total window spend — personal use counts as orphan here, but doesn't dilute the Cost tab's work-spend share).

Efficiency

Three proxies for “done efficiently,” all relative to your tenant, not fixed cut-offs. Where there's enough history, each also shows the person's prior equal-length window (“was …”) — improvement is judged against themselves, not only the cohort. A thin prior window (fewer than 5 sessions) is never shown as a trend.

Cost vs peers on similar work

The person's tokens ÷ the median cost of work at that complexity tier across your tenant. Around 1× is in-band; lower is better.

Repeated or duplicate work

Share of tokens spent re-doing near-duplicate work (the same prompt cluster hit again). Lower is better.

Overspent on simple tasks

Low-tier conversations costing more than the 90th percentile for that tier — expensive effort on cheap tasks. (Model-based right-sizing — a deep model on a rename — needs model data not yet on the telemetry path.) Lower is better.

Agentic share

The share of a person's output tokens spent on the agentic machinery — tool calls, their arguments, and extended thinking — rather than plain answer text (agentic_out ÷ total_out). High means tool-heavy, autonomous work; low means mostly writing or answering. It's the raw input to the gate's spin signal, which fires only when a person's agentic share sits above your tenant's 75th percentile. (The Cost tab's “agentic share of spend” is the same idea measured over total tokens and cost-weighted.)

Why the spend is high

Everyone reviewed is already a heavy spender, so raw size separates no one. The read looks at the shape of the spend on two axes — signs of progress vs signs of spinning — and routes into three outcomes.

Signs of progress

Depth (genuinely complex work), novelty (new ground, not re-tread), convergence (closing in over time).

Signs of spinning

Repetition (same cluster), spin (tool-loop overhead above your tenant's p75), static surface (no new ground), tier mismatch (low-tier over-spend).

Genuinely hard problem

Strong progress signals, little spinning → supports the request.

Unclear — your call

Signals conflict → routed to you. The default when it's genuinely unclear.

Rework without progress

Three or more spinning signals converge with weak progress → worth a look. Deliberately hard to trigger: falsely flagging a good engineer on a hard problem is worse than funding some waste, so a thin history is never flagged.

What the gate can't see: whether quiet work succeeded or was abandoned — both make a topic go silent. Telling them apart needs a session→outcome link that doesn't exist yet; until then, “advancing toward a dead end” can read as a hard problem.

UI conventions

Drill modal

Click any bar, chip, or table row to open a filtered list of conversations with sortable columns. Compose multiple filters by drilling repeatedly (the filter description across the top shows what's applied).

Stacked-by-effort bars

On the Dashboard, the "By topic", "By complexity", and "By output type" panels show bars where length = volume and internal split = effort mix. Hover any segment for full numbers.

Heatmap on hover

Hovering any person row pops up their 30-day calendar (each cell is one day; intensity is conversations that day).

MTD

Month-to-date — total since the 1st of the current month. Compared against last month at the same point.

Drilldown card

Each person row expands inline to show topic mix, output mix, monthly trend, sub-topics, top tools, and 10 most complex conversations.

Connecting a data source

In Settings → Data sources, each provider connects one of three ways. Live API uses a read-only admin key to sync on a recurring schedule, bounded by the provider's API retention window (scheduled API sync is implemented for Anthropic today). Live telemetry (Anthropic only) streams OpenTelemetry events — the only way to capture Cowork; forward-only, reconstructed on a rolling hourly import. Manual upload takes a one-time export file (like the conversations.json that powered this dashboard) and imports its full history — useful for backfill or when outbound API keys are blocked. Sync schedules are tunable under Settings → Notifications & schedules. Full setup steps + a per-product capability matrix: the data sources guide ↗.

Agent opportunities

The Agent opportunities section at the bottom of the Action plan surfaces recurring work patterns that could be handed to a custom agent or skill. Candidates are found in two stages: a $0 heuristic clusters conversations by topic + sub-topic, then an LLM scores each cluster's fit for automation. Distinct from the Explore tab's Data Signals (management nudges) — this one answers "what could we build?"

Ranking

Candidates are ranked by frequency × cross-team adoption × complexity fit — how often the pattern recurs, how many people share it, and how well it suits an agent.

