We build the governed metric layer first, so the board pack, the dashboards and the models all read the same number. Scoped before we quote, senior hands only, and a written stop point at every stage of the work.
30 minutes, camera optional, no slide deck. Replies within one working day. One new engagement starts at a time.
You see real numbers in your own data in week two, not a slide deck.
Written scope first, then a fixed quote against it. No open-ended discovery.
If the numbers cannot be reconciled to your finance system, we keep going at no charge. In writing, in the contract.
Repository, semantic model, documentation and definitions all move to your tenant at handover.
Which market is dragging group margin down?
26.6% in 2025 against 29.1% in 2024, down 2.5 points
What the model is telling you: Four markets moved inside a point. DE accounts for essentially all of the group decline, so the next question is inside one market rather than across five.
select m.market,
sum(r.net_revenue) as net_revenue,
sum(f.cogs + f.inbound_freight) as cost_of_sales,
1 - sum(f.cogs + f.inbound_freight)
/ nullif(sum(r.net_revenue), 0) as gross_margin
from analytics.fct_order_lines f
join analytics.dim_market m using (market_key)
join analytics.fct_invoiced_net r using (order_line_key)
where f.iso_week between {{start_week}} and {{end_week}}
group by 1
order by gross_margin asc;Computed in your browser from 520 bundled weekly snapshots (520 rows scanned for this answer) over market x ISO week grain. Northline Distribution AB (synthetic) is a synthetic data set generated for this demo, not client data. Definitions, owner and source shown are the ones the layer would carry in a real deployment.
This is the actual demo, not a screenshot. Ask it a question and it recomputes from the bundled series.
We are rarely the first call. We are the call after the first attempt has been running for a year and the numbers still do not agree.
Finance and commercial arrive with different revenue figures and both are defensible. The rest of the meeting is spent on reconciliation.
Reporting was built, licensed and launched. The Monday meeting still starts with someone updating a spreadsheet.
A proof of concept impressed the board and has been in review for four months. Nobody can say what it costs to run or who owns it.
One analyst understands the queries. Their holiday is a business continuity risk and everyone knows it.
The numbers tie out eventually, by hand, after three weeks of extract reconciliation and a lot of goodwill.
The common cause: reporting was built on top of whatever definitions were available at the time, and nobody wrote the arithmetic down. Fix that first and the rest gets cheaper.
Start at the metric layerEvery engagement runs through the same order: define the numbers, model them once, build on top, then hand over. The AI practice is not an exception.
One definition per number, written down, owned by a named person, and enforced in a semantic layer instead of a spreadsheet.
Pipelines that fail loudly instead of silently, a warehouse that answers a query before the meeting ends, and lineage you can hand to an auditor.
Fewer reports, read more often. Reporting that survives the person who built it leaving.
Eight service lines, sold the same way as everything else: a short fixed-scope readiness step, then a pilot that has to beat your current process, then production only if it does.
The second failure is the expensive one. It happens when an upstream column changes shape, a definition moves, or an access rule the prototype never had to respect turns out to be the whole problem. We start where those things live: the data contract, the metric definition, and the approval gate.
“Add AI to every process.”
Most processes we see should be fixed with a data contract and a workflow change. We will name those in the readiness sprint and quote nothing for them.
“Agents that run themselves.”
An agent with write access and no approval gate is an incident waiting for a timestamp. We build the gate first, always.
“AI-powered insights.”
Insights come from a defined metric and a decision owner. The model is the cheap part. If that sentence was the pitch, the project has no measurable outcome.
Three of eight samples. Each one leads with a number, states how it was measured, and prints the sample size where the number depends on one.
How to read these samples. Illustrative example. This is a composite of the kind of engagement we deliver, built from our own method and representative market benchmarks. It is not a specific client project and the figures are directional, not verified client results. We can walk you through anonymised, NDA-covered engagements on a call: the numbers, the stack, and what did not work.
A Nordic distributor, five markets, around 210 employees
A utility supplier, around 38,000 supplier invoices a year
A medtech manufacturer, 340-document regulatory corpus
A named method with a written exit criterion at each gate. You can stop at any gate and keep what was built up to that point.
