Want the long version? These guides go deeper than the definitions here.
Connecting a tool for the first time — the step-by-step setup guides.
Topics
Every conversation gets exactly one topic, decided by reading what it was actually about rather than by who filed it or what they called it — so nobody can change their numbers by renaming their work. Personal conversations are set aside from every summary by default; the note at the top of the page says how many there were.
Sub-topics
A finer split within each topic. Unlike topics, these are not a fixed list — they're drawn from your own work, then near-duplicates are merged so the same thing under three names doesn't look like three things. That means this list is yours and will look nothing like another company's. Click any sub-topic to read its conversations.
Output types
What the conversation actually produced — the thing you'd point at afterwards. Each conversation gets one, chosen from the fixed list below, so "we made 40 decks this month" is a question with an answer.
Complexity
How much the AI actually did — how many tools it reached for, whether it stopped to think things through, and whether it had files to work from. It deliberately ignores how much typing was involved: getting a hard job done in one well-aimed sentence is complex work, not simple work. That's effort, measured separately below.
how many tool calls (each one counts 1, up to 30)
+ extended thinking (counts 8 if it happened at all)
+ attachments (each one counts 2, up to 5)
- Simple
- Under 5 · a question and an answer, nothing else involved.
- Medium
- 5–14 · everyday assisted work with a few steps to it.
- Complex
- 15–29 · multi-step work with several tools in play.
- Deep
- 30 and up · long, tool-heavy work, usually with the AI thinking things through.
Effort
How much the person put in — how many times they came back, how much they wrote and handed over, and how many files they brought. This is the other half of the picture: two conversations can take the same effort and be worlds apart in complexity, and the pair together is far more telling than either alone.
how many times the person wrote (each counts 1.5, up to 20 turns)
+ how much they wrote or handed over (up to 30 — the largest term,
maxed out at ~12,000 input tokens.
Only what the PERSON wrote counts:
anything a tool handed back to the
AI mid-task is not their effort)
+ attachments (each one counts 3, up to 5)
- Quick
- Under 6 · asked once, gave little context.
- Light
- 6–17 · a few rounds of polishing.
- Substantial
- 18–34 · a real working session with genuine revisions.
- Deep
- 35 and up · a long collaboration, lots of back-and-forth, often with files.
- Unknown
- Not reported · this conversation came from a source that does not send us what the person wrote. We show the gap rather than guess a level, so it will appear as a plain segment on the effort bars.
Tokens
Tokens are the units AI providers bill in: roughly, pieces of words. A conversation that arrives as live telemetry, or in a Claude Code upload, carries the provider's own billed counts. A claude.ai or ChatGPT export carries only the text, so we count it with a real tokenizer rather than a "characters ÷ 4" rule of thumb. Either way the count is split by kind, because each kind is priced differently.
- Input tokens
- Everything the model reads on a turn: the conversation so far, the new message, and anything a tool handed back. These pile up as a conversation gets longer, which is why long sessions cost more than their word count suggests. On live telemetry this counts only the input that was not read from cache.
- Output tokens
- Everything the model writes: its replies, the tool calls it makes, and its extended thinking. There are usually far fewer of these, but each one costs several times more.
- Cache reads
- Input the provider had already stored from an earlier turn and reads back instead of processing again. A long agentic session re-reads its whole context on every turn, so this is usually by far the largest count. Each one is billed at a small fraction of the input rate; the fraction depends on the model. Live telemetry and Claude Code uploads report it. A claude.ai or ChatGPT export does not, so a conversation from one shows cache as not reported.
- Cache writes
- Input the provider stores so later turns can read it back cheaply. Each one is billed above the input rate: more for a cache that lasts an hour than for one that lasts five minutes.
- Agentic tokens
- The slice of input and output that is machinery rather than conversation. It is measured two ways, and the Cost tab keeps them apart. In an uploaded export it is the tool calls, what tools returned, and extended thinking, counted in the transcript. In live telemetry it is every model call after the first in each prompt, since those are the calls the work made on its own. Counting it separately is how the Cost tab can show what share of a bill went to autonomous work versus plain writing and answering.
- What isn't counted
- Images have no text to count, so they add nothing here — a picture-heavy conversation reads cheaper than it was. A single very large tool result is also capped, so one enormous file read can't dominate a month. Both make totals conservative rather than inflated.
Cost model
Published per-million-token rates × the tokens your team used. Reading and writing are priced separately because they differ by three to five times.
How sure we are — the three levels
The three levels say how well we know which model did the work — not how well we know the price. Every dollar we show, at every level, is a published list price applied to your token counts; none of them is read off your provider invoice, which is why each carries a ~ and an est. The Cost tab grades the window you have selected against these levels, and See by source there labels every provider with one of them — per provider, never for the company as a whole.
- ○ Estimated
- We can see the conversations but not which model ran them, so one mid-range list rate for that provider is applied to all of them. Good enough to compare topics, people and months against each other; treat the absolute total as a ballpark. This is where an uploaded export starts.
- ◐ Model mix
- Your provider's usage reporting tells us which models your team actually ran each day, so each day is priced on its own real blend of list rates. Token counts are calibrated against the provider's own reported totals too, which removes most of the remaining drift — the closest any level gets to your bill, though the dollars are still list prices rather than billed ones. The half-filled mark is also used for a partly-priced source — see below.
- ● Actual model
- Every conversation carries the model that actually handled it, so no model is blended or assumed and each is priced at the list rate we have on record for the day it ran. It is the highest level available, and a Compliance API connection or live telemetry unlocks it — but it is actual model, not actual cost: the dollars are still our list prices, not your invoice.