Stars / score

A 1–5 star strength rating (with an underlying score) from the LLM's agent-fit evaluation. More stars = a stronger automation candidate.

Tagline & problem

A one-line summary of the candidate, plus the underlying problem the recurring pattern represents.

Input / output shape

What the agent would take in (shape_in) and produce (shape_out) — the rough I/O contract, shown under "Show spec".

Why it's strong

The reasoning behind the rating — what makes this pattern a good fit for automation.

Risks

Caveats to weigh before building — accuracy needs, edge cases, or where a human should stay in the loop.

Savings estimate

A rough estimate of the time or cost an agent could save, where the data supports one.

Contributors & examples

The people who do this work today, plus example conversations behind the candidate — click "Show spec" to drill into them.

Actions

The Action plan is a synthesis layer over the Data Signals on the Explore tab: it bundles the many signals about the same person, topic, or group into a short, ranked list of concrete next steps — "coach this person", "de-risk this topic", "run an AI 101 for this group". Each action cites the signals and conversations behind it (click a card to drill in), names the score dimension it moves, and shows the points it would add — and the plan is ranked by those points alone, biggest first. Actions that name an individual are admin-only.

Every play carries one category (the colored chip), which sets the verb and the kind of follow-up doc you can generate:

Coaching

A 1:1 with one person — they've plateaued, gone quiet, or their results lag their effort. Generates a talking-points doc for the conversation.

Structure

Spread or de-risk concentrated work — a bus-factor risk, an orphaned topic, or a team that needs an agreed workflow or standard. Generates talking points.

Automation

Build or roll out a tool — recurring manual work, reinvented prompts, or an automation candidate. Pairs with the PRD / skill in the Agent opportunities section below.

Cost

Trim spend or improve cost-efficiency — a person or pair driving disproportionate cost, an over-provisioned model tier, or the same work duplicated across two providers. Generates a talking-points brief.

Enablement

Run an "AI 101" for a group who share a low-complexity, low-agentic pattern in a topic. The card names the members for admins; the headline stays count-based everywhere else. Generates a session outline.

Governance

Review sensitive exposure (HR / legal / finance / security volume) with the data owner.

Win

Recognize a positive pattern — a comeback, a first adopter, or a winning workflow worth spreading. Visible to everyone.

Priority

The pill (0–100) is rolled up from the underlying signals — a play backed by more, higher-severity signals ranks higher. The shown list is kept deliberately diverse, so one category can't crowd out the others.

Evidence

"N signals · M conversations" — click a card to drill into the exact conversations behind it, the same way signals do.

Generate a doc

Coaching/structure/cost/enablement plays have a 📄 button that drafts a ready-to-use Markdown doc — review it in place, then download or copy.

Who can see what

Plays that name an individual or a group are restricted to admins (and holders of the sensitive-data permission). Group member names never leave in an email, Slack, or a downloaded doc — only in the in-app drill.

Data Signals

The Explore tab surfaces these management signals — deterministic patterns detected across all conversations ($0, no LLM), ranked by severity. Each carries a lane (the colored chip below). A signal points at a topic, a person, or a pattern, and links to the exact conversations behind it (click any card to drill in). Signals that name an individual or touch sensitive content are tagged admin-only for future role-based visibility.

Trend usage shifts over time

Topic trend

A topic's usage jumped or dropped sharply month-over-month.

Person trend

A person's overall usage jumped or dropped sharply month-over-month.

New topic

A brand-new topic appeared this month with real volume.

Topic went quiet

A previously-active topic dropped off — automated, moved, or abandoned?

Consolidating

A topic is shifting into fewer, longer sessions — ripe to formalize into a workflow.

Output mix shift

A topic's mix of output types changed notably (e.g. a surge in one artifact).

Tool-use spike

Agentic/tool use in a topic jumped — a pattern worth formalizing into a tool.

Seasonal

A topic recurs on a predictable cycle — pre-stage or automate ahead of the next peak.

Person individual trajectories

Rising complexity

A person's work is steadily getting more complex — showcase it or have them mentor.

Stuck on simple

A tenured person has stayed on simple tasks — a nudge could unlock higher-leverage use.

