AI data analysis without waiting for an analyst
Your numbers sit in the CRM, the accounting system, the shop, a few spreadsheets and Analytics, and answering one question from the board takes two hours of exports. We pull them into one database, clean them and connect a dashboard that refreshes itself. On top of that sits an AI agent that answers a question about numbers in 5 to 10 seconds, without opening a single system.
- 10 to 20places where a B2B company keeps data, the count we usually find in the workshop: CRM, accounting, the shop, Analytics, spreadsheets, the shared inbox
- 5 to 10 secondshow long the AI agent takes to answer a question about numbers, instead of two hours of exports and spreadsheet stitching
- 2 to 4 weeksfrom workshop to database, pipeline and the first dashboard on your data, refreshed automatically
The problem
Why does a simple question about numbers take two hours to answer?
The CEO asks how sales are tracking this quarter against the last one. The analyst opens the CRM, exports to a spreadsheet, logs into the accounting system for invoices, adds campaign leads from Analytics and cost from the ad platform. Two hours later the report is ready, but the quarter closed a fortnight ago and the decisions went ahead without it.
In the workshop we usually find 10 to 20 places where a company keeps data: CRM, accounting, the shop, Analytics, ad platforms, the task system, the shared inbox, and a spreadsheet that quietly works as a second CRM. Each one answers a different question and none of them sees the whole. To answer a single business question, somebody has to open several panels, export, join by hand and recalculate.
The data that finally gets collected is rarely clean. In a study published in Harvard Business Review, where 75 managers scored the data their own departments produced, only 3 percent of the collections cleared a basic quality bar, and on average 47 percent of newly created records carried at least one error that affected the work. A dashboard built on data like that misleads faster than a spreadsheet, because it looks authoritative.
- The board decides on a report that describes last month
- The analyst spends two hours on exports instead of on analysis
- Sales and finance walk into the same meeting with two different revenue figures
- A drop in conversion surfaces at the quarterly review, not in the week it started
Why now
What changed, so a data warehouse is no longer a quarter-long project?
BigQuery is Google's data warehouse with no server to keep alive. It scales on its own and handles petabytes, and you pay for the data a query processes rather than for a machine standing ready, so for most B2B companies the monthly bill stays small. A few years ago the same result meant buying hardware and hiring someone to watch it.
The second change is interfaces. CRM, accounting, BaseLinker, Shopify, Analytics and Search Console all expose an API, so data can be pulled on a schedule or in reaction to an event. A pipeline in Make, n8n or a custom integration does that without a person, which removes the step where somebody has to export.
The third change is the AI agent. A dashboard answers the questions someone anticipated while building it. Why margin dropped in July, and only in that one segment, was not on that list. BOND, our company AI agent, writes the query against the database itself and answers in a sentence in 5 to 10 seconds, together with the table it counted from.
What changes
What do the board and the team actually get?
One revenue figure for the whole company
Data from the CRM, accounting and the shop lands in one database and goes through the same transformation, so sales and finance count from the same source. Currencies, time zones, field names and metric definitions are flattened to one format.
Metric definitions written down in a document we point at whenever a number is disputed
A dashboard that loads itself
Looker Studio, Google Sheets or a dedicated panel wired to the database. In the morning you see today, not the last export somebody ran on Friday.
Schedule from once a day to every 15 minutes, depending on how fast the data moves
Ask about numbers in a plain sentence
What is our lead to contract conversion this quarter. Which client grew most year on year. What did we pay per order in August. The AI agent writes the query against the database itself and answers in 5 to 10 seconds.
The answer names the table and the period it counted from, so it can be checked
Anomalies report themselves
We monitor the metrics chosen in the workshop and send an alert when conversion falls, cost per acquisition rises or delivery slips. The alert lands where the team is already looking.
Notification in Slack, Microsoft Teams or email on the day it happens, not at the monthly review
The analyst goes back to analysis
Exports, spreadsheet stitching and copying columns across are handled by the pipeline. The person keeps the questions that need thinking rather than clicking.
Two hours for a quarterly report drop to one dashboard refresh
The history of your data stays with you
The database sits in your own Google Cloud project or on your PostgreSQL and grows from the day of rollout. When a source system shortens how far back you can look, the warehouse still holds the full record.