Levels are per provider, not per company — Claude can sit at Actual model while a second provider is still Estimated, and the tab says so. Connecting a source only ever raises what a provider is capable of; nothing you add takes a level away. Setting these up: the data sources guide ↗.
A provider can be part-way, and the label says so rather than rounding up. A level describes the dollars, and a conversation only reaches its provider's level if we can price it that way — Actual model needs that conversation's own model to be one we hold a rate for, Model mix needs its day to be in your provider's usage reporting. Anything that misses falls to the provider's fallback rate. So a provider whose conversations only partly qualify reads ◐ Actual model for 2 of 4 conversations — the half-filled mark, because part of it is, and the count and dollars behind it in the line beside it. If none of them qualify, it reads ○ Estimated and the line says which connection it has and why nothing landed: the connection is still there, but no dollar under it was priced any better than Estimated, and showing you the higher badge would be claiming an accuracy these numbers do not have.
Rates we price with (per million tokens)
Published list prices, kept current with each provider's pricing page. This table is today's — a past conversation keeps the list price we have on record for its own date, so a rate that changed since will not match the row here. If your company has negotiated rates these will read high; the comparisons between topics, teams and months hold either way.
These dollars are measured AI usage, priced at list. They are a different thing entirely from the labor dollars on the Action plan and Tools, which value people's time at the hourly rate you set.
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 offers an advisory read — it never approves anything; that happens in your own budget tool, which is why you type the “since” date in yourself (we can't see the approval).
The read
It reads efficiency, not return on investment: whether the spend looks like real work done efficiently — not what that work produced, which we can't yet trace from a session to the thing it shipped. Every threshold comes from your own team's numbers, never a fixed industry figure.
- 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.
What the work was
The person's tokens in the window, grouped by the subtopic each conversation was classified under, largest first. It describes the work and judges nothing, and it needs no company profile. The five largest are named and the rest are folded into Other. Subtopic not shown counts work at its topic when it has no subtopic, or when its subtopic is one you may not read: a sensitive category you are not cleared for, or a conversation not yet checked for sensitivity. It is the same rule that hides those subtopics everywhere else in the app. Unclassified is work the classifier has not placed, and it always gets its own row. Whether this work served a company objective is the separate section below it.
Alignment — is it real work?
The share of a person's tokens that went to work matching one of your company objectives (you write these in Configuration → Company profile). A conversation on a topic an objective names specifically gets full credit; one that only falls under a goal you set for the whole company gets partial credit — otherwise one company-wide goal would make everything read as perfectly aligned. Personal conversations never count as aligned, whatever the goal says. Whatever's left over is unclaimed spend. Topics come from what the conversation was actually about, so nobody can improve their own number by labeling their work differently. No company profile, no alignment figure. This shows where the money went; it never pushes anything up or down your action plan. The Cost tab's Aligned column and its “Share going to priority work” tile credit each conversation exactly the same way — they only differ in what they divide by (the Cost tab compares against work spend, this review against every dollar in the window, so personal use lowers the number here but not there).
Efficiency
Three stand-ins for “done efficiently,” all measured against your own team rather than a fixed cut-off. Where there's enough history each one also shows the same person's previous stretch of equal length (“was …”), so improvement is judged against themselves and not only against colleagues. Fewer than five earlier sessions is never shown as a trend.
- Cost vs peers on similar work
- Their tokens compared with what work of the same complexity typically costs across your company. About the same is normal; less is better.
- Repeated or duplicate work
- The share of tokens spent going over near-identical ground again. Less is better.
- Overspent on simple tasks
- Simple conversations that cost more than almost all other simple conversations — expensive effort on cheap tasks. Less is better. (Spotting the other kind of mismatch — a top-end model doing a rename — needs per-conversation model data we don't have on every source yet.)
Agentic share
Of everything the AI wrote for this person, the share that was machinery — tool calls and extended thinking — rather than plain answer text. High means tool-heavy, autonomous work; low means mostly writing and answering. Neither is good or bad on its own; it only becomes a warning sign when someone's share sits in the top quarter for your team, which is what feeds the “spinning” read below. (The Cost tab shows the same idea across all tokens, weighted by dollars.)
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 (the same prompt over and over), tool loops (machinery overhead — measured against the top quarter for your team once there is enough of your data to compare, and against a standard bar until then), no new ground, and simple tasks costing too much.
- 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 signs of spinning line up with little sign of progress → worth a look. Deliberately hard to trigger: wrongly flagging someone doing genuinely hard work is worse than funding some waste, so a short history is never flagged at all.
What this can't see: whether quiet work finished or was given up on — both look the same from here. Telling them apart needs a link from a session to what it produced, which doesn't exist yet. Until it does, steady progress toward a dead end reads as a hard problem.
Getting around
- The tabs, in order
- The rail is a ladder of scope, narrowest first, and only the tabs your role and your organization's plan allow are shown. My AI is your own month and nobody else's. Company is the summary for everyone. My team appears if you lead one, and names it — Sales team for a single department, My teams for several. Then, under Insights: People holds the score and the per-person views, Explore the raw data and the detailed signals, Cost the spend, Tools every AI skill your team uses, and Agents the audit of the agents your team runs. Under Workbench — the doing, rather than the reading — Action plan is what to do about it all, and Capabilities is where your own skills are proposed, proven and published. Under Outcomes, Revenue appears for people cleared to see customer and deal information.
