News · 3 Jun 2026 · MTW Editorial Team
Microsoft Data Formulator is a free, open-source desktop tool that lets you build charts and explore data by describing what you want in plain English, and version 0.7 turns it from a clever research prototype into something a working UK analyst can genuinely use. It is a Microsoft Research project released under the permissive MIT licence, it runs entirely on your own machine, and it connects to whichever AI model you already pay for rather than locking you into one vendor. For analysts, small businesses and students who want AI-assisted charting without a new subscription, this is one of the more interesting releases of the year.
Key facts: what Data Formulator 0.7 is and why it matters now
Version 0.7.0 landed on 28 May 2026. The headline is that the tool now reads from real databases and warehouses, not just files you drag in, and it wraps everything in a single conversational agent that remembers the thread of your analysis. You install it with one command, point it at a model and a dataset, and ask for the chart you want. The software writes the data transformations and the visualisation specification for you, then shows its working so you can correct it.
The reason it matters now is cost and control. Most AI charting features are arriving as paid add-ons inside subscriptions, a pattern we covered in our look at the Microsoft 365 Copilot UK price rise on 1 July 2026. Data Formulator takes the opposite route: the application is free, and you only pay for the model tokens you actually consume. For a UK sole trader or a university student that difference can be the gap between using AI for data work and not bothering at all.
What you actually get in the box
The 0.7 release adds several pieces that, taken together, change how the tool feels day to day. There are governed data connectors for databases, warehouses, BI systems and object stores, so you can pull from a live source rather than exporting a spreadsheet first. A data-loading agent helps discover tables and extract the right ones. A unified conversational agent now carries thread memory, so a follow-up like “now split that by region” understands what “that” refers to.
On the charting side there is a semantic engine offering more than 30 chart types, covering the usual bar, line and scatter plots but also area, streamgraph, candlestick, radar, maps and single-figure KPI cards. A style-refinement agent lets you adjust colours, labels and layout by asking in words instead of hunting through menus. Finally, there is report creation and export to image or PDF, plus persistent sessions and workspaces so your analysis survives a restart. The interface ships in English and Chinese.
Installing it on a UK machine: the exact commands
Setup is deliberately light. You need Python 3.11 or newer on the machine. The quickest route, with no permanent install, is to run uvx data_formulator, which fetches and launches the app in one step using the uv package runner. If you prefer a standard install, use pip install data_formulator and then start it. Either way the application opens in your browser at localhost:5567, served from your own computer rather than the cloud.
Once it is running you tell it which model to use. Data Formulator connects to OpenAI, Azure OpenAI, Anthropic and locally run Ollama models through LiteLLM, which means you bring your own model and your own key. If you already have an Anthropic key, the setup mirrors the approach we describe in our guide to Claude for UK accountants, where the model does the reasoning and you keep the data in your own hands. There is no Data Formulator account and no separate licence fee.
The data-governance question every UK business must answer
Running the app locally does not mean your data stays local. When you ask for a chart, the tool sends a description of your columns, and sometimes sample values, to whichever model you have configured. If that model is a hosted service such as OpenAI or Anthropic, your business data leaves the UK device and travels to a third party. Under UK GDPR that is a processing activity you are responsible for, and the Information Commissioner’s Office expects you to know where personal data goes and on what lawful basis.
The practical answer is the local-model route. Point Data Formulator at an Ollama model running on your own hardware and nothing leaves the building, which suits sensitive client records, health data or anything covered by a confidentiality clause. For lower-risk numbers, a hosted model on a business tier with a no-training-on-your-data commitment may be acceptable, but you should record that decision. UK firms weighing these trade-offs at scale will recognise the pattern from our analysis of what UK enterprises can learn from the 743,000-seat Copilot rollout, where governance, not capability, was the limiting factor.
A short checklist keeps you on the right side of the rules. First, decide whether the dataset contains personal data; if it does, a hosted model needs a lawful basis and, in many cases, a data processing agreement with the provider. Second, prefer pseudonymised or aggregated columns where you can, so the model never sees a name or an account number it does not need. Third, write down which model you used for which task, because the ability to show that decision trail is exactly what the regulator looks for. None of this is onerous, but it should happen before the first real upload, not after.
It is also worth being clear about what the tool does not do. Data Formulator is not a data store, a sharing platform or a managed service, so there is no central admin console to lock down access or audit who ran what. That responsibility stays with you and your existing systems. For a solo analyst that is fine; for a regulated team it means the tool belongs in a controlled sandbox rather than wired straight into production data.
Microsoft’s own framing of how organisations are adopting AI sets useful context for where a tool like this fits. The point for a UK reader is that the governance work sits with you, the user, regardless of how polished the software is. A free tool removes the cost barrier but not the responsibility, and that is the part worth slowing down for before you connect a live customer database.
Data Formulator versus Power BI Copilot: who each one suits
The obvious comparison is Power BI Copilot, Microsoft’s paid AI layer inside its mainstream BI platform. The two are aimed at different people. Power BI Copilot suits organisations already standardised on Microsoft 365 and the Power Platform, where governance, sharing and dashboards are managed centrally and the AI features arrive as part of a licence you already hold. It is polished, supported and built for teams.
