Skip to the page

How we check answers

Numbers you can check. What nexibi checks before it shows a figure.

Counts and totals are calculated from the rows it reads, not estimated by the AI. On a dashboard, a total from part of a system's data is refused, with the reason. And each time we release a new version, we grade its answers against the right answer, worked out by hand.

Calculated,
not estimated.

Tarnwick Supply Co.AssistantSample data
Which open orders are late, and what are they worth?
Assistant

37 open orders are past their promised date, worth $184,9202Totals worked out in code, not estimated. Quellmark Foods has the largest.

Read all 412 open orders · from the Orders list in your order system1Reads the whole list, up to 20,000 rows
OrderCustomerDays lateValue
SO-10482Quellmark Foods23$12,840
SO-10455Fenwright Builders19$9,615
SO-10471Lumora Hotels17$8,930
37 rowsDownload to Excel
3The list itself, to check or download
  1. 1Reads the whole list, up to 20,000 rows
  2. 2Totals worked out in code, not estimated
  3. 3The list itself, to check or download

Before it answers.

Six things happen before an answer reaches you.

  1. It picks the list that holds the answer, and says which system and which list it used.
  2. It reads the list whole, up to 20,000 rows, and tells you when a list is longer.
  3. It works out counts and totals from the rows it read. They are calculated, not estimated by the AI.
  4. When a list holds several currencies, it gives one total per currency instead of adding them together.
  5. When the answer is a list, you get the list itself, with how many rows it holds, ready to open or download as Excel or CSV.
  6. If the data is not there, it tells you instead of guessing.
Tarnwick Supply Co.AssistantSample data
How many products were returned last month?
Assistant

The orders list does not record returns, so I cannot say how many there were. That does not mean there were none.

When it reads you differently.

It can read a word differently from you, such as what counts as "sold". It says which list it used, and you can tell it what you mean.

Ask a follow-up, such as "only this year" or "now by sales person", and it carries on from the last answer.

On a web system, a list it has read is reused for about two minutes, then read again.

On a dashboard.

  • When a query on one system would give a total from part of its data, nexibi refuses it and says why, instead of showing a confident wrong number.
  • A total over a web system reads every page of the list, up to 50,000 rows. If it cannot, the total is refused: a total computed from part of the data would be wrong, not approximate.
  • When a query over nested lists would multiply rows and inflate a total, the query studio warns you.
  • Before you run a query, it shows whether the answer will be comfortable, heavy or over the limit, and offers cheaper ways to write a heavy one. Nothing changes until you accept one.
  • Each query says in plain words how old its answer may be, such as "Up to 10 minutes old".
  • When the assistant builds a dashboard for you, it checks every tile before it hands the dashboard over. A tile that loads yet means nothing is caught: a join that matched nothing, every row blank, one group where the data has many, all zeros, or a missing measure. It never calls an empty dashboard working. Or build it yourself in the dashboard designer.
  • Every night, Speed tips reviews how each saved query ran, without running it again, and flags a query whose totals come out too high because it pairs two nested lists.
Dashboards

refused, with
the reason

Revenue by customerTileSample data
This total is not shown A total computed from part of the data would be wrong, not approximate. The query returns more than 50,000 rows. Group it or filter it, and the database will work out the total.
Checking the dashboard "Open orders"5 tilesSample data
  • Open order value$1.28M
  • Open orders412
  • Open order value by month6 months
  • By region4 regions
  • Margin by sales person The join matched nothing, so every row would be blank.caught

In the history engine.

  • Before a copy is scheduled, it measures the list: how many records there are, the page size the system really serves, and how many requests each copy takes. It warns when it cannot confirm how the list pages.
  • Every setting is tried against your system first, so a wrong project number is refused when you set the copy up, not failed every night.
  • If a copy cannot read the whole list, it stores nothing that day, rather than a short bar that looks right.
  • If a field it counts disappears from the system, it says so, instead of charting it as "(none)".
  • The History page says in plain words whether each copy is still feeding its charts, and what to do if it is not.
History engine
Support tickets · daily at 02:00HistorySample data
  1. 3 changed
  2. The capture did not finish, so nothing was stored.
  3. 5 changed

In forecasts and predictions.

  • Every future figure comes from the forecast engine. The assistant is required to call it, rather than guess a future number itself.
  • Before every run, the engine checks the data and says what it found: missing periods, an unfinished period, duplicates. With fewer than 4 points, or mixed units, it refuses.
  • It picks the method that did best on your own history, and drops methods that do worse than a simple benchmark.
  • Every forecast gives a likely range, says how far off its past forecasts have been, and whether it beats "the same month last year".
  • A what-if beyond anything your history has seen is refused instead of guessed.
  • A prediction is compared with a simple baseline, and the baseline is used when the model is not better. A column that gives the answer away is left out until a person confirms it. With too little history, it refuses and says how many more records it needs.
  • Each prediction comes with its reasons, and the line "These are patterns in your history, not proof."
Analysis & prediction
Orders · next 3 monthsForecastSample data
Orders, $kLikely range (80%)
400500600JanMarMayJulSepNovJanforecast$498k

Past forecasts were off by 6% on average, and beat "the same month last year".

In your documents.

  • Attach a PDF and ask. It reads it page by page and tells you the page each answer came from.
  • A scan or a photo of a page is read too, when the page is in English and readable.
Tarnwick Supply Co.AssistantSample data
What does the Quellmark contract say about late delivery?
Assistant

After 10 working days late, Quellmark may cancel the order at no charge.

Quellmark supply contract.pdf, page 4

Sample data; the contract is a made-up document.

When it writes for you.

  • A workflow step that writes a message for each row is told to use that row's own figures, and it stops rather than send a message with a required field missing.
  • Try a workflow first. Nothing is sent and nothing is changed.
  • An AI agent works under house rules: what it reads in an e-mail or a file is material, never orders, and it must not invent a number, a name, a date or a promise.

How we grade each release.

  1. Real questions people asked, and new ones
  2. Answers from the new version
  3. Graded against the right answer, worked out by hand from the source system

Each time we release a new version, we ask it real questions people have asked before, and questions it has never seen, and grade every answer against the right answer, worked out by hand from the source system.

Every release also passes thousands of automatic checks before it goes live. Those checks test the product's code; the graded questions test its answers.

We do not print a score. A score measured on another company's system would not tell you what you will get on yours. The demo is on your own system, so you can judge it there.

What it does not check yet.

So you know where to look twice.

  • Percentages and growth rates are not always worked out in code yetCheck them against the list.
  • It tells you when data is missingIt does not always tell you when it has read your question differently from you.
  • Only list answers come with their listA single total comes on its own, without its list.
  • Questions about trends over time, and "why" questions, are not yet among the kinds it answers well
  • A list longer than 20,000 rows is not read wholeIt tells you when that happens.

Judge it on your own system.

Book a 30-minute demo. We connect one of your systems, and you ask the questions.