Classify people in People
Let the org's own model sort everyone in Venom People by sector, persona, seniority and opportunity, judge how relevant each person is to your focus, and search and sort by it, under a daily $ cap.
Classification reads what People already knows about a person (role, company, location, tags, how you met, notes, the event) and asks the org's model five things:
- sector, persona and seniority, each one value from a list you control;
- opportunities: what this person could be for you (customer, partner, investor, hire, connector, advisor, competitor, vendor), zero or more;
- relevance: high, medium, low or unknown, judged only against your focus, with a one-line reason;
- a short headline, up to three topics, a suggested next step, and how confident it is;
- in the same call, only when the text says it: the country they live or work in, gender (from pronouns or a title), and up to three personality words. These feed People's derived fields, where the AI's answer beats the rules and your own value beats both. A person classified before this was added gets them on their next classification (a change to the person, or the re-check after the days you set); nobody is re-sent just for these.
The result shows on the person and in search, and you can filter and sort by it. It is a suggestion: change any field by hand and your value stays, even when the person is classified again.
Turn it on
It is off by default. An owner or admin turns it on under People → Settings (/people/settings), Classification:
| Setting | Default | What it does |
|---|---|---|
| Focus | empty | What you care about, in plain words ("sovereign AI for banks in Asia"). Relevance is judged against it; without one, relevance is unknown unless a note says why someone matters |
| Lists | the defaults below | The values for sector, persona, seniority and opportunity. other (and unknown for seniority) is always kept as the fallback |
| Gear | the org's default | Which gear's model answers: any of your organisation's gears, or a venom built-in (qwen, china, …) unless your organisation allows only its own. The run uses its summary model, else its main model. A gear that cannot run now is listed greyed out with the reason, and a saved one that stops working stays selected with its reason |
| Schedule | daily | off, hourly or daily: how often the scheduler classifies new and changed people |
| Daily cap | $1 | Classification's own spend per UTC day |
| People per run | 200 | A run stops after this many |
| Batch size | 8 | People per model call |
| Reclassify after | 30 days | A person unchanged for this long is checked again |
| Topics as tags | off | Adds up to three topic tags to the person, reusing tags you already have first |
| Sector as a tag | off | Also adds the sector (never other) as a tag |
| Minimum confidence for tags | 0.7 | Topics and the sector become tags only from an answer at least this confident |
| Learn from corrections | on | Shows the model your hand corrections as examples (below) |
| Examples per prompt | 6 | At most this many corrections per prompt, 0 to 20 |
| Suggest new categories after a run | on | One model call after a run looks for missing list values (below) |
| Category scans per day | 2 | At most this many scans a UTC day, after runs and on demand together; 0 turns them off |
Changing the focus or a list starts a new taxonomy version: everyone is classified again on the next run. Turning it off stops a running run; the classifications already made stay.
Default lists:
sector: ai-infra, fintech, banking, asset-management, web3-crypto, vc-investing, legal-compliance,
government-public, consulting-services, media-marketing, academia-research, healthcare, other
persona: founder, investor, executive, engineer, product, bd-sales, advisor-consultant,
researcher-academic, student, government, service-provider, other
seniority: c-level, partner-director, manager, individual-contributor, student, unknown
opportunity: customer, partner, investor, hire, connector, advisor, competitor, vendorRuns
Estimate shows how many people are waiting and the calls, tokens and $ range a run would take (from the gear's model prices; tokens only when no price is known). It spends nothing.
Run now starts a run. The page runs it in steps and shows the progress; you can leave, and the scheduler finishes it. One run per org at a time. A run classifies only people who are new, changed since they were last classified, or older than Reclassify after: an unchanged person is skipped without a model call.
A run stops at whichever comes first, and says which:
| Stop reason | Means |
|---|---|
done |
No one left to classify |
limit |
It reached People per run |
classify_cap |
Classification's daily cap is spent |
people_cap |
People's own daily AI cap (shared with cards, enrichment and the other People features) is spent |
model_error |
The model answered with an error; the run's error says which |
no_model |
The gear cannot run now; the run's error gives the same reason the settings show (no provider key for its model, an offline daemon, a wildcard that matches nothing) |
off |
Someone turned classification off |
The runs table lists each run's people, calls and cost, with a link to its requests in the log explorer
(session:"people-classify:<run id>").
