Analysis Machine
Chat with your data, perform deep analysis, and get answers to your burning questions in minutes, not days or weeks.
What it is
Not a dashboard and not text-to-SQL. You ask a commercial question in plain words; it works out what you actually mean, asks back when the question is vague, runs the queries, and writes up an answer with its working shown.
It can do that because your business is described to it once — the processes, the rules, the exceptions — and read from the systems you already run.
How it works
01 / Ask
No query language, no dashboard to find first, no request sitting in an analyst’s queue. Ask the way you’d ask a colleague, and follow up the same way.
02 / Describe
Processes, rules and the exceptions everyone knows but nobody documented are held as an ontology, reviewed before anything runs on it.
03 / Connect
Read the data lake, APIs, systems of record or vendor portals the business actually runs on, in real time.
04 / Build
Turn an answer into a visualisation, analysis, dashboard or app, versioned with an owner and audience.
05 / Model
Model the decision before the spend is committed, then put an app on top of it and share it with the team.
06 / Trust
Every answer shows which sources, rules, queries and ontology version produced it. Every edit is auditable.
Industries
The problem
Promotions get the attention, and they are genuinely hard. But in the businesses we have worked in, the larger share of volume moves without one — and that part often gets forecast by copying last year forward.
Retailer POS, distributor feeds, shipments and finance rarely agree, so the weekly number gets argued about before anyone acts on it.
How Analysis Machine helps
Real impact
980 → 40
One legacy report re-modelled from 980 measures to 40, surfacing calculation errors that had sat unnoticed.
5×
Price elasticity re-fit on the business’s own retail data came back roughly five times the textbook default.
20 min
A full weekly account report, produced by one team while the others were still on the road.
Anatomy
Sources
The system is source agnostic: traditional data infrastructure, ERPs, CRMs, APIs and even manual imports. AI-powered custom importers clean and structure data that needs to be used organisationally.
Warehouse
Snowflake, Databricks, Fabric, BigQuery, Redshift or Postgres. Read-only SQL under the same row-level rules as any other consumer. Existing dbt or metrics layers remain the definitions of record.
The ontology
A lightweight model of how your company operates in the real world, including the messy parts: which SKU hierarchy is real, which revenue table finance trusts, and the exceptions everyone knows.
Reasoning
A question becomes a plan, a person validates or adjusts it, the plan becomes SQL, the SQL runs, and the result is written up. Every step is recorded.
Data security
AI models read the ontology, plan and write SQL, but the data itself is never sent to the model. Analysis runs as code inside a secure sandbox.
Surfaces
A web workspace, a Slack or Teams bot, and an API. Most companies start with the workspace and end up living in the bot.
Models
Model choice is yours: Anthropic, OpenAI, Bedrock, Azure, Vertex or something you host—interchangeably, in production.
Alongside your BI
Power BI, Tableau and Looker keep running, but you will probably find yourself using them only for the trickiest questions.
Data, access & governance
Access is controlled by whatever your warehouse already says. The user’s identity passes through; if someone cannot see the Germany rows in Snowflake, they cannot get them by asking nicely.
This is not a multi-tenant SaaS app. Everything is hosted on your infrastructure, under your control, and nothing is used to train the AI models.
Every answer shows its query and sources and says plainly when the data is thin. An answer you cannot check is worse than no answer.
Every question, query and result is logged against the person who asked. The log is yours, in your warehouse, and queryable like anything else.
Read-only by design. Remove it tomorrow and your stack runs as it did before. Write access is a different beast altogether.
It gets the room to the same set of facts faster. Forecasting, planning and judgement remain with your people.
Deployment
Inside
Your cloud account. Your region. Single tenant.
Identity
Your IdP, through SAML or OIDC. SCIM to provision. Nobody gets a second login.
Analysis Machine
Reasoning, the ontology and the clause index, running as your workload on your compute.
Queries
Run in your warehouse, on your credits, under the same row-level rules as any other consumer.
Logs
Every question, query and result, written to your own tables. Queryable like anything else.
Outside
Machine & Folk
What we provide
What stays in your account
Integration
Read-only SQL as a service account. Several at once, so a question can cross two.
Their own APIs, or the tables you already land in the warehouse.
Indexed in place. Nothing is copied into a second store.
Read as the definitions of record, not rebuilt beside them.
Pushed to where people already work. Same answer, same trace, whichever surface.
Latency & freshness
~1
A simple query resolves to SQL fairly quickly. Of course sometimes the simplest questions are the hardest to answer.
2–10
A hard question can be very hard—the kind your team struggles to answer. Here, the tool is a thought partner and collaborator rather than a gofer.
live
Not every source is live. Some refresh weekly, monthly or quarterly. The system understands that and does not answer on data it does not have.
none
We save the SQL, not the data. The same question tomorrow is guaranteed to get tomorrow’s data.
10s
On a mid-sized warehouse. The system scales according to your needs and usage.
mins
Changes to the ontology or agent rules take effect immediately. No upgrade or redeploy is required.
How it is deployed
Unlike SaaS applications, this software is completely tailored to your business. Installation mostly involves connecting to your data sources and tuning the ontology to your business processes.
Weeks 1–2
We sit with the people who need answers and write down what they actually ask. No requirements document.
Weeks 3–4
Connected to one real source, answering a handful of real questions. Wrong often enough to be useful.
Weeks 5–6
A first team uses it for real work. What they ask next decides what gets built next.
After
Your engineers have been involved since week one. We stay as long as that is useful and no longer.
See it working
01 · Watch
Fifteen minutes, on our data. What a question looks like going in and an answer coming out, end to end.
Watch the video02 · Meet
We can tailor a presentation to your business. Share your data schema with us and it’ll feel like home.
Tell us the question your team can’t currently answer. We’ll tell you honestly whether this is the way to answer it — sometimes it isn’t.
Because it's not for everyone. But if you're interested in the cutting edge of AI transformation, you could do worse than our monthly missives.
By subscribing, you agree to our privacy policy. Unsubscribe anytime.