Dormant

A once-active person has gone idle — re-engage before they churn.

High effort, few tools

Heavy work done with few tools — agentic features could speed them up.

Chat-only

Real volume with zero tool use, on a team where peers do use tools — the capability is proven next to them.

Prompt retyper

One person retypes a near-identical prompt over and over — save it as a skill.

Plateau

A once-rising person's complexity rose then went flat — a stretch project could restart growth.

Low artifact yield

Heavy effort but few concrete artifacts — check for a blocker, tool, or template gap.

Cost token spend & model fit

Topic cost concentration

A topic is a large share of token spend — highest-ROI target for a shared template.

Person cost concentration

A person is a large share of token spend — confirm it's intentional heavy use.

Token-efficiency outlier

A person uses far more tokens per conversation than peers on the same topic.

Budget review candidate

Heavy recent spend the thrash-vs-hard-problem gate reads as circling or can't classify — open their budget review (Cost tab → By user) before approving more tokens. Hard-problem and thin-history people are never flagged.

Artifact cost outlier

A team spends far more tokens per artifact (memo, deck, analysis…) than the org median for that output type — start with the two driving workflows. Teams doing measurably heavier work are never flagged. (Computed from token counts × output classification; team-aggregate — the per-person figure is admin-only.)

Model downgrade

Lots of simple-tier work on an expensive (Opus-class) model — a cheaper one would do. (Live-API tenants with per-conversation model data only.)

Model underpowered

Complex/deep work concentrated on the cheapest model available — a more capable one is likely worth it. (Covers only conversations with per-conversation model data; claude.ai uploads carry no model.)

Model-mix savings

Org-level run-rate: routing routine (Simple/Medium) work to the cheapest in-family model would save a meaningful monthly amount. (Tenants with a Usage or Compliance API connection only; estimates marked ~.)

Risk concentration & provisioning

Hidden expert

Deep work in a topic is held by only one or two people — a single point of failure.

Bus factor

One person owns a large share of a topic — de-risk the silo.

Hard-case outliers

A mostly-simple topic with a few hard outliers — the hard cases may need a dedicated tool.

Complexity variance

Complexity varies widely across people on one topic — share the efficient approach.

Over-provisioned thinking

Extended thinking used on simple tasks — right-size the model to cut cost.

Effort ↔ tier mismatch

High effort spent on simple-tier tasks — a template or faster model could cut the grind.

Abandoned long convs

Long conversations that produced no artifact — an unmet need or a stalling workflow.

No agreed workflow

No consistent output pattern across a team for a topic — align on a shared approach.

Output anomaly

A topic usually yields one output type but some conversations yield another — mislabeled or cross-functional?

Automation candidate

High-effort, zero-tool manual grind — a tool or skill could absorb it.

Under-provisioned thinking

Deep-tier work running without extended thinking — likely a quality lift.

Knowledge hub

Many people's work resembles one person's — a de-facto expert (and a dependency).

Orphan topic

A topic with real volume but no clear owner — assign one before it falls through the cracks.

Workload when & how hard people work

Off-hours load

A meaningful share of work happens outside the team's active hours — check on workload.

Weekend load

A meaningful share of activity lands on weekends — check workload and balance.

Crunch

Off-hours/weekend load jumped sharply vs baseline — possible crunch before it burns people out.

Working-hours creep

A person's active-hours window is widening over months — check workload.

Friction workflows that fight back

Retry / rework

Repeated short same-day attempts on a topic — the workflow or tool likely needs fixing.

Connect collaboration opportunities

Cross-pollination

Two people both heavy in a topic but with divergent approaches — introduce them.

Reinvention

Two people solving near-identical problems separately — point the later one at the earlier result.

Tools capabilities to spread

Undiffused tool

A useful capability is stuck with one person — spread it to the team.

Shared prompt template

Several people reuse a near-identical prompt — turn it into a shared template or command.

Roster membership hygiene

Underutilized seat

A person has a seat but almost no activity — onboard or reclaim it.

Role adoption gap

A person lags peers in the same role — a possible onboarding or enablement gap.

Multi-source provider spread

Multi-provider overlap

One person's work on a topic is split across multiple providers — consider consolidating.

