A follow-on to "Who Needs the Statisticians Workbench?" — the many kinds of work that benefit from inexpensive statistical reporting, and how local AI makes it genuinely private.
The previous document, Who Needs the Statisticians Workbench?, made the case for one audience in particular — middle management of corporate departments, together with managers of smaller organisations. That is where we believe the maximum benefit lies. But the Workbench's core value — sophisticated statistical reporting at a price and speed ordinary budgets can afford — extends well beyond that one audience.
This document lays out the potential uses of the Workbench across a wider range of human activity, and then focuses on the aspect that makes it suitable for some of the most sensitive of those uses: privacy. Because the Workbench can be configured to use a local AI service rather than a cloud one, it can keep analysis — data and all — entirely within an organisation's own four walls.
Teaching statistics is hard because the tools are usually either too simple to be real, or too expensive and complex to put in front of every student. The Workbench offers a middle path: a real statistical platform, with genuinely sophisticated methods, that students can actually use.
An instructor can demonstrate a t-test, an ANOVA, or a regression on real data and show the plain-language interpretation alongside the numbers, so the meaning of the result is as visible as the calculation. Students can repeat the analysis themselves without needing to install a specialist package or learn a programming language first. And because the AI concierge can explain what is happening at each step, the system doubles as a tutor — a student who is unsure which test to use, or what a p-value means, can ask, in plain language, and get an answer.
For institutions this also removes a cost barrier: a statistical learning environment that runs on ordinary, self-hosted infrastructure, rather than per-seat licences for commercial software.
One of the least visible but most valuable problems in a large organisation is that the same number gets calculated differently by different people. Two analysts asked the same question — "what was our average resolution time?", "is this region outperforming that one?" — can produce different answers because they made different methodological choices along the way.
The Workbench can act as a standardising layer. Because the statistical methods are built in and applied the same way every time, an organisation can define a common approach — the same test, the same treatment of the data, the same interpretation — and have it applied consistently across every department. The same report run by two different teams produces the same result, because the underlying engine is deterministic. This turns reporting from a matter of individual analyst judgement into a repeatable, auditable process, which is exactly what an organisation needs when reports start being used to compare departments or to make decisions.
Much real research is done by small teams with no dedicated statistician — a clinical pilot study, a field trial, a survey, a longitudinal observation of some cohort. These projects genuinely need statistical analysis, but the budget for a biostatistician is often absent, and the analysis ends up either done crudely in a spreadsheet or not done at all.
The Workbench gives such a team access to the standard inferential methods — t-tests, one- and two-way ANOVA, chi-square, regression, correlation — with plain-language interpretation, so the team can do a proper initial analysis themselves. For work that ultimately needs a professional statistician, the Workbench is still valuable: it lets the team explore and prepare the data properly first, so that the expensive specialist time is spent on the hard questions rather than on basic cleaning and description. And because the analysis can be kept private, the raw research data does not have to leave the research group.
Modern farming produces a great deal of data — yields, input costs, weather, soil test results, livestock records, trial plots comparing seed varieties or fertiliser regimes — but the people making the decisions are usually not statisticians. A farmer or farm manager wants to know: did this variety actually outperform the other one, or was the difference just this season's weather? Is the fertiliser change really paying for itself? Which paddock is genuinely under-performing?
These are precisely the questions the Workbench's methods are designed to answer, and at a cost a farm operation can justify. Analysis can run on the farm's own data, without sending it to an external service, and the results are interpreted in language a manager can act on. The non-destructive data preparation is well suited to the job too: trial data can be filtered and cleaned for analysis without disturbing the original records.
Government departments face a combination of pressures that makes inexpensive statistical reporting especially relevant: they are publicly accountable for measurable outcomes, they must be able to defend their figures, and their budgets for external consultants are increasingly constrained. At the same time, much of the data they hold is sensitive — information about citizens, service users, or internal operations that should not be handed to a third party.
The Workbench addresses both sides of this. It provides the statistical methods a department needs for reporting — performance measurement, comparing cohorts or regions, tracking outcomes over time — without the cost of an external engagement. And because it is self-hosted and can run with a local AI service, the analysis can be carried out without the data leaving the department's control. For public bodies, where both accountability and privacy are non-negotiable, that combination is hard to match.
The five uses above are the ones we were asked to lay out. But the same logic — sophisticated statistics, low cost, understandable results, and privacy — applies far more broadly. Any field where someone must make decisions from data, but cannot afford a statistician, is a candidate. Others that come to mind:
| Field | What inexpensive statistical reporting enables |
|---|---|
| Clinical practice & small healthcare | Clinics and practices auditing their own outcomes, tracking quality-improvement initiatives, and analysing patient data — without sending medical data to a third party. |
| Non-profits & NGOs | Measuring the impact of programmes, evaluating whether an intervention actually changed outcomes, and reporting to donors with rigour rather than anecdote. |
| Sports & coaching | Amateur and semi-professional clubs analysing performance data, comparing training regimes, and testing whether a change genuinely improves results. |
| Retail & e-commerce | Small businesses analysing sales patterns, testing whether a promotion or a price change actually lifted revenue, and understanding what predicts a repeat customer. |
| Manufacturing & quality control | Analysing defect rates, testing whether a process or supplier change improved quality, and tracking variation over time. |
| Environmental monitoring | Conservation groups and local agencies analysing weather, wildlife, or pollution data to detect real trends amid natural variation. |
| Journalism & data reporting | Small newsrooms testing whether a claim about the data is actually supported, rather than relying on a press release's summary. |
| Human resources & people analytics | Departments analysing retention, hiring, and performance data — where confidentiality of employee records makes local, private analysis essential. |
| Veterinary practice & animal care | Practices analysing treatment outcomes and caseload patterns in their own records. |
| Hospitality & tourism | Operators analysing occupancy, booking patterns, and the effect of pricing changes on demand. |
The common thread is not the subject matter — it is the situation. In every one of these fields there is someone who must make a decision from data, who would benefit from real statistical rigour, and who cannot justify hiring a statistician or engaging a consulting firm to get it. That is the recurring gap, and it is the gap the Workbench is designed to fill.
Several of the uses above — healthcare, government, human resources, research involving personal data — share a hard constraint that a manager in any other field might not face: the data cannot leave the organisation. Patient records, citizen data, employee records, and unpublished research are