Agentic product analytics

Funnels, retention and journeys on the event data already in your cloud — the analytics your PMs know, without a second copy of your events. An AI analyst that investigates the why is in build.

See the product

Available today dashboards · funnels · retention · trends · journeys · cohorts · chart narration, on ClickHouse. AI analyst in build · other warehouses on request · status as of October 2026

01

The product analytics UI

An AI analyst is only as good as the tool underneath it. This is the full product-analytics seat — the one your team already knows how to drive.

Dashboards

The weekly review your team already runs — on your own warehouse.

Trends, funnels, retention and journeys on one board, reading your events where they land. Nothing was re-instrumented and nothing was copied out to build this.

Live todayA rendered illustration on a sample dataset of views in the current release.

Illustration: a weekly checkout-health dashboard with daily orders over 30 days, a five-step checkout funnel converting at 7.4%, a weekly retention grid with W4 at 31%, and a journey diagram from app open.

The tool

Everything your PMs use today — on your data.

A replacement for the product-analytics seat: funnels, retention, metrics, journeys and dynamic segments today; alerts and the Slack daily review arrive with Watch, which is next. Fluent on day one for a team that knows Mixpanel or Amplitude. Not session replay, not feature flags.

Funnel: product viewed to order placed over 30 days on sample data — 43,140 users enter, 3,192 convert (7.4%); the largest drop is the first step, at 58%.

Product purchase flow
Last 30 days · Conversion over steps
Conversion
7.4%
product_viewed → order_placed
Converted
3,192
of 43,140
100%43,140
−58%
42.0%18,119
−45%
23.1%9,965
−36%
14.8%6,385
−50%
7.4%3,192
product_viewedadded_to_cartcheckout_startedpayment_startedorder_placed
02

The AI analyst

Every product team has more questions than analyst hours. This is what we are building to answer them — shipping in stages, with its working shown every time.

The target experience

One question, asked the way a PM would ask it.

No SQL, no ticket, no waiting for someone to have a free afternoon. The analyst finds the drop, isolates the segment, ranks the likely causes — and shows you the chart it read.

In buildAn illustration of what we are building, on a sample dataset — not a recording of the current product.

Illustration: asked why activation fell, the analyst reports a drop from 31.4% to 24.5% (6.9 points, 22% relative), isolated to iOS at the OTP step, names the iOS 4.12 release as the leading hypothesis with the checks still to run, and attaches a chart of OTP verification by platform.

1Detects the drop2Finds the segment and the step3Ranks likely causes as hypotheses — with the chart it came from
Who it is for, and what it does

Answers for everyone. Analysts for the hard problems.

Most questions a product team asks are simple — and still wait a week for an analyst. The AI analyst is being built to answer those in seconds, with receipts. Your analysts get their time back for the work that deserves it.

Ask In buildanswer · build · explain
Watch Nextmonitor · daily review
Dig Laterroot cause · test a hypothesis
WhereIn the productSlack PlannedMCP — Claude, Cursor, ChatGPT Planned
Live todaydashboards · funnels · retention · trends · journeys · cohorts · chart narration

Shipping in this order: catalog → Ask → MCP → Watch → Dig. Status as of October 2026; we'll tell you on the call exactly what's live the week you'd go in.

Everyone can askPMs, growth, founders. Plain language, no SQL, no ticket.
The queue shrinksBuilt to answer the trivial and the mid-complex in seconds, not a sprint.
Analysts move upOff the number-pulling, onto the strategy they were hired for.

Illustration of three example exchanges on sample data: a question asked in the product, a daily review posted to Slack (planned), and a retention question asked over MCP (planned).

How you check the number

Resolved through the catalogAnswers are designed to resolve through one catalog of your metrics, cohorts and funnels — each tested against your real data and approved by a person, with the version used shown.
Receipts on every answerThe chart it came from, the window, the segment. Check it in a glance; adjust it if the question wasn't quite what you meant.
It proposes, you approveReading and analysing run unattended. Saving a definition, or sending anything to other people, waits for you.
The catalog agent In build

Where do the definitions come from? Your data — checked by a person.

Point a language model at raw tables and it guesses: at the column that sounds right, at what “active” means. We constrain it instead. Answers resolve through one catalog of your metrics, dimensions, cohorts and context — drafted from your own warehouse, then signed off by someone on your side.

01Indexer

The warehouse indexer reads what you have

Schemas, tables, columns, volumes, relationships and how they are actually populated — profiled where they already sit, read-only. It learns that plan_tier holds four real values, not that it sounds like a plan.

events1.4B rows · 212 types
users38 columns profiled
ordersjoins on user_id
02Agent

The catalog agent drafts your semantic layer

From that index plus your own docs, it proposes the metrics your team argues about, the dimensions each can legitimately be split by, the cohorts that matter, and the business context behind them. Every definition is run against real data before it is offered to you.