Why gates rather than phases. A phase runs until the budget ends. A gate has a test you can fail. Gate 02 does not close until last year-end reconciles, and if it does not, the next phase is free until it does. That sentence is in the contract, not in a proposal.
Handover is a gate, not a favour. Gate 05 closes when two people on your team can add a metric end to end while we watch.
We define the numbers before we touch a database. Every metric gets a plain-language definition, a formula, an owner and the decision it is meant to drive.
Stop point. Stop here and you keep the sheet. Several clients have.
The definitions become a semantic layer with documented grain, relationships and tests. This is the part that makes reporting and AI agree with each other later.
Stop point. Stop here and you own a model someone else can build on.
Reporting, pipelines or AI sit on top of the model. Page design starts from a decision and an owner, not from a chart type. Anything that does not change a decision does not get built.
Stop point. Stop here and the artefacts ship documented and tested.
We measure adoption, reconciliation and the original business case, and we report the parts that did not work alongside the ones that did.
Stop point. If reconciliation fails, the next phase is free until it passes. That is in the contract.
Repository, model, documentation, definitions, credentials and a recorded walkthrough move to your tenant. Named people in your team can change a definition without calling us.
Stop point. No retainer is required to keep the lights on. Most clients keep one anyway.
Not positioning language. Each of these is a commitment you can hold us to, and a reason to say no to us early.
No price list, because no two clients run the same systems. You get a written scope with deliverables, dates and assumptions, then a fixed quote against it within two working days. You are never asked to sign on the call.
Written scope and fixed quote before any commitment.
Week two produces numbers from your data, in your environment, that a decision owner has looked at. If by week two you have only seen slides, we have not started.
Gate 02 exit criterion: last year-end reconciles.
The person who scopes the work builds it. No account manager between you and the engineer, and no junior learning your business on your invoice.
Fixed-scope engagements are capped at two concurrent projects.
Repository, semantic model, definitions, documentation and credentials move to your tenant at handover. No proprietary runtime, no licence back to us, no data hostage.
Gate 05: two of your people add a metric unaided before final payment.
If the data does not support the use case, or the model does not beat your current process, that is a written deliverable. Two of the three answers in a readiness sprint are reasons not to build.
AI readiness sprint ends build / fix the foundation / stop.
Every project starts at the metric layer because that is where reporting and AI both break. It is a slower first fortnight and a cheaper rest of the year.
Gate 01 runs before any pipeline work.
Two companies can ask for the same dashboard and need a week apart in effort, so we scope first and price per client. What you get is a scope document you can read in one sitting, and a fixed quote you can hold us to.
What the problem is, which service line fits, and whether a smaller first step settles it.
Deliverables, gates, assumptions and what is out of scope, within two working days.
Priced against that scope, per client. Larger builds are split across gates with a milestone each.
If your question is not here, ask it directly and we will put the answer on this page.
Client by client, after scoping. Two companies can ask for the same dashboard and need a week apart in effort, so a published price list would either be wrong or would push us to cut corners to protect it. What you get instead: a written scope with deliverables, dates and assumptions, then a fixed quote against that scope within two working days. The number does not move unless you change the scope.
A hire costs 12 to 18 months of salary before they know your business, and they leave with the context. We are useful when you need one specific thing built properly: a metric layer, a warehouse, a compliant reporting setup, an AI workflow that survives production. If what you actually need is a permanent analyst, we will say so, and we do run fractional retainers when that fits better.
Usually the semantic model. Reports inherit whatever definitions were available when they were built, so the same metric drifts across pages. We harden the model, add descriptions and hierarchies, and put a verified-answer set in place, which is also the prerequisite for Copilot giving correct answers rather than confident wrong ones.
Every AI engagement starts with a short fixed-scope readiness step that ends in a decision rather than a commitment. Nothing is billed for a phase that does not meet its written exit criterion, and if a pilot cannot beat the process it is meant to replace on your own data, you do not buy the build. We put that in writing, including when the honest answer is not to build it at all.
Bring it to a 30-minute call. We will tell you what usually causes that specific disagreement, what settling it takes, and whether you need us at all. If you do not, you will still leave with the definition sheet.
Remote team, operating from Dhaka, Bangladesh · Asia/Dhaka (UTC+6), with a standing overlap into European mornings · Replies within one working day.