Product · Lynk Data Agents

Any agent can answer business questions. What happens when it’s wrong?

Read the whole investigation, not just the answer

Here’s how a Lynk data agent answers a question like the one below — working from your company’s brain, the written-down definitions and rules your data team owns. The plan it chooses, every query it runs, the finding it stands behind: all on the record, nothing to take on faith.

conversations/ndr-dip-mid-market
The ask

“Why did NDR dip in Q2 for mid-market accounts?”

a Head of Customer Success, in Lynk chat

It checks

“Your fiscal year starts February 1, so I’m reading Q2 as May–July. Say the word if you meant calendar Q2.”

The fiscal-year rule lives in the brain — stated, not assumed silently

The plan

Loads the churn-investigation skill from the brain: separate churn from contraction first, cut by segment and plan tier, check recent pricing changes.

The queries

Four checks, in parallel — every number compiled from named metrics in the brain, never improvised SQL.

ndr · by segment — contraction_revenue · by plan tier — logo_churn · by cohort — downgrades · pre/post price change  ✓ 4/4

The finding

“NDR held everywhere except mid-market: down 4 points — and it isn’t churn. Contraction on the legacy Pro plan doubled after April’s price change; 31 accounts downgraded rather than leave. Charts and method attached.”

Every conversation lands in this form — reviewable in Conversations, down to the SQL behind each number.

The best analyst, out of the box.
Answers your company can stand behind.

Lynk data agents are analyst and data-scientist agents in Lynk chat, the app where anyone in your company asks questions in plain language. You ask; Lynk brings the right agent to the question. Every answer is built on your company’s brain — your definitions, your rules, your warehouse — so business users get answers while the data team controls what they’re made of.

Working on day one

Point Lynk at your warehouse, bring up the brain, and first answers arrive in minutes. No agent to design, no pipeline to build.

Reasoning you can read

The agent shows its plan, states its assumptions, and logs every query it runs. When a number matters, you open the trace instead of redoing the analysis.

Your team controls what it knows

Every definition, skill, and policy the agent uses is a file in your Git repo. It chooses among your definitions — it never invents its own.

Three ways out of the queue — and why they didn’t hold

Each one trades the bottleneck for a different problem. None of them produces an answer you can check.

Chat-with-your-data tools

An LLM pointed at the warehouse gives confident answers with invisible logic. It’s wrong just often enough that every number needs re-checking — which is the queue again, with extra steps.

Building your own analyst agent

Orchestration, context, evaluation, monitoring, caching, versioning — months of infrastructure before the first trusted answer. And then your team maintains it.

More dashboards

Dashboards answer the questions someone predicted. The question that matters is the follow-up — “why did it move?” — and that one starts a ticket.

How the NDR answer was made

The trace you read above, step by step — the loop a Lynk data agent runs on every question.

  1. 01

    The ask

    A Head of Customer Success types a question into Lynk chat. Ambiguity gets handled first: the agent knows the fiscal calendar from the brain, states its reading of “Q2,” and moves on — or asks, when stating isn’t safe.

  2. 02

    It plans like your best analyst

    The agent loads the churn-investigation skill — an investigation runbook that lives in the brain as a file your team owns and edits: separate churn from contraction, cut by segment, check what changed recently.

  3. 03

    Governed queries, in parallel

    Each check compiles from named metrics — ndr, contraction_revenue — into deterministic SQL on your warehouse, under access control. Four queries at once; none of them improvised.

  4. 04

    Charts and an explanation

    The answer comes back the way an analyst would give it: the finding, the charts behind it, and the method — what was checked, and what was ruled out.

  5. 05

    The trace goes on record

    The full conversation lands in Conversations: question, plan, queries, answer. Review it, correct it, and fold what you learn back into the brain — the agent gets better exactly where it was wrong.

skills/churn-investigation/SKILL.md
# Churn investigation

When a retention metric moves, find *where*
and *what kind* before asking why.

1. Split the move: logo churn vs dollar
   contraction vs expansion. Different
   owners, different fixes.
2. Cut by segment and plan tier before
   anything else — blended NDR hides more
   than it shows.
3. Check pricing and packaging changes in
   the window. Compare cohorts before and
   after.
4. Report the split, not the average.
✓ skill applied · 4 queries from named metrics · full trace in Conversations

What’s computed, what’s judgment, what’s yours

An answer you can defend needs a boundary you can point to.

The numbers — deterministic

Every figure compiles from named definitions in the brain.

The same definition compiles to the same SQL, runs on your warehouse under access control, and is logged with the answer. No model ever makes up a metric.

The investigation — the agent’s

It plans the checks, picks the cuts, and writes the explanation.

Which skill applies, what to rule out first, how to present the finding — that’s the agent’s analyst judgment, exercised entirely inside your definitions and policies.

The knowledge — yours

The brain the agent works from is files in your Git repo.

Definitions, skills, and policies change by pull request — the brain is yours to open, edit, and veto. What you correct in review lands back in it.

Built to earn your data team’s sign-off

The questions a data leader asks before letting an agent talk to the business — answered straight.

01

Value in days, not quarters

Day one is connecting your warehouse and the brain: first questions answered in five to ten minutes, excellent results within one to two weeks — with a dedicated Lynk data engineer working alongside your team.

02

Production stays on the record

Every production conversation is logged in Conversations. When an answer misses, you can see exactly why — and the fix becomes part of the brain.

03

Root-cause today, dashboards next

The analyst agent answers and investigates day to day. The data-scientist agent takes the heavier work — root-cause, churn, and fraud investigations over large datasets — running today as skills in the brain. Dashboard building from a plain-language prompt is next on the roadmap: we’d rather tell you that than demo it.

04

Your warehouse, your rules

Runs on Snowflake and BigQuery with read-only, compiled queries and access control in the query path. The heavy lifting stays in your warehouse; your data is never copied out.