Activation ratesignup → first_order ≤ 24h
Checkout conversion5-step funnel
W4 retentionweekly cohorts
03Approval

A person on your side approves it

Your analyst edits the wording, fixes the edge case, rejects what is wrong. Nothing enters the catalog without that sign-off, and each approval is versioned. Definitions stay readable text you can diff and revert.

12 definitions approvedby your analyst
2 sent backwording edited
1 rejectedwrong join

Why this narrows what the AI can get wrong

Less is inferred at question timeWhat a metric is, how it joins, what may be split by what — decided once, in the catalog, and approved by a person. The question still has to be interpreted, but it picks from checked definitions, not raw columns.
Every answer ships its receiptThe chart, the window, the segment, the definition and version it resolved to, and how fresh the data is. No number should need taking on faith; checking it is meant to be a glance, not an investigation.
It is allowed to say noA question the catalog cannot answer comes back as exactly that, with what is missing. An analyst who says “I don't know” is worth more than one who is confidently wrong.

A catalog constrains what gets computed; it can't prove a cause. Explanations come back as ranked hypotheses, with the next check to run.

When it can't“I can't answer that yet — refund rate isn't in your catalog. I can draft it from refund_requested and order_placed for your analyst to approve.”
03

How we deliver it

Privacy-first by construction, not by policy. The architecture that keeps your events in your cloud is the same one that takes the vendor's copy off the bill.

Privacy-first architecture

Why does your data have to leave to be analysed? It doesn't.

Today
Your cloud
Your warehouseevents land here
a second copy→
Vendor's serversAmplitude, Mixpanel$ billed by volume

A copy of every event leaves your cloud. You pay to store it twice — so you ration what you track.

→
With DepSmith
Your cloud / VPC
Your warehouseevents stay here
←reads in place
DepSmithdeployed here

The analytics moves to the data. Raw events are not copied out; what the AI layer sends to a model is listed on the Security page.

ClickHouse · liveSnowflake · on requestDatabricks · on requestBigQuery · on requestRedshift · on requestS3 · on requestPostgres · on request
Live = supported in the current release. On request = scoped per customer and may need custom work. S3 needs a query engine in front of it. Connector details
→
How it benefits you
0copies of your raw events outside your cloud
50–70%expected saving against incumbent list price, for the analytics replaced — see the method

If your events don't land in a warehouse yet, there is a collection path — we'll tell you which one you're on before the call ends.  ·  Who runs what in your account — and a hosted option, still being defined — is on the Deployment page →

Security & data residency

What can we see? As little as the architecture allows.

Deployed in your cloudIn the customer-cloud deployment, DepSmith runs in your account, under your IAM. There is no DepSmith-side environment holding your events.
No copy of your event dataRead-only against the store you already run. Raw events are not replicated to us — that removes one class of exposure, not every one.
Your credentials, your perimeterWarehouse credentials stay inside your account. How DepSmith users map to warehouse roles is set out on the Security page.
The AI model boundaryRaw tables are not sent to a model. An AI answer does need the model to see the question, the relevant definitions and aggregated results. The provider, its region and bring-your-own-key terms are being finalised with design partners — you see the exact flow before you connect anything.

No compliance badges until we hold them, and we display none today. Keeping raw events in your account simplifies residency questions under DPDP; it is not, by itself, compliance. Full Security page →

Pricing

Where does the saving actually come from?

Today
Vendor bill — grows with events held
With DepSmith
Vendor billYour warehouse — same either wayDepSmith softwareAdded query & model cost
The warehouse baseline is the same either way. What changes: the vendor bill goes, our software is added, and so is the extra query and model work the analytics runs in your account. Illustrative proportions, not a quote.
How we charge
  • Not per event. Track everything; the bill does not decide what you measure.
  • No charge for retention. We don't store your data, so we don't bill you for holding it.
  • You pay us for the software. Storage and the extra compute our queries use are billed by your cloud provider, not by us — and they count in the comparison.

Design-partner pricing on the call. The pricing unit — per workspace, per seat or a flat tier — is not fixed yet; we'll tell you which applies before you commit.

50–70% is the expected range against incumbent list pricing for the analytics replaced, not a measured result. How we calculate it, and what it leaves out →

How the options compare

How is this different from what you use today?