- Tab names and group names
- Some tabs sit under a heading in the rail, and the two are deliberately not the same word: a heading names an area of the product, a tab names one page. The heading Workbench names the place you work — the Action plan of what to do, and the Capabilities you build to do it — while Insights next door is what you read. Outcomes stands over Revenue and is the wider word on purpose: it is where the other views of what AI-assisted work actually produced will appear. And a heading is not a promise that everything under it is yours. Two of them hold tabs sold separately — Agents under Insights, and Capabilities under Workbench — so you may well see a heading with only some of its tabs beneath it, or a single tab under a heading named for more than you have.
- Clicking a number
- Almost every figure here opens the conversations behind it. Click a bar, a chip, a table row, or a card and you get a filtered, sortable list — click again from inside it to narrow further. The line across the top always says which filters are applied.
- Bars split by effort
- On Explore, the "Team work by topic", "Team work by complexity" and "Team output types" panels use bar length for volume and the shading inside for the effort mix. Hover any segment for the numbers. (Cost’s "By topic" and "By complexity tier" charts look similar but plot spend, not effort.)
- Hovering a person
- Hovering a person's row pops up their last 30 days as a calendar — one cell per day, darker on busier days. Clicking the row expands it into their topics, outputs, trend, tools and hardest conversations.
- MTD
- Month-to-date — everything since the 1st, compared with last month at the same point rather than with a full month (which would always look like a drop).
- ⓘ and ↗
- An ⓘ opens this glossary at the matching definition. An arrow like How Tools works ↗ opens a longer guide in a new tab.
- Switching organizations
- If your account belongs to more than one organization, the block at the foot of the left rail shows which one you're looking at and lets you switch. Each one is completely separate — its own data, and your own role and permissions within it, which may differ from organization to organization.
- Connecting a data source
- In Configuration → Data & sources, each provider connects one of three ways. A live API connection uses a read-only key to pull new activity on a schedule, going back as far as that provider keeps history. Live telemetry has the tools report activity as it happens — the only way to see Cowork, and it starts from the day you turn it on rather than filling in the past. A manual upload takes an export file and brings in everything in it, which is how you backfill history or work around a blocked outbound key. You can change how often we sync under Configuration → Notifications. Step-by-step setup and a table of what each product can and can't report: the data sources guide ↗.
- What SightLift can see
- What you connect decides what we can measure, and some things simply aren't visible without a particular connection. Which level unlocks what, what we store, and how it's kept separate from every other company: data access & privacy ↗.
My AI
Your own page, and the only one that is about you rather than about the company. It shows your last month of AI work: what you worked on, how it breaks down by complexity and effort, the skills you reached for, and a calendar of your active days. Everyone gets it, whatever their role. It reads only your own activity — the server decides whose page it is, so there is nothing here that can be pointed at a colleague.
- Your month in AI
- The summary at the top: conversations, AI-active workdays, and the tool steps that ran on your behalf. The same measures used everywhere else in the product, narrowed to you.
- Your activity
- A calendar of the last 30 days, one cell per day, darker on busier days — the same strip that pops up when someone hovers your row on People.
- Three wins and one thing worth trying
- A deliberate cap, not a target. Up to three things that went well, and at most one suggestion — because six things to do differently is not a plan and one is. Fewer than three wins shows fewer rather than filler.
- How your teammates work differently
- Your work set against a comparison group. That group is your department when your organization's chart covers it, and the whole company when it does not — it widens rather than disappearing, and the label always says which one you are being compared with.
- Skills you have used
- The published capabilities and outside skills your own conversations reached for, so you can see what you are already running and what your team has published that you are not.
- Weekly email
- Your own subscription to the weekly digest, turned on or off here. It is your choice alone; an admin can see who subscribes but this switch is yours.
- “We do not know which activity is yours”
- A different thing from a quiet month, and kept on a separate screen for that reason. It means your account in your AI tool has not been matched to you here, so there is activity we cannot attribute. You can flag the match for an admin to review. A quiet month is a quiet month; this is a gap in the wiring, and reading one as the other is how a person concludes the product has nothing on them and never flags the match that would fix it.
- No score for you
- The AI-Use Score is never computed for an individual, so there is none on this page and there never will be. Nothing here ranks you against a named colleague either.
My team
The tab a leader gets for the departments they actually lead — not the company-wide People view, and not gated on the Insights module. It names itself after what you lead: Sales team when that is one department, My teams when it is several, and one block per department when you lead more than one.
- At a glance
- That department's AI-Use Score and which way it moved, using the same renderer and the same deadband as the score ribbon — so a team's score reads identically wherever you meet it, and a move under 3 points shows as flat here too.
- Highlights
- The good news worth passing on: the Win actions raised about people in the departments you lead. Already narrowed to your departments before it reaches you.
- Opportunities this week
- The handful of actions from the Action plan that are about your people, so a leader does not have to read the whole company's list to find their own.
- Skill adoption
- How far each published capability has spread in your department, with the headers and units spelled out. The denominator is everyone in the department, including people who have not started — which is the point of it. Both columns are all time.
- When a department is too small
- A team summary needs a minimum number of people; below that, the block says the department is too small to summarize and shows nothing instead. It is not an error and nothing is wrong — a summary of two people is a description of a person, and the same privacy floor applies here as everywhere else in the product.
Agent opportunities
Work your team does over and over that a custom agent or skill could take on. We find them in two passes: first by grouping conversations that cover the same ground, then by having a model read each group and judge how well it would suit automation. Every candidate lands in the Detected column of Capabilities with its full write-up; the strongest also become Automation actions on your Action plan. Both come from the same detection. This answers "what could we build?" — the Data Signals under Explore → Signals answer "what should we do about the people?"
- Ranking
- How often the work comes up, how many people across the team do it, and how well it would suit an agent — multiplied together, so something frequent but only one person's job doesn't beat something frequent and shared.
