The product layer
for AI agents.

Locus reads every conversation your agent has had, groups users by what they actually do, and shows the picture in one view. The first memo is free.

locus.app · your-agent / overview
OverviewGroupsUsersConversationsUse casesDriftMemosSettings
every user · last 7 days
240 users · 1,920 conversations
Code-first32%Writer21%Researcher17%Analyst13%Generalist10%Advice-seeker7%

Groups

See who your users are by what they actually do. Cohorts that match behaviour, not plan tier.

Intents

Read every conversation. Auto-classify into intents and use cases. Find what they're really asking for.

Drift

Track value, not just volume. Get told when your agent is losing trust, before retention reflects it.

User Visibility

See your users by what they actually do.

Locus groups every user by behaviour, surfaces their cohorts, and shows you which group is growing or shrinking week over week.

GroupsCohortsPower usersTrust scoreDriftFilters
/groups
Groups · 240 users
sorted by size
  • Code-first
    Engineers. Most of their sessions are code. A learning tail and a small data tail are common.
    77
  • Writer
    Drafting and editing prose. Often pairs with creative brainstorming.
    50
  • Researcher
    Asks questions about topics. Reads and summarizes. High question-rate, low tool use.
    41
  • Analyst
    Heavy data + code. Spreadsheets, SQL, numeric work.
    31
  • Generalist
    Broad. No single use case dominates. Often the curious first-month user.
    24
  • Advice-seeker
    Decisions, relationships, career moves. Talks out loud, asks for perspective.
    17
Intent Detection

Read every conversation. Find what users are actually asking for.

Locus auto-classifies every conversation into intents and use cases. You see what they're trying to do, not just what they did.

Intent mapUse casesSentimentResolutionTool callsSearch
/conversations
Intent breakdown
Code & programming22%
Writing & editing20%
Research & summarization14%
Learning & explanation12%
Data & analysis10%
Troubleshooting & advice7%
Latest conversations
  • u-0042 · Researcher
    Summarize the methodology section of this paper for me
    Research12s ago
  • u-0118 · Code-first
    Why is this React component re-rendering twice on mount
    Code1m ago
  • u-0073 · Writer
    Edit this paragraph for clarity, keep the tone the same
    Writing3m ago
  • u-0201 · Analyst
    Convert this CSV into a SQL insert statement, infer types
    Data4m ago
Drift Tracking

Track value, not just volume.

Locus watches whether your agent is creating user value and flags drift before it shows up in retention.

DriftRetentionAcceptanceShadow reworkCohort tracking
/drift
Trust score · 12 weeks
14↘ −36% over 12w
Wk 1 · 22%Wk 6 · 30%Wk 12 · 14%
Drift detected: Code-first users dropping in trust 4 weeks before retention reflects it.
What you walk away with

One memo. Three to six pages. Written for product.

Six sections, every one something you can take into a roadmap meeting tomorrow. No traces, no token counts, no dashboards.

Cover · table of contents
v1
Prepared for

The product team at [your company]

  • 01Intent map§
  • 02Completion vs. value§
  • 03Signal-from-noise§
  • 04Trust and shadow rework§
  • 05Value drift§
  • 06What to instrument next§
Excerpt · §04 trust and shadow rework
redacted
Finding 04.2

Acceptance is not a value signal in the workflow.

31% accepted-then-edited12% sent · then ticket57% genuine value

Of users who accepted the agent's draft on a first attempt, a third re-edited more than half the output before sending it. Acceptance was a politeness signal — not a value signal.

Sample findings

What recent memos found.

Each one is a real excerpt from a snapshot memo, names redacted. The shape of the pattern is what carries over to your team.

44%

44% of completed runs ended with the user editing more than half the output before sending it.

§02 Completion vs. valueredacted
31% rework

Acceptance was a politeness signal, not a value signal. Users accepted, then redid the work elsewhere.

§04 Trustredacted
−36%

Code-first users dropped 36% in trust over 12 weeks, four weeks before retention reflected it.

§05 Value driftredacted
5x growth

An emerging ‘edit my email’ cluster grew 5x in three weeks. The team had no name for it yet.

§01 Intent mapredacted
inverse

High-usage users were not happy users. The strongest engagement scores tied to the highest rephrase rates.

§03 Signal-from-noiseredacted
3 to add

Three product events would have caught the trust drop a month earlier. None of them existed in the codebase.

§06 Instrument nextredacted
On-prem

Your data, on your side of the wall.

For most teams a sanitized sample is enough. For teams under stricter rules, the entire pipeline runs inside your VPC. Nothing ever leaves your network.

SOC 2 Type IIGDPRDPA on requestNo model trainingBYO key managementBYO audit logging
/onprem
Locus pipeline imageonce · signed
Your VPC
no egress
  • 01
    Sample lands
    Sanitized runs land in an object store inside your VPC.
  • 02
    Pipeline runs
    Locus containers run the read in your environment. Nothing leaves.
  • 03
    Memo emits
    PDF + markdown drop into a folder your team controls.
after the runMemo · PDF + MD
Common questions

What we get asked most.

What if my agent isn't fully in production yet?

If you're past internal testing and have at least a few hundred real-user runs over the last month, the snapshot will work. Below that, we'll tell you on the call — there's not enough behaviour to read, and a memo would be guessing.

What if we can't share any data at all?

Two paths. We can do a structured discovery audit using only synthetic examples and your team's notes — narrower findings but still a usable read. Or we hand you a container image and you run the pipeline inside your VPC. Nothing leaves your network in either case.

How big a sample is enough?

100 to 500 sanitized runs is the sweet spot. Less and the patterns are noisy. More and we don't read the extra carefully — we cap at 500 to keep the memo specific.

What happens after the free first pass?

If the snapshot is useful, we offer a four-week paid pilot — bigger sample, two follow-up memos, drift tracking, and a final read-out. Pricing depends on volume. We only suggest it if the first memo earned the next conversation.

Why not just hire a consultant or do this in-house?

You can. Most teams that try end up reading twenty conversations a week and calling that research. The snapshot exists because reading at the volume that produces real patterns takes a system, not a person.

Thirty minutes · one memo · free first pass

See what every user of your agent does.

Pick a time. We'll walk through what a snapshot would look like for your product, on your terms.