You own the cloud project, the bill goes to Google directly, with no margin from us
Proof
Do you work on data this way yourselves?
Yes. The AI agent inside our founders' agency has been loading the warehouse and answering from it since March 2026. Below is the scale of that work, together with what the number does NOT prove.
Built for ourselves. These are not client deployments but implementations inside the founders' own companies. We say so plainly, because proof you cannot verify is not proof.
JustBottheir agentJustIdea Agency, a performance marketing agency of around 30 people
- 23
- people from the company work with the agent
- 38
- systems wired in permanently
- 112
- processes run without a human
- 475
- thousand lines of the agent’s code
The agent is wired into a BigQuery data warehouse and into the ad platforms, the CRM, the task system, the SEO tool, Analytics and Search Console. First commit 1 March 2026, today 23 people use it.
8 of 38 systems it works in
Google Ads
Meta Ads- HubSpot
- ActiveCollab
- Ahrefs
GA4
Search Console
BigQuery
- Before
- Every question about numbers started with an export from several panels and a stitched spreadsheet. Ad data and sales data did not sit in one place, so a year on year comparison was assembled by hand.
- After
- The warehouse loads itself from the system APIs, and questions about data go to the agent as a sentence. A person writes a sentence rather than SQL and gets the answer together with its source.
3,506 operations against the data warehouse in 30 days, run without a person
And in your company
What if our data is a mess?
Four sentences we hear most often, and what we say to them.
Our data is in such a state that there is nothing to analyse.
That is a normal starting point and it changes the order of the work, not the plan. The workshop opens with a review of sources: what is in the CRM, what is in a spreadsheet, and what exists only in a salesperson's head. Cleaning and standardising is a separate step before the dashboard, because a panel on dirty data misleads faster than a spreadsheet.
We have Looker Studio and it did not help.
Because a visualisation tool does not fix the absence of a shared database. If every report pulls from a different export, every report shows a different number. The order is the other way round: one database and one metric definition first, the panel second. At that point your current tool is usually enough and we do not ask you to replace it.
We have no analyst and nobody who owns data.
You do not need one to use this. The dashboard refreshes itself and questions go to the agent as a plain sentence. The pipeline is maintained by us or by your IT provider once we hand over the documentation. For the workshop, one person who knows where the data sits and who uses it is enough.
We will not put financial and customer data into the cloud.
The database sits in your Google Cloud project or on your PostgreSQL, not with us. Which fields leave the source system at all, who has access to them and how long we keep them is agreed in the workshop. Personal data is processed under the <a href="https://eur-lex.europa.eu/eli/reg/2016/679/oj" rel="noopener">GDPR, Regulation (EU) 2016/679</a>, and where you sell outside the EU we add the rules that apply on those markets.
- Warehouse and databases
- BigQuery, PostgreSQL, SQL databases
- CRM
- HubSpot, Pipedrive, Salesforce, LiveSpace
- Accounting and invoicing
- Xero, sevDesk, Fakturownia, Comarch Optima
- Online sales
- Shopify, WooCommerce, PrestaShop, BaseLinker, Allegro
- Analytics and dashboards
- GA4, Search Console, Looker Studio, Google Sheets, a dedicated panel
- Automation layer
- Make, n8n, Zapier, custom API integrations
Where the data comes from, how often, and which question it answers
The sources that most often enter the warehouse. We set the load frequency per source, because order data goes stale in hours and a budget plan in months.
| Data source | How often we load it | Which question it answers |
|---|---|---|
| CRM: HubSpot, Pipedrive, Salesforce, LiveSpace | hourly or on an event | how many open opportunities, what the lead to contract conversion is |
| Accounting: Xero, sevDesk, Fakturownia, Comarch Optima | once a day | what was actually invoiced, who has not paid on time |
| Online sales: Shopify, WooCommerce, PrestaShop, BaseLinker, Allegro | from every 15 minutes to once a day | sales per product and channel, returns, margin after discounts |
| Ad platforms and GA4 | once a day | cost per lead, which channel the order came from |
| Search Console | once a day | which queries bring traffic, which page lost position |
| Task system: ClickUp, Jira, Asana | once a day | how many hours went into a client, where a project is slipping |
| Team spreadsheets in Google Sheets | on every change | targets, budgets and the data that exists in no system |
These rows are a starting point. A first rollout usually covers two to four sources, the ones that answer the questions asked most often. The rest go in when they turn out to be needed.