Data quality trust the signals

High uncategorized

A large share of a person's work is unclassified — a blind spot to review before trusting other signals.

Win recognition

First deep task

A person just completed their first deep-tier task — recognize the milestone.

Comeback

A dormant person re-engaged after a long gap — welcome them back and learn what worked.

First adopter

The first person org-wide to use an output type new to the team — have them show others.

Efficient producer

Converts effort into concrete artifacts at an unusually high rate — worth learning from.

Governance sensitive activity admin-only

Sensitive category exposure

A sensitivity category (HR, legal, finance, security) has enough volume to warrant a governance heads-up — confirm access controls.

Sensitive activity by person

An individual with a notable count of sensitive conversations — confirm handling and access. (Categorized by the conversation classifier, with a keyword-lexicon fallback; only a derived category label is stored, never the text.)

Do-not-paste patterns in chats

Conversations contain a pattern your AI use policy says should stay out of chats (an email address, phone number, or card/ID number). (Matched by pattern shape against the classes you declare in your policy; only the class name is stored, never the matched text.)

The signals below join AI usage with sales outcomes from your CRM (HubSpot) — they appear only when a CRM is connected and the data is matched. All are descriptive associations, not causal claims, and are admin/RevOps-only.

Outcomes plays to spread (team-level) admin-only · needs a connected CRM

Winning workflow

The AI output type top-revenue reps produce far more of than the lightest-usage quartile — a play to capture and spread.

Winning play (subtopic)

The specific sales subtopic top-revenue reps use AI for far more than the lightest quartile — the play to codify and coach.

Agentic leverage

Top-revenue reps run more tool/agentic AI workflows on sales work — automate the steps the winners automate.

Cycle advantage

The heaviest-AI-usage quartile closes deals materially faster — package their workflow as a team playbook.

Usage without lift

Top-usage reps use far more AI but show no win-rate edge — check whether it's going to the highest-leverage work.

Outcomes people to act on admin-only · needs a connected CRM

Coaching candidate

A strong closer sitting in the lightest AI-usage quartile — a quick adoption win to lift their output.

Adoption win

A rep who recently ramped AI on sales work and is closing well — a within-rep adoption story to replicate.

Effort misallocation

A rep pours high-effort AI into sales work but converts below the team median — the workflow may be wrong.

Reinvention drag

A below-median closer keeps re-deriving AI work others already did — hand them the team's proven templates.

Expert, low yield

A de-facto AI expert whose own deals lag — coach on execution, or lean on them to teach the winning play.

Engagement decline

A rep's sales-topic AI usage is falling while results stay weak — an early churn signal to act on.

Denial loop

The auto-approve permission classifier blocked the same tool repeatedly — pre-approve it or serve it as a capability. (From Claude Code / Cowork OTEL tool-decision events; team-aggregate, no individual named.)

Error-retry loop

A tool errors and gets retried in the same session over and over — a broken step burning model turns. (From OTEL tool-result + api_request events; the wasted spend is the real tokens/cost between the error and the retry.)

Cache-inefficient sessions

Long sessions re-send full context every turn instead of reading it from cache — the fresh-input tokens are the cost driver. (From OTEL api_request cache-read/usage fields; team-aggregate.)

AI-Use Score

A number from 0 to 100 for how well your team uses AI — not just how much. It's the average of the five dimensions below (each also scored 0–100), using the weights you set. Every point traces back to something we counted in your own data — conversations, seats, dollars, shared-tool use — and the score page always shows the handful of numbers behind each dimension. Shown per team and for the whole org as a tier ribbon; the Weekly / Monthly / Quarterly toggle zooms the lookback (≈6 months / ≈18 months / ≈3 years). It's a directional management signal — read the tier and the direction, not the exact number. There is no individual score, and there never will be (see Privacy).

A dimension we can't measure yet is skipped, never guessed — the average is taken over the rest, so an unactivated Reuse or Governance dimension can't drag the number down.

The five dimensions

Adoption is the whole team on board?