Data Formulator suits the individual analyst, the curious student and the small business that wants to experiment without committing to a platform or a per-seat fee. Because it is open-source and model-agnostic, it is also a sensible teaching tool: you can see the transformations it generates and learn from them. The honest limitation is that it is a research project, so you should not expect enterprise support, service-level agreements or the collaboration features a paid suite gives you. Buyers comparing the wider Microsoft toolset may also want our take on what the Microsoft Build 2026 keynote means for businesses.
Specifications and key takeaways at a glance
| Spec | Detail |
|---|---|
| Version and date | 0.7.0, released 28 May 2026 |
| Maker | Microsoft Research |
| Licence | MIT (free, open-source) |
| Cost | Free software; you pay only for model tokens |
| Install | uvx data_formulator or pip install data_formulator |
| Requirement | Python 3.11 or newer |
| Runs at | localhost:5567, locally in your browser |
| Models | OpenAI, Azure, Anthropic, Ollama via LiteLLM |
| Chart types | 30-plus via a semantic engine |
| Export | Reports to image or PDF |
The table above is the short version for anyone scanning. The single most important line is the licence: MIT means you can use it commercially, modify it and run it on as many machines as you like with no fee. The second most important is the model line, because your choice there decides both your running cost and your data-governance posture.
How it fits a typical UK analyst’s week
Picture a finance analyst at a mid-sized UK firm with quarterly sales data in a warehouse. With Data Formulator they connect the database, ask for revenue by region as a map, then refine it by saying “show year-on-year change and use a colourblind-safe palette”. The style-refinement agent handles the formatting, and the report exports to PDF for the board pack. The same workflow suits a student analysing open council data for a dissertation, or a marketer slicing campaign numbers.
The thread memory is what makes this feel different from one-shot chart generators. Because the conversational agent remembers earlier steps, you build an analysis rather than firing isolated requests. If your work involves writing up findings as well as charting them, you may also find value in our roundup of the best AI writing assistant in the UK for 2026, since the two tasks often sit side by side. For SMEs scoping any AI build, the practical lessons in our Code with Claude lessons for UK SMEs piece apply equally here.
A realistic first session goes like this. You spend five minutes on install, a couple more pointing the tool at a model and a key, then you connect a data source. The data-loading agent surfaces the tables it can see, you pick one, and you ask your first question in plain words. The early charts are rarely perfect, which is the point: you correct them in conversation, watching the generated transformations so you understand what the tool changed. Within half an hour most analysts have a working chart and a feel for where the agent is reliable and where it needs a firmer hand.
The honest caveat is that the quality of the output tracks the quality of the model behind it. A small local model keeps your data private but may misread an ambiguous request, while a large hosted model handles nuance better at the cost of sending data outside your machine. That trade-off is the single decision that shapes the whole experience, and it is worth testing both on a non-sensitive sample dataset before you commit a real workflow to either.
Where to buy or check next in the UK
There is nothing to buy: Data Formulator itself is free from the official Microsoft repository on GitHub, where you download it, read the documentation and check the latest release notes before installing. What you may need to budget for is the surrounding kit. If you plan to run a local Ollama model for governance reasons, check that your machine has enough memory and, ideally, a capable GPU.
For hardware, named UK retailers such as Currys, John Lewis and Argos list business laptops and desktops with the RAM headroom local models need, so compare price, delivery and returns windows before committing. Amazon UK and Very are worth checking for the same machines on finance terms, while AO.com often competes on delivery. If you want a Microsoft-first setup, the brand’s own UK store and our guide to Microsoft Surface for Business in 2026 cover warranty and support cover. Buyers weighing a cheaper route should also read our Googlebook versus Chromebook comparison before assuming a lightweight laptop will cope. Whatever you choose, confirm the warranty length and the returns policy in writing.
Our verdict: who should install it and who should wait
Our view is that Data Formulator 0.7 is an easy recommendation for individual analysts, students and curious small businesses who already pay for a model or can run one locally. The price is right, the install is trivial, and the thread-aware agent genuinely speeds up exploratory charting. It is the most approachable version yet of a tool that used to feel like a lab experiment.
We would tell larger teams that depend on central governance, support contracts and collaboration to wait, or to treat this as a sandbox alongside Power BI rather than a replacement. The thing that would change our recommendation is official support and a managed deployment path; until then, the research-project status is the catch. Install it to learn and to prototype, govern your data deliberately, and keep your paid BI platform for anything mission-critical.
What we like and what we’d watch
| What we like | What we’d watch |
|---|---|
| Free MIT licence with no per-seat fee | Research project, so no formal support or SLA |
| Bring your own model, including local Ollama | Hosted models send data to a third party |
| Thread memory makes multi-step analysis natural | Limited team collaboration versus a paid suite |
The balance is clear. The strengths are about freedom and cost, and the cautions are about support and data flow. Neither side is a deal-breaker for the audience this tool is aimed at, provided you go in with eyes open about governance.
Reader discussion
Leave a comment
Comments are moderated. Keep it useful, accurate, and on topic.