How we do it
Five steps from exports to a database that loads itself
Every step ends with something you can judge before you decide on the next one.
Workshop
We write down the questions that come up most often in the company and check where the data behind the answers actually sits. The output is a list of sources, including the spreadsheets that quietly act as systems. We agree metric definitions, because revenue in sales and revenue in accounting are usually two different numbers.
Database and pipeline
We set up the warehouse in your project: BigQuery or PostgreSQL. The pipeline pulls data from the system APIs on a schedule set per source. This is the stage where gaps and errors in the source data surface, so we split them straight away into what gets fixed at the source and what gets cleaned in transformation.
Transformation and the first dashboard
We bring fields, currencies, time zones and names to a common format. The first dashboard shows the metrics from the workshop question list and runs alongside your existing reports, so the numbers can be compared before anyone switches off the old spreadsheet.
AI layer and alerts
We connect the AI agent to the database together with a description of the data, so questions go in as sentences rather than SQL. We set alert thresholds: conversion falling, cost per acquisition rising, delivery slipping. We test the answers on questions where we already know the result from existing reports.
Measurement after 30 and 90 days
We check which views the team actually uses, which ones nobody opens, and whether the alerts fire too often. A dashboard nobody returns to looks identical in a report to a successful one, which is why we look at usage rather than at the number of charts.
Before and after
How the Monday results meeting changes
Before
- A question about numbers triggers two hours of exports at the analyst's desk
- Sales and finance quote two different revenue figures and the argument eats half the meeting
- The report describes last month, because that is how long it takes to assemble
- A drop in conversion comes out once the quarter is already closed
After
- A question about numbers comes back in 5 to 10 seconds, as a sentence and with its source
- One database and one written metric definition, so there is one number
- The dashboard shows today, because it loads itself from the system APIs
- The alert about falling conversion arrives on the day the drop started
The first dashboard on your data stands 2 to 4 weeks after the workshop. The full AI agent layer usually takes around 2 months, and we judge usage and the sense of the alert thresholds after 30 and 90 days.
For the person signing this off
What the CEO or the CFO gets out of it
You stop opening the results conversation by asking whose number is the real one. Metric definitions are written down, data comes from one place, and when somebody quotes a different figure it is immediately clear which source it came from. That shortens the board meeting and takes you out of the role of referee.
The second thing matters more on big decisions. You know what data you are deciding on and how old it is, because the dashboard shows the time of the last refresh. In a conversation with an investor or a bank that is the difference between a number you can defend and a number from a spreadsheet whose author is on holiday.
- The database sits in your own cloud project, the bill goes to Google directly, with no margin from us
- Metric definitions are written down, so the number does not depend on who calculated it
- The new dashboard runs alongside the old reports until the numbers agree
- The workshop ends with a document you can execute without us
What it costs
The workshop has a fixed price, the rollout is quoted after it
The process workshop costs EUR 3,500 net per week. With data it usually runs one to two weeks, because several systems have to be looked at and metric definitions agreed across departments. We quote the rollout after the workshop, once we know how many sources are in play and what state the data is in.
The warehouse bill goes to Google directly. You pay for the data a query processes, not for keeping a server alive, and nothing of ours sits on top of that invoice. The AI model cost is a vendor cost as well, in the range of EUR 0.005 to 0.07 per query depending on the model, and it is billed to you by the provider.
Process workshop
EUR 3,500 net
per week, with data usually one to two weeks
A self-contained piece of work. It ends with a list of questions, a map of data sources, metric definitions and a plan you can execute without us.