Three ingredients, all counted from your usage: Activation (of the seats you pay for, how many were actually used this month), Regularity (how many days per month active people use AI — a daily habit beats a monthly binge), and Balance (whether usage is spread across the team or concentrated in one or two heavy users). Goes up when quiet seats come back to life, people build a habit, and usage stops depending on a single hero.

Sophistication is AI doing real work?

Two ingredients, graded on every conversation: Complexity (how big the task was — a quick lookup grades low, a multi-step build or analysis grades high) and Output (code, analysis, plans, and documents count for more than a quick email rewrite). We grade the work, not the typing — using a skill to do a complex job in one step counts as the complex job. Goes up when routine users graduate to bigger tasks and more conversations end in substantial output.

Efficiency how much of your spend does real work?

Six leaks, all counted in dollars and compared to your total spend: retry loops, abandoned sessions, error churn, re-sent context, model overkill (premium models on routine tasks — for most teams the biggest leak), and non-work spend (only if your AI use policy says it counts). We grade the leaks, not the bill: 100 means no measurable waste, and spending more never lowers this score — only wasteful spend does. Goes up when the waste dollars shrink — each Efficiency action names the exact leak and what it costs per month.

Reuse does the team run on shared, proven tools?

Two ingredients, counted across your shared tools (skills, templates, automations): Used, not redone (when work matches a shared tool the team already has, how often the tool was used instead of the work being done by hand) and Coverage (of the repeat work we can see, how much has a shared tool at all). We count the habit, not the logo — skills your team built before SightLift count from day one. Goes up when people adopt the standard tool instead of hand-rolling the work, and when repeated workflows get captured and published to your toolbox.
Not scored until we can see shared-tool activity or your first published capability — until then the score is simply the average of the other dimensions.

Governance is sensitive work in the right hands?

Two ingredients, measured only when you give us the context to judge fairly: Right hands (once we know people's functions — from your HR system or the role map in Settings — the share of sensitive work in legal, HR, finance, and security handled by people whose function covers it) and Policy compliance (once we have your AI use policy — sensitive work staying on company accounts, AI work staying on approved tools, personal data staying out of chats). Your governance score is the average of the rules we're actively checking, and each violation pattern becomes an action on your plan. Without roles or a policy, the dimension isn't scored and sensitive findings appear as notes instead. Goes up when sensitive work lands with the right people and policy violations fade to zero.

Weights you control

The score uses your weights — equal by default (each dimension counts the same). Set them in Settings → Company profile → Score weights to match the behaviors you want to drive, and change the balance as your priorities evolve. Weights are a strategy lever, not a way to inflate the number: raising a weight makes that dimension matter more and makes its gaps cost more.

How actions connect to the score

Every action on your Action plan names the dimension it moves and shows the points it could add — computed from your own numbers, with the math on the card (e.g. "re-activating 4 quiet seats raises Activation from 67% to 100% — about +3 points"). The plan is ranked by those points, biggest first. Two kinds of projection, always labeled:

Measured

Computed directly from observed counts — bypasses of a shared tool, waste dollars, quiet seats.

Estimate

Relies on one stated assumption, shown on the card (for example, "assumes this group's routine work moves up one complexity tier").

Some cards also show a dollar figure (like the monthly cost of manual work an automation would remove). Points rank the plan; dollars size the prize.

Tiers & movement

Banding is deliberate, so a 1-point wiggle can't read as a trend:

Needs focus

< 50

Developing

50–59

Solid

60–69

Strong

70+

↑ / → / ↓

Period-over-period direction (vs the previous bucket at the current grain). A change under 3 points shows as → (flat) on purpose — noise shouldn't look like movement.

Reading the ribbon

Teams

Grouped by explicit role/department where available, otherwise by each person's main work area (their dominant topic). The access role (admin/member/viewer) is never the grouping.

Hover a cell

Shows the subscore breakdown behind that score, each dimension with its own contribution.

Click a cell

Opens the conversations behind that team-month.

Not enough data

Teams below the confidence floor show "not enough data" rather than a shaky number — too few people or conversations to score reliably.

Privacy

Scores exist at the team and company level only. There is no individual score, and there never will be — managers drill into examples of work patterns, never into a ranked list of people. CRM outcomes, where connected, only help validate the weights — they are never an input to the score.