- Stars
- A one-to-five rating of how strong a candidate this is. More stars, better fit for automation.
- Tagline & problem
- One line saying what it is, plus the underlying problem this repeated work represents.
- What goes in, what comes out
- Roughly what you'd hand the agent and what it would hand back — enough to tell whether it's worth building, shown on the candidate's detail panel.
- Why it's strong
- The reasoning behind the rating: what specifically makes this pattern a good fit.
- Risks
- What to weigh before building — how accurate it needs to be, awkward cases, and where a person should stay in the loop.
- Savings estimate
- A rough figure for the time or cost an agent could save, shown only where the data supports one. Shown in dollars, that's the estimated time valued at the hourly rate you set.
- Who does this today
- The people doing this work now, plus real examples behind the candidate — click "Show spec" to read them.
Capabilities
A capability is one repeatable piece of work, packaged so your team can reuse it instead of redoing it. The Capabilities tab, under the Workbench heading in the rail, is where they're proposed, proven and published; the Tools tab is where you see how the published ones are doing. You don't build or configure any of this yourself — SightLift proposes them from work your team already repeats and asks you to approve the ones worth keeping. The full walkthrough is How capabilities work ↗.
Nothing is ever proposed off one person's habit: the same work has to show up several times, across more than one person, before it becomes a candidate.
- Automation
- Work where the same input reliably gives the same result — pull a list, reconcile a file, format an export. It runs as plain code with no AI involved at the time it runs, which is where the largest savings come from.
- Skill
- Work that needs judgment — drafting, summarizing, deciding. The steps are consistent even though the thinking isn't mechanical, so it runs as a guided recipe rather than fixed code. SightLift sizes the model for each step; your own tools run it, on your own account.
The columns on the board
Capabilities move left to right, becoming more trustworthy at each step. Nothing publishes itself — a person signs off before anything is served. A column only appears when something is sitting in it, so you won't see all of these at once.
- Detected
- We spotted repeated work and proposed something. Nothing is committed. Your move: promote it, or leave it.
- Testing
- Being checked, not yet trusted — either replayed against your team's own past results, or running quietly alongside real work while it collects results people confirm. Usually nothing for you to do; you're told when it's ready.
- Verified
- It passed, and the evidence is attached. Your move: approve it to publish, or hold.
- Live
- Published into your tools and being measured — how often it's used, how often people keep the result, and what it's actually saving.
- Needs attention
- Something live has drifted away from what people do now. The work changed underneath it. Your move: re-check it against fresh examples, or retire it.
- Retired
- Unpublished and no longer served — retiring it removes the tool from your team's assistants. Nothing is deleted: it stays on the board with its proof intact, so you keep a record of what you used to run.
How strongly something is proven
The proof is the whole point — it's what lets you trust a capability without reading its internals. We say plainly which kind you're getting, and never dress one up as another.
- Verified on real work
- It reproduced your team's own past results exactly. The strongest proof there is, and only automations can earn it.
- Confirmed in live use
- It earned its proof from real results people confirmed, because there was nothing in the past to replay against. An honest notch below the above.
- Quality-checked
- A graded assessment, used for skills. Judgment work can't be certified by reproducing it, so a skill is graded and never called "proven" — and it says so wherever it's served.
- Failing forward
- When a candidate can't reproduce every example, it isn't thrown away — it's offered as a skill instead, with the differences shown. A "no" to automation is really "this part needs judgment."
Testing runs in a sealed sandbox: no data leaves it, and nothing runs against anything live. Every proof keeps a de-identified copy of the examples it passed, so it stays valid even if the original conversations are later deleted — and the person who approved it and the date are always attached.
Labels you'll see on a card
These are the small labels that appear on individual cards. Each one is meant to read on its own — this is the reference, not the explanation you need to make sense of them.
- replaying past results
- We're re-running it against work your team already did, to see whether it reproduces the same results. Nothing is served while this happens.
- testing on live work
- It runs alongside real work without acting on anything, so your team can confirm its results before it's published.
- ran on your real inputs
- The test used the same inputs your team actually used — not invented examples.
- same result every run
- It produced an identical result on every run, with no model involved at the time it ran.
- only N examples so far
- The proof rests on a small number of examples. It's honest, just early — it strengthens on its own as your team confirms more runs.
- only N people so far
- Only that many people did this work, so there's less evidence behind it than usual. Worth confirming the result before relying on it.
- no decision in N weeks
- This has sat in Detected that long without being promoted or turned down. If it isn't worth building, deciding against it clears it off the board and stops it counting as outstanding work.
- requested by your agents
- Your team's agents asked for something like this at the point of use and nothing matched. A request from your tools, not a pattern we spotted.
- added by you
- Someone on your team wrote this one, rather than SightLift detecting it from repeated work.
- Not repeated enough yet
- Work we found repeating, but not often enough or across enough people to be worth building. It isn't counted as outstanding work and we won't raise it as an action — you can still build it if you disagree.
Agents
The register of every agent your team runs, wherever it was built, and what its runs actually did. This is the one tab about work SightLift did not do itself: the agents live in your own systems and we read their runs. It is a separate purchase — a team that does not have it opens the tab and gets a page describing what the audit is, rather than an empty register. Its loop is four steps: read the runs, grade them against criteria you write, change the criteria, then re-grade and compare.
The vocabulary
- Agent
- One automation your team runs, named and counted here whichever system it was built in. The register says which system each came from.
- Run
- One execution of one agent. A run carries what it produced, how long it took, how many model calls it made, how much of its input was reused from cache, and what it cost.