- The list of questions the company has to be able to answer, and how often each is asked
- Map of data sources, including the spreadsheets acting as quiet systems
- Metric definitions agreed between sales, finance and marketing
- Recommendation: BigQuery, PostgreSQL, or staying with a tidy spreadsheet
- The indicators to measure after 30 and 90 days
- Most popular
Workshop and warehouse rollout
Quoted after the workshop
scope follows from the number of sources and the state of the data
The option most companies choose. Database, pipeline and the first dashboard on your data 2 to 4 weeks after the workshop.
- Everything in the workshop
- Warehouse in your own project: BigQuery or PostgreSQL
- Pipeline from the system APIs in Make, n8n or a custom integration
- Transformations and a dashboard in Looker Studio, a spreadsheet or a dedicated panel
- Anomaly alerts to Slack, Microsoft Teams or email
- Documentation and team training on how to use it
Warehouse and an AI agent on your data
Quoted after the workshop
the agent layer usually takes around 2 months
Everything in the rollout plus BOND, the company AI agent that answers questions about your company data in 5 to 10 seconds.
- Everything in the warehouse rollout
- AI agent with access to the database, questions in English and in Polish
- Answers name the table and the period, so they can be checked
- Access for the board and the team, with no login to the source systems
- Model cost paid to the vendor directly, with no margin from us
The workshop ends with a document: the list of questions, the map of sources and the metric definitions. You can take it to another supplier or run the rollout yourself. And if it turns out that with two sources and one question a week a tidy spreadsheet with automatic loading is enough, we say so and we do not sell you a warehouse.
Frequently asked questions
What is AI data analysis?
It is a service where we first collect data from your systems into one database (BigQuery or PostgreSQL), clean and standardise it, and then put dashboards and an AI agent on top. The agent answers questions about numbers as a plain sentence in 5 to 10 seconds, because it writes the query against the database itself. An AI layer with no ordered data underneath has nothing to answer from, which is why the order is this way round and not the other.
How long does a data analysis rollout take?
A week or two of workshop, then 2 to 4 weeks for the database, the pipeline and the first dashboard on your data. The full AI agent layer usually takes around 2 months, because it adds a description of the data, access rules and testing answers against real questions. You see the first working piece within 14 days of the workshop.
Why BigQuery and not a database on our own server?
BigQuery needs no server to keep alive, scales on its own and handles petabytes of data, and you pay for the data a query processes rather than for machine uptime. Where the data has to stay on your infrastructure we set up PostgreSQL and the pipeline works the same way. We give the recommendation in the workshop, after looking at the volume and the sensitivity of the data.
Do we have to replace our dashboards?
No. Looker Studio and Google Sheets, the ones you already use, get connected to the new database, so the panels stay and what changes is where they take data from. We build a dedicated panel only when the existing tool cannot show something. The new source runs alongside the old reports until the numbers agree.
Where does the data sit and who has access to it?
The database sits in your Google Cloud project or on your PostgreSQL, so you own the data, not us. Access is granted per person and per dataset, and the AI agent answers only from what the person asking is entitled to see. The scope of fields leaving the source systems is agreed in the workshop, and personal data is processed under the GDPR.
How do we know the AI agent will not invent a number?
The agent does not guess. It writes a query against the database and returns the result together with the table and the period it counted from, so every answer can be checked. In testing we ask it questions where we already know the result from existing reports and compare. When a question is imprecise or the data does not reach that far back, the agent says so instead of producing a figure.
Does this make sense for a small company with three systems?
Sometimes it does not, and we say that in the consultation. With two or three sources and one question a week, a tidy spreadsheet with automatic loading is enough, cheaper and faster to stand up. A warehouse starts to pay off when the questions come several times a day, the sources go past four, or two departments quote different numbers.
Let us find the questions your data cannot answer today
The free consultation takes 30 minutes and you write directly to Paweł, the founder. You list the three questions that come up most often in your company and say where you look for the answers today. We show which ones can be served from a single source straight away and which need the data tidied first. We can demo the AI agent on sample data, without touching your systems.
We run at most two rollouts at a time, so that someone is actually sitting on each of them.
Book a free consultationPS. If someone at your company built a report out of four exports this week, that is exactly the problem. Data collected by hand is always a few days behind the decision it is meant to support. After the rollout the same question comes back in 5 to 10 seconds, and the dashboard loads itself.