Skip to content

Data & Intelligence

AI analytics: ask in plain words, get the governed number

People want to ask “how did Riverside do last month?” without learning a reporting tool. AI analytics lets them, but the model doesn't calculate anything: it turns the question into a query on your defined metrics, shows exactly what it used, and says plainly when a question is outside what the data can answer. We build it on your defined metrics and test it on the questions your team really asks.
Talk through your dataAsk it three questions ↓
operations managerHow did Riverside do in September compared with August?Completed jobs · Revenue (invoiced)RiversideSep vs AugSepAugCompleted jobs1218Revenue (invoiced)$38.9k$44.2kmetrics: completed_jobs, revenue_invoiced · as of Oct 1, 06:10Why are Riverside customers unhappy?No metric answers “why”.Available: review scores and complaints by location.

Questions shouldn't need a report request

Most questions about the business are simple and specific: how did a location do, which service grew, how many quotes are open. Today each one becomes a request to whoever runs reports, or a guess.

AI analytics answers them directly, from the same metric definitions your reports use. The model only works out what's being asked. The numbers come from the metric layer, and every answer shows what it used.

Today, and with AI analytics

Today

  1. 01Questions go to whoever builds reports, and wait.
  2. 02Or someone exports data and works it out themselves.
  3. 03A general chatbot will answer, with numbers it made up.
  4. 04Nobody can check how an answer was reached.

Simple questions are slow, and fast answers can't be trusted.

With AI analytics

  1. 01Ask in plain words, where you already work.
  2. 02The question becomes a query on your defined metrics, shown to you.
  3. 03The answer comes with its metrics, filters and freshness.
  4. 04Ambiguous questions get a clarifying question; unanswerable ones get a plain “can't”.

Fast answers you can check.

Three questions

The same assistant, three kinds of question.

operations manager

How did Riverside do in September compared with August?

Riverside completed 12 jobs in September, down from 18 in August. Revenue (invoiced) was $38.9k, down from $44.2k.

Sources

  • completed_jobs · Riverside · Sep vs Aug
  • revenue_invoiced · Riverside · Sep vs Aug
  • data as of Oct 1, 06:10

Every figure names the metric it came from, so it matches the report exactly.

The model reads. The metric layer counts.

Understand the questionThe modelLanguage is what it's good at
Choose metrics, filters and periodsThe model, from the defined list onlyIt can't invent a metric
Check accessThe platformAnswers stay inside what the asker may see
CalculateThe metric layerNumbers come from definitions, never from the model
Write the answerThe model, from the returned figuresPlain words around exact numbers
Show what was usedThe platformAnyone can check the answer

What it won't do

  1. Invent a number

    Metrics only

    Every figure comes from a defined metric. If there isn't one, there's no figure.

  2. Hide how it got there

    The query is shown

    Metrics, filters, period and freshness appear with every answer.

  3. Answer outside access

    Same rules as reports

    A technician asking about revenue gets what a technician may see.

  4. Guess at why

    Says so instead

    Causes need people. It shows what the data has and leaves the reason to them.

The parts it's built from

Each part is built and tested on its own, so you can start with one and add the rest.

  1. Understand

    LLM Application Development

    The question, in context, with the list of metrics.

  2. Map to metrics

    Business Intelligence

    • the question
    • defined metrics and dimensions
    • the asker's access

    no matching metric → say so, offer what exists

  3. Calculate

    Data Modeling

    The metric layer runs the query.

  4. Explain

    LLM Application Development

    Plain words around the returned figures.

Underneath every stepData Quality & ValidationThe metrics it answers from are checked every morning.

What changes from one business to the next

MetricsWhich defined metrics and dimensions it can use
Where it livesA chat workspace, the dashboard, or a weekly email that answers follow-up questions
AccessWho may ask about what
TestingA set of your real questions, checked on every change

A common starting point is a hosted language model with tool calling, your metric definitions in dbt or SQL views on PostgreSQL, and Langfuse for tracing and test questions. These are starting candidates, not requirements.

Is AI analytics the right system?

Good fit

  • Your metrics are defined and trusted already.
  • People ask simple questions often and wait for answers.
  • Someone will review the test questions when metrics change.

Avoid when

  • The numbers aren't agreed yet. It would answer quickly from disagreeing data.
  • You want it to explain causes on its own.

How we'd approach it

Most of the early work is defining the metrics, deciding where it lives and who may ask, and writing the questions it's tested on.

  1. 01

    Understand

    As much as the problem needs

    We map how the work runs today: who does what, which systems hold the records, and where it breaks.

  2. 02

    Design

    Decided before it's built

    Ownership, rules, integrations and exceptions, agreed with you and written down with acceptance criteria.

  3. 03

    Build

    Connected to what you use

    We build the scoped system, connect it to the tools you already run, and test it on the cases that matter.

Related systems

The capabilities underneath

Questions about AI analytics

Can the model get the numbers wrong?

It can pick the wrong metric, which is why the metric and filters are shown and tested. It can't miscalculate, because it doesn't calculate.

What if our metrics aren't defined yet?

Then that's the first step. Without definitions it would give fluent answers from disagreeing data.

Does our data go to the model provider?

The question and the returned figures do, under the provider's business terms. Raw records stay in your database.

Can it make charts?

Yes, from the same query, when a chart answers better than a sentence.

Let people ask the data in plain words

Tell us which questions people wait on an analyst for today. We'll show you which ones an assistant could answer, and how it shows its working.

Start a project