- Trigger
- What starts an agent: a schedule (one or several), or an event or a person. Read the wording carefully — where it says inferred, not read from the system, we worked the trigger out from the runs because the tool does not report one. For Claude Code that is every row, because its telemetry carries nothing marking a run as scheduled. “Not recorded” means there was nothing to infer from.
- Criteria
- What you say a good run looks like. Criteria live on the agent, not on the run, and every change makes a new version — so a pass rate always says which criteria produced it, and the run charts mark where a version changed. That is what makes a jump in the pass rate attributable to your criteria changing rather than to the agent's behavior changing.
- Whose verdict it is
- Where the system that ran the agent graded it too, that verdict is shown as theirs and never merged into yours. Two graders, two vocabularies, kept apart on purpose: a run can pass yours and fail theirs, and averaging them would hide exactly that.
How a run is graded
- Passed
- The run met the criteria in force at the time.
- Partial
- It met some of them. Kept as its own outcome rather than folded into either neighbor, because a chart that buries it is the screen deciding for you whether a partial run passed.
- Failed
- It did not meet the criteria. Worth asking whether the agent needs fixing or the criteria do — the tab is built to tell those two apart.
- Not graded
- Nothing has graded this run: no criteria applied to it, or the grading has not happened yet. Not a failure, and not counted as one.
- Did not start
- A trigger fired and produced no run at all. We cannot tell you why: from here, a machine that was off, a paused schedule and an outright failure look identical. Shown rather than dropped, because a trigger that never produces a run is the failure a pass rate cannot see.
What the figures will and will not say
- Spend on runs
- Split into what the runs that passed cost and what was spent on the runs that did not — the second being the more useful number, since it is what the failures cost you. “Spend per run that passed” is what a working run costs.
- When a total is a floor
- A run missing a vendor's price is left out of every published total, and the figure says so and names the vendor. So the number is a floor, not a guess, and never a silent zero.
- When no total is shown at all
- Rather than adding up things that cannot be added, the tab states the reason: runs priced in more than one currency, runs carrying no currency, or no run with a complete price. An absent figure here is a refusal, not a nil result.
- Counts cover what we can read
- A count of agents is a count of the agents we can enumerate. Where we can neither read the runs nor count the agents, no number is quoted at all rather than one we cannot stand behind — so nothing on this tab should be read as “this is everything your team runs”.
- History starts when telemetry did
- Runs are read from live telemetry, so the record begins the day it was turned on and does not fill in the past. A short history is a short history, not a quiet team.
- Schedule health
- Paused and archived schedules are counted beside the agent they fire, since there is no separate place for them. A status we inferred is labeled as inferred rather than presented as one the vendor confirmed.
Actions
The Action plan takes the many Data Signals about the same person, topic or group and bundles them into a short, ranked list of things to actually do: "coach this person", "spread this topic beyond one owner", "run an AI 101 for this group". Every action shows the signals and conversations it came from (click a card to read them), says which part of the AI-Use Score it would move, and how many points it would add. The queue groups them by the part of the score they move, weakest part first, and puts the biggest gain first inside each group. Actions that name an individual are shown only to admins.
Each one carries a single category — the colored chip — which decides what kind of thing it's asking of you and what document you can generate from it:
- 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 out work that sits with too few people, or agree on a way of working — one person holding a whole topic, a topic nobody owns, or a team all doing the same thing differently. Generates talking points.
- Automation
- Build or roll out a tool — work being done by hand over and over, the same prompt reinvented by different people, or a strong automation candidate. The candidate's own panel holds the write-up you'd build from.
- Cost
- Cut spend or waste — one or two people driving an outsized share of the bill, a top-end model doing routine work, or the same job being done twice on two different providers. Generates a talking-points brief.
- Enablement
- Run an "AI 101" for a group stuck on small, hands-on tasks in the same area. Admins see who's in the group; everyone else sees only how many. Generates a session outline.
- Governance
- Sit down with whoever owns the data and review how much sensitive work is happening and, for legal, finance, personnel, security and strategic work, who's doing it.
- Win
- Recognize a positive pattern — a comeback, a first adopter, or a winning workflow worth spreading. Visible to everyone.
- Where actions appear
- Most sit in the ranked queue. Three kinds have their own sections instead. Revenue actions are ordered by what they are worth on sales results, and Agents to improve also appear on that agent’s page under What to change; neither moves the AI-Use Score. Worth sharing holds Win actions: good things that already happened, to pass on rather than do. Marking an action done or dismissing it keeps it on record with who made the call.
- Priority
- The 0–100 pill adds up the signals underneath: more signals, and more serious ones, rank higher. The list you see is kept deliberately mixed, so one category can't crowd out everything else.
- Evidence
- How many signals and how many conversations sit behind the action. Click the card to read them.
- Generate a doc
- Coaching, structure, cost and enablement actions have a 📄 button that drafts a ready-to-use document — read it in place, then copy or download it.
- Who can see what
- Anything naming a person or the members of a group is limited to admins and to people given sensitive-data access. Those names never travel — not in an email, a Slack message, or a downloaded document. You only ever see them inside the app.
Dollars on an action — your company rate
Some automation actions carry a dollar figure ("~$4,200/mo of manual time", "~$800 of avoidable rework"), as do the savings on a published capability. These are labor dollars, and the arithmetic is deliberately simple: estimated time × your company rate. The time comes from your own data: how long this work is reckoned to take, multiplied by how often we actually saw it happen. The rate is a company-wide number an admin sets in Configuration → Cost → Budget and labor rates: an average hourly cost for the people doing this work, salary plus everything on top of it. Until someone sets it, a default is used, and every figure that depends on it says which rate it used when you hover. A team can carry its own rate, because a support hour and an engineering hour do not cost the same, and a team figure names the rate that priced it. Rates are dated, so changing one prices from that date forward and leaves what you have already reported where it was.