How it learns
The section How classification learns, under the classification settings, holds four things.
Corrections become examples
When you set a field by hand and it differs from what the model said, People keeps it as a correction: a short summary of what the model read (role, company, location, tags, the start of the notes; never the name or handle values), what it answered and what you chose. The next run shows the model up to Examples per prompt of your corrections for sector, persona, seniority, relevance and opportunities, as examples to follow for similar people. Free-text fields (headline, reason, next step, topics) count for the quality signal but are not shown as examples, so the model does not copy them.
Examples are picked once per run and owner (the newest correction per person and field, values still in your lists, spread over fields and people), so every batch of that owner's people sends the same prompt and the provider can cache it. Your corrections are used only for your own people: a classification is its owner's only. A correction stops being an example when you clear or change the hand value. The estimate counts the examples' tokens.
Suggested categories
After a run finishes, if people landed in other or below 0.5 confidence since the last look, one model call reads up
to 60 of them (role, company, headline, the current categories and tags; never names, notes or handles) and proposes
new sector, persona or opportunity values, each with how many people it fits and up to three examples (role and
company). It needs at least three such people and a value must fit at least two. Look for new categories does the
same on demand; Estimate a scan first shows the people and $ and spends nothing.
Add to list puts the value in its list before other. That is a new taxonomy version, so the next runs classify
everyone again, as any list change does. Dismiss means the value is never proposed again. Scans stop at
Category scans per day and at both daily caps.
Auto-tags
With Topics as tags or Sector as a tag on, a run adds those tags to people whose answer is at least the Minimum confidence for tags. Preview tags shows which tags the people already classified would get, and how many people each, with the settings on screen and without writing anything. Add these tags now adds them with the saved settings. Tags are only ever added, never removed, and adding them does not make anyone pending again.
Is it getting better?
For the last 30 days, the table shows each field's override rate (corrections ÷ people classified) for people classified without examples and with them. A lower rate with examples means your corrections are helping.
What is sent to the model
Only the fields listed above, with each person under a reference (p1, p2, …). Handle values (emails, phone
numbers, profile links) are never sent, only which kinds a person has. Notes are cut at 1,500 characters. The call goes
through the org's own route with People's internal key, so it is booked, capped and logged like any other request.
Who sees what
A classification belongs to the person's owner: only they see it, filter by it or change it. Classification runs over every member's people in the org, each batch holding one owner's people.
Search and sort
The People search takes sector, persona, seniority, relevance and opportunity filters, and sort: relevance
(high first, then by confidence). They apply to your own people only.
Browse
The list at /people uses the classification once your people have one:
- Filters: Sector, Persona, Seniority, Relevance and Opportunity sit with the other filters (see
Filters), each value with how many of your people have it. Seniority and Sector filter People's
derived fields: the classification's answer when there is one, else the rules' reading of the job title and company,
so they work before classification is on. Seniority is a scale from 1 (student or intern) to 6 (C-level or founder);
a word instead (
seniority=c-level, or a value of your own list such asvp) matches what the classification said, as Insights counts it, and a word nobody has matches nobody. Sectorotheris the people the classification read and found no sector for (and no rule did either), apart from those it knows nothing about. A filter that is on shows as a chip above the list; click the chip to remove it. They are in the URL (/people?sector=fintech&relevance=high), so a filtered list can be bookmarked or shared. On a phone they fold behind the filter button. - Sort: Relevance puts high first, then medium, low, unknown and people never classified, each by confidence.
- Rows: the headline replaces the role under the name (the company stays after it), with a small relevance badge. Wider screens add the sector and persona as chips.
- Group by: none, sector, persona, seniority, relevance, company or the event you met them at (
group=in the URL). Each group has a header with its count that stays at the top while its people scroll by. Sector, persona and seniority follow the order of your lists; people with no value come last ("Not classified", "No company", "No event"). - Board: List and Board switch the view (
view=board). The board's columns follow the grouping, or relevance when there is none, and each card shows the name, the headline and the next step. On a phone the columns scroll sideways inside the list.
The list's grouping covers every match (the server sends each group's people together, and the header counts the whole group); the board shows the first 200 that match, plus any the list loaded before you switched.