Mixpanel / AmplitudePostHog self-hosted (open-source edition)Build it in-houseDepSmith
Where your event data livesTheir servers, in a region you pickYour infrastructureYoursYour cloud account, where it already lands
Who operates the infrastructureThe vendorYou — ClickHouse, Kafka, the restYouComponents run in your account; the operating split is set per deployment — see Deployment
What you are billed forEvent volume (Mixpanel) or monthly tracked users (Amplitude)Your infrastructure, and the team to run itA platform teamThe software; your cloud bills the compute
Runs in your own cloud account✗ regional hosting (e.g. EU, India) is offered✓✓✓
AI analystYes, over their hosted copyOffered on PostHog Cloud; not established for self-hostedBuild itIn build — in the product first; Slack and MCP planned

On building it yourself. It works — teams have done it. The price is a dedicated platform team, the better part of a year, and a standing capability gap versus the tool their PMs had before. That team is the cost.

Compared in October 2026 from each vendor's public documentation: Mixpanel pricing, Mixpanel data residency, Amplitude MTU billing, Amplitude AI data handling, PostHog self-hosting, PostHog AI. Plans change — check the source.

How to get started

What happens after you say yes?

1

Connect

Read-only, to where your events already land. Inside your cloud.

prod_eventsread-only ✓
2

The catalog is proposed In build

The analyst reads your events and your docs and drafts your metrics, cohorts and funnels — each tested against real data.

Activation ratetested ✓
Checkout conversiontested ✓
W4 retentiontested ✓
3

You approve

A person on your side signs off each definition. Nothing enters the catalog without this step.

12 definitions approvedby your analyst
4

Ask

Dashboards on day one; questions in the product, Slack or MCP as each ships.

Why was activation down yesterday?
−22% · iOS · OTP step · since 14:20

Typical: days, not a quarter, once your events land in a supported warehouse and access is approved — the security review on your side comes first and is often the longest step. History in that warehouse is queryable from the first day. Coming from Mixpanel or Amplitude? The migration path, and how to leave →

Start with 30 minutes. A walkthrough on sample data and an honest read on fit. Your own data comes in a working session, once access is approved.
Who this is for

Is this a fit for you yet?

A good fit if

  • You are a consumer product at real scale.
  • Your event data already lands somewhere you control.
  • Your analytics bill is a line item someone asks about.
  • Your team already knows what they want out of a product analytics tool.
  • Keeping events in your own cloud would help with residency or security reviews — welcome, not required.

Not yet a fit if

  • Your events live in a store we don't support yet, and you can't wait for it.
  • You are early enough that an incumbent's free tier is genuinely fine.
  • Analytics is an insignificant line item for you, and budget is not a worry.

We will tell you so on the call.

Questions we get on every call

The practical questions

Does our data leave our cloud?

Raw events don't. DepSmith is deployed inside your cloud account and reads your events where they already land; we never take a copy. Some derived data does travel: an AI answer needs the model to see the question, the relevant definitions and aggregated results, and Slack or MCP — when you turn them on — receive answers. The model provider and its terms are being finalised; the Security page lists every flow and what is still open.

We don't have a warehouse yet — can we still use it?

Yes, with one more step. If your events don't already land somewhere you control, there is a collection path that gets them there first. We'll tell you which path you're on before the first call ends.

Does our historical data work on day one?

Yes, when your events already land in a warehouse you control — we read what you hold, so there is no backfill and no gap. If your history lives only inside Mixpanel or Amplitude, there is an import step; we'll confirm the path for your setup on the call.

Is this a replacement for Mixpanel or Amplitude?

For the product analytics your PMs use daily — funnels, retention, trends, journeys, cohorts, dashboards — that is what it is built to replace. Familiar charts don't guarantee identical numbers: identity resolution, time zones and funnel windows differ between tools, so a pilot reconciles a sample of your existing reports first. It is not session replay or feature flags.

Who builds the metric catalog?

The catalog agent — in build — proposes it from your data and your docs: metrics, dimensions, cohorts, funnels, each tested against real data. A person on your side approves, and approvals are versioned. Nothing enters the catalog without that step.

How is this different from self-hosting an open-source tool?

Self-hosting solves residency and hands you a stack to operate — ClickHouse, Kafka, the rest — plus their SDK to adopt. Here there is no new event pipeline to adopt and the data doesn't move; who operates the DepSmith components in your account is set per deployment — see Deployment.

How long to go live?

Days, for the common case where your events already land in a supported warehouse and your security review is done. The steps: connect read-only, propose definitions, your analyst approves, first dashboards. How long each takes depends on your access process, so after the first call we give you a dated plan rather than a number here.

What if we want to leave?

Your raw events never moved, so there is nothing to retrieve. Catalog definitions are readable text you can export. Saved reports, cohort logic, permissions and configuration are application state — what exports today, and in what format, is on the Migration & exit page; not all of it does yet. Then you remove the deployment.

See what runs today.

Thirty minutes: a walkthrough on sample data, a look at where your events live, and an honest read on whether this is a fit yet. If it is, the next step is a working session on your own data, once access is approved.

Not ready to talk? Leave an email and we will send you the occasional real update — no drip sequence. How we use it.