This is money your team could redirect, not money that appears in a bank account — it values hours at what you told us an hour costs. It is also entirely separate from the dollars on the Cost tab and in an action's score impact, which are measured AI spend and carry no wage assumption at all.
Revenue & accounts
The Revenue tab asks where AI-assisted work goes across your customers, and what it changes. Accounts are picked up from conversations with no setup at all; uploading your account list or connecting your CRM firms up the matches and unlocks the deal views. It appears for admins and for people an admin has given revenue access (Configuration → Users & roles) — and only once there's account or deal data to show, so a new grant on an empty tenant isn't broken, there's just nothing to put in it yet.
Everything here that compares outcomes is a description of what went together, not proof that one caused the other. Rows are your customers — this tab never ranks, totals or compares your own staff. "Worked by" is a fact about the account, the same way your CRM shows who touched it.
The two views
- Accounts
- Every customer your team has done AI-assisted work for, how much, who worked it, and when. Available with no setup — this is the view you get on day one.
- Outcomes
- What the prep on your own deals is worth, and what to scale. Which prep patterns show up in won deals more than lost ones, by topic and by served skill; each topic's leading metric week by week; how closed deals with AI-assisted work compare with those without; and which reps apply the winning patterns, as a fact and never a ranking. Each section needs its own data and says so plainly when it's missing. Accounts getting no AI-assisted work are a filter on the Accounts list: No AI work yet.
How sure we are it's really that account
Three ways an account gets identified, and the difference is always visible on the row — a guess never gets styled like a certainty.
- From a web or email domain
- The conversation mentioned a domain that belongs to a known account. The strongest match, and it renders plainly.
- From your list
- The name matches an account you uploaded or that came from your CRM. Also renders plainly.
- Detected ~
- A company-shaped name that keeps coming up, with nothing to anchor it to. Shown with a ~ and a "detected" chip because it is a guess — it can pick up a partner, a competitor, or a name that just looks like a company. Confirm the ones that are real in Configuration → Company profile → Accounts & matching; until you do, they never write back anywhere.
Terms on this tab
- Contact
- A person at an account, from the CRM record. These are your customers’ people; your own team appears as worked by. Contacts arrive only through the CRM sync, so an account SightLift has never synced has none.
- Stage
- The deal’s own stage, in your CRM’s stage names, never renamed here. A stage your CRM has not given us a name for shows as its stage ID. SightLift reads the stage on each sync and records each change it sees, so a deal that moved back a step is kept as a move rather than overwritten. For each deal it also asks your CRM once for the stage history from before SightLift first saw it. Where work is dated earlier than any stage SightLift holds for the deal, the stage reads as unknown rather than guessed. The Accounts list asks two stage questions: the Stage filter finds accounts by where an open deal stands now, and Stage mix shows the stage each conversation sat at on its own date.
- Play
- A prep pattern your team has decided is practice: at a stage, on open deals above your large-deal line, the prep of a topic or a served skill, aimed at a contact role, within some days of entering the stage. A manager blesses it from Outcomes, from a pattern with lift or by declaring it, and becomes its one owner; a rep can propose one. Its page shows where it is applied and where it is missing, with one action per missing account in the Action plan. A person closes a play; nothing closes it on its own. A play is a different thing from an Action plan action.
- Single-threaded
- Only one person has done AI-assisted work on this account. Not a problem by itself, but worth knowing before that person takes leave or moves on.
- Shared sessions
- A conversation that covered more than one account. Its effort and dollars are split across them, so nothing gets double-counted.
- The 12-week strip
- Account work comes in bursts rather than a steady trickle, so recent weeks are drawn as separate cells instead of a line. A line would imply a trend that isn't really there.
- Untouched large deals
- Open deals above the tab’s large-deal threshold with no AI-assisted work behind them. That threshold is your own: SightLift sets it near the top quarter of your priced open pipeline (rounded down to a round number, so the board is a little wider than a strict quartile), falls back to a shipped figure while there are too few deals to do that, and the view states which wherever it uses it. Not a verdict on the deal — a list of the biggest places your team's proven approaches aren't being used.
- Handoff brief
- A one-page summary of everything the tab knows about an account, for handing over to a colleague. It's built only from facts already on screen, and nothing leaves until you've read it and said so.
- Small accounts still show
- Unlike the people-based views, there is no minimum here — a customer appears even if one person worked it once. These are companies, not employees, so the privacy floors that protect individuals don't apply. Sensitive conversations are still held back, and the tab says how many.
Data Signals
Under Explore → Signals you'll find these — patterns found by counting, the same way every time, across every conversation, and sorted by how much they matter. Each one is grouped into a family (the colored chips below). A signal points at a topic, a person or a habit, and always links to the exact conversations behind it; click any card to read them. Signals that name someone, or that touch sensitive work, are shown only to admins.
These are the raw findings. The Action plan is where many signals about the same person or topic get bundled into one thing to actually do.
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, and the data does not say whether it was automated, moved elsewhere, or abandoned.
- Consolidating
- A topic is shifting into fewer, longer sessions.
- Output mix shift
- A topic's mix of output types changed notably (e.g. a surge in one artifact).
- Tool-use spike
- Agentic and tool use in a topic jumped. Measured month over month: the last full month against the one before it.
- Seasonal
- A topic recurs on a predictable cycle.
Person individual trajectories
- Rising complexity
- A person's work is steadily getting more complex.