Insights
/people/insights (Insights in People's bar) sums up your own people:
| Card | Shows |
|---|---|
| What People knows | For country, seniority, gender and sector: how many of your people have a value, and N unknown, a link to the list of exactly those people (/people?owner=me&country=unknown). Pick them all there and set the value with Set seniority, country, gender, sector |
| Pipeline | A funnel of your people: each counts once, at the furthest step they reached. Won: in a won deal. In a deal: an open deal past its pipeline's first stage. Talking: they wrote within 30 days. Reached out: you wrote last, no reply yet. Target: in an open deal at its pipeline's first stage and none of the above. Deals are the ones you can see. Each count opens exactly those people (/people?step=talking). Targets, deals and wins need a Funnel pipeline |
| Coverage | How many of your people are classified, and the last run with how it ended |
| Relevance, Sector, Persona, Seniority | People per value as bars; a bar opens the list filtered by it, and the list shows exactly that many. Sector counts the derived sector (the classification's, else the rules'), so it covers people not classified yet; the others count what the classification said |
| Top opportunities | The eight opportunities most of your people are tagged with |
| Follow up | High relevance, a next step, and no touch in 30 days, longest untouched first; "See all" opens the high-relevance list by last touch |
The bars are plain HTML and CSS: the page loads nothing from anywhere else. With classification off it says so, and owners and admins get a link to the settings.
On a person
A person you own shows an AI profile under the context card while classification is on: the headline, the relevance with its reason, sector, persona and seniority, opportunities, topics, the next step, the confidence and "Classified 2 days ago by <model>". A small dot marks a field you set by hand.
- Edit changes any field. What you save is an override: later runs keep it. Use the AI's answers clears your overrides so the model's answer shows again.
- Reclassify classifies this person now, as a one-person run that counts against the caps above. A person who has not been classified yet shows one quiet line, "Not classified yet", with Classify now, plus a link to the settings for owners and admins.
The card is hidden for people other members shared with you, and for everyone while classification is off.
For agents (MCP)
On the People MCP server (/workspace/mcp/people):
| Tool | Does |
|---|---|
people_classifier_get |
The settings, today's spend against both caps, how many are classified, a running run |
people_classifier_set |
Change the settings (owners and admins) |
people_classify_run |
dry_run: true for the estimate; otherwise start or continue (run) a run, or classify one of your people now (person) |
people_classifier_history |
The run history, paged with cursor and limit |
people_classification_set |
Set or clear (null) fields of a person's classification by hand |
people_correction_list |
Your corrections: what the model said, what you chose, whether it is still an example |
people_quality_get |
The override rate per field over 30 days, without and with examples |
people_proposal_list |
Suggested categories (status: pending by default, accepted, dismissed, all) |
people_taxonomy_propose |
dry_run: true for the estimate; otherwise one scan (owners and admins) |
people_proposal_accept |
Add a suggested value to its list (owners and admins) |
people_proposal_dismiss |
Dismiss a suggested value for good (owners and admins) |
people_autotag_apply |
dry_run: true previews the tags (other fields try settings); otherwise adds them (owners and admins) |
people_person_search |
Takes the filters and sort: relevance above |
The REST twins are GET/PUT /api/v1/people/classify/setting, GET/POST /api/v1/people/classify/runs,
PUT /api/v1/people/{person}/classification, GET /api/v1/people/classify/corrections,
GET /api/v1/people/classify/quality, GET/POST /api/v1/people/classify/proposals,
POST /api/v1/people/classify/proposals/{proposal}/accept and …/dismiss, and POST /api/v1/people/classify/tags.
Every model call, runs and scans alike, goes through the org's route with People's internal key, so it shows on
Requests with its cost: a run under session:"people-classify:<run id>", a scan under
session:"people-taxonomy:<scan id>". Both count against classification's daily cap and People's daily AI cap.
Operators
The scheduler is /api/cron/people-classify, every 10 minutes (CRON_SECRET): it starts scheduled runs and finishes
runs a browser left. Tables people_classifications and people_classify_runs (migration 0272);
people_classify_corrections, people_taxonomy_proposals and the runs' chosen examples (migration 0273).