- Stuck on simple
- A tenured person's work has stayed on simple tasks.
- Dormant
- A once-active person has gone idle.
- High effort, few tools
- Heavy work done with few tools.
- Chat-only
- Real volume with zero tool use, on a team where peers do use tools.
- Prompt retyper
- One person retypes a near-identical prompt over and over.
- Plateau
- A once-rising person's complexity rose and then went flat.
- Low artifact yield
- Heavy effort, with few of those conversations producing a concrete artifact.
Cost token spend & model fit
- Topic cost concentration
- A topic is a large share of token spend.
- Person cost concentration
- A person is a large share of token spend.
- Token-efficiency outlier
- A person uses far more tokens per conversation than peers on the same topic.
- Budget review candidate
- Heavy recent spend that looks like circling, or that we can't call either way. People doing genuinely hard work, and people with too little history to judge, 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. Teams doing measurably heavier work are never flagged. (Counted per team, never per person — the individual figure is admin-only.)
- Model downgrade
- Lots of simple-tier work on an expensive (Opus-class) model. (Only where we can see which model ran each conversation.)
- Model underpowered
- Complex and deep work concentrated on the cheapest model available. (Only where we can see which model ran each conversation — a claude.ai export doesn't say.)
- Model-mix savings
- Org-level run-rate: routing routine (Simple/Medium) work to a cheaper model in the same family would save a meaningful monthly amount. (Needs a live connection to your provider's usage reporting; the figure is an estimate and is marked as one.)
Risk concentration & provisioning
- Hidden expert
- Deep work in a topic is held by a single person or a pair.
- Bus factor
- One person owns a large share of a topic.
- Hard-case outliers
- A mostly-simple topic with a few hard outliers.
- Complexity variance
- Complexity varies widely across the people working on one topic.
- Over-provisioned thinking
- Extended thinking left on for one-line lookups: questions answered in a single exchange, with no tools.
- Effort ↔ tier mismatch
- High effort spent on simple-tier tasks.
- Abandoned long conversations
- Long conversations that produced nothing concrete.
- No agreed workflow
- No consistent output pattern across a team for a topic.
- Output anomaly
- A topic usually yields one output type, and some conversations yield another.
- Automation candidate
- High-effort work in a topic with no tool use at all.
- Under-provisioned thinking
- Deep-tier work running without extended thinking.
- Knowledge hub
- Many people's work resembles one person's.
- Orphan topic
- A topic with real volume and no clear owner.
Workload when & how hard people work
- Off-hours load
- A meaningful share of work happens outside the team's active hours.
- Weekend load
- A meaningful share of activity lands on weekends.
- Crunch
- Off-hours and weekend load jumped sharply against the team's own baseline.
- Working-hours creep
- A person's active-hours window has widened between the earlier and more recent halves of their work.
Friction workflows that fight back
- Retry / rework
- Repeated short same-day attempts on a topic.
Connect collaboration opportunities
- Cross-pollination
- A pair of people both heavy in a topic, working on it in divergent ways. Measured over every conversation on record.
- Reinvention
- A pair of people solving near-identical problems separately.
Tools capabilities to spread
- Undiffused tool
- A useful capability is being used by only one person.
- Shared prompt template
- Several people reuse a near-identical prompt.
Roster membership hygiene
- Underutilized seat
- An account has a seat but almost no activity.
- Role adoption gap
- A person is well below peers in the same role on AI usage.
Multi-source tool spread
- Multi-tool overlap
- One person's work on a topic is split across multiple tools.
Data quality trust the signals
- High uncategorized
- A large share of a person's work is unclassified, which is a blind spot for every other signal about them.
Win recognition
- First deep task
- A person just completed their first deep-tier task. Measured over the last full month in the data.
- Comeback
- A dormant person re-engaged after a long gap. Measured over the last full month in the data, after 3 or more months with no conversations.
- First adopter
- The first person org-wide to use an output type new to the team. Measured over the last 2 full months in the data.
- Efficient producer
- Effort turns into concrete artifacts at an unusually high rate. Measured over every conversation on record.
Governance sensitive activity admin-only
- Sensitive category exposure
- Sensitive work in one category (Legal, Finance, Personnel, Security, Strategic Internal or Strategic External) by people whose clearance in Settings does not cover it. Only work outside a recorded clearance counts, so work a person is cleared for raises nothing, and a person with no clearance recorded is never counted. Personal and Health, union, religion are never counted, because nobody can be cleared for them. A category is reported only when it spans enough people that nobody is singled out, and never per person. When your policy checks sensitive routing, the same people can also have their own alerts. Those may count only the last 30 days, where this one counts all their work, so the two need not add up. This one is kept beside them because it shows one category reaching several people at once.
- Sensitive work by someone not cleared
- A person did sensitive work in a category their clearance in Settings does not cover. A person with no clearance recorded is never counted. (The category comes from reading the conversation, with a keyword check as a backstop; only the category name 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, card/ID number, or a secret such as an API key). (Matched on the shape of the pattern against the kinds you named in your policy; only the kind is stored, never what was matched.)
- A rule your admins added
- Conversations matched a rule an admin added by hand in Settings, checked by its own kind: a word, a shape, a tool, a model, an account, the time, a department, a clearance, a cost, a label SightLift already gives, or SightLift's model. (A word your admins listed is shown to admins only. A rule added by hand raises alerts and does not count toward the governance score.)
- Sensitive work off the roster
- Sensitive conversations ran on accounts that aren't on your uploaded roster: either personal logins, or people missing from the roster file. Personal and Health, union, religion are never counted, because taking one person off the roster would show how much of that work is theirs. (We can only see whether an account appears in the roster you uploaded, so the action asks which of the two it is rather than assuming.)
- Source we can't check yet
- Work ran on a source SightLift can't match to a tool family: either one we haven't classified yet, or a conversation that reached us with no source recorded at all. Either way it counts against your approved-tools score until we fix it, and that part is on us. (This one is on us, not you — it stays a plain note and is never ranked or priced, because classifying the tool might reveal a family your policy doesn't approve rather than clearing the gap.)
- Tool outside the approved list
- Work ran on a tool your AI use policy doesn't list under approved_sources. (Only tools we can identify are named here — work on a source we haven't classified counts against the check but is never attributed to you.)
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. Measured over every conversation on record, for the reps matched to the deals your CRM outcomes hold: the last 12 months a CRM sync reads, or the months an imported file lists.
- Winning play (subtopic)
- The specific sales subtopic top-revenue reps use AI for far more than the lightest quartile. Measured over every sales conversation on record, for the reps matched to the deals your CRM outcomes hold: the last 12 months a CRM sync reads, or the months an imported file lists.
- Agentic leverage
- Top-revenue reps run more tool and agentic AI workflows on sales work. Measured over every sales conversation on record, for the reps matched to the deals your CRM outcomes hold: the last 12 months a CRM sync reads, or the months an imported file lists.
- Cycle advantage
- The heaviest-AI-usage quartile closes deals materially faster. Measured over the won deals among the deals your CRM outcomes hold: the last 12 months a CRM sync reads, or the months an imported file lists.
- Usage without lift
- Top-usage reps use far more AI and show no win-rate edge for it.
Outcomes people to act on admin-only · needs a connected CRM
- Coaching candidate
- A strong closer sitting in the lightest AI-usage quartile.
- Adoption win
- A rep who recently ramped AI on sales work and is closing well. Measured over the later half of the months on record against the earlier half, with the close rate over the deals your CRM outcomes hold: the last 12 months a CRM sync reads, or the months an imported file lists.
- Effort misallocation
- A rep puts high-effort AI into sales work and converts below the team median.
- Reinvention drag
- A rep whose AI work repeats work others already did, and who converts below the team median.
- Expert, low yield
- A person whose AI work many others resemble, whose own deals lag the team median.
- Engagement decline
- A rep's sales-topic AI usage is falling while results stay below the median.
- Denial loop
- The same tool got blocked for approval over and over. (From live telemetry; counted per team, nobody named.)
- Error-retry loop
- A tool errors and gets retried in the same session over and over. (From live telemetry; the wasted spend is the real tokens burned between the error and the retry.)
- Repeated context sessions
- Long sessions send the whole conversation again every turn instead of reusing what was already sent, and that repetition is the cost. The saving beside it is what caching that repeated context would return at the published cache rates, net of the premium charged on the turn that writes the cache. It assumes the repeated part really is a stable prefix, which per-session totals cannot prove, so it is the ceiling on a suggestion rather than a figure we have collected. (From live telemetry; counted per team.)
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 ribbon; the Weekly / Monthly / Quarterly toggle zooms the lookback (≈6 months / ≈18 months / ≈3 years). It's a steer, not a measurement instrument — read which way it's moving, not a 1-point difference between two teams. There is no individual score, and there never will be (see Privacy). The full walkthrough is How your AI-Use Score works ↗.
composite = weighted average of the dimensions we can measure
(weights equal by default — set them in Settings)
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 (of the workdays in the month, how many your active people use AI on — a daily habit beats a monthly binge), and Depth (of the seats you pay for, how many belong to someone for whom AI is part of the working week: at least three workdays in five, in at least two different places). Goes up when quiet seats come back to life, people build a habit, and heavy use spreads beyond a few individuals.
- Sophistication is AI doing real work?
- Two ingredients: Advanced practice (of the seats you pay for, how many belong to someone whose work goes through a skill, a shared tool your team chose, or a sub-agent — it counts people, not conversations, and the tools that ship with the assistant don’t count) 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 advanced practice spreads to more people, and when 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 each person's Function Sensitivity — set in Configuration → Company profile → Org & roles — the share of sensitive work in legal, finance, personnel, security and strategic matters handled by people cleared for 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 Configuration → 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
- Worked out straight from things we counted — times a shared tool was skipped, wasted dollars, unused seats.
- Estimate
- Rests on one assumption, which is always written on the card — for example, "assumes this group's routine work moves up one level of complexity".
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.
Movement
We report where the score is and what it did, and we don't grade it. A composite out of 100 is a measurement; a grading word laid over it is a verdict we'd be handing you about your own team — so the number stays plain and the change sits beside it. Movement is deliberately banded, so a 1-point wiggle can't read as a trend:
- ↑ / → / ↓
- Period-over-period direction on the ribbon (vs the previous bucket at the current grain). A change under 3 points shows as → (flat) on purpose — noise shouldn't look like movement.
- ▲ 3 this month
- The same rule on the score itself, with the size of the move spelled out. Under 3 points it reads steady rather than a number, for the reason above; before there is a previous period to compare against, it says so instead of reporting zero.
Reading the ribbon
- Teams
- Grouped by the department you set in Configuration → Company profile → Org & roles where you've set one, and otherwise by what each person mostly works on. Someone's access level in SightLift — admin, contributor, user — is never used as a team.
- Hover a cell
- Shows what's behind that score — each of the five dimensions and what it contributed.
- Click a cell
- Opens the conversations behind that team-month.
- Not enough data
- A team with too few people or too few conversations to score reliably says so, instead of showing a number that would wobble on one person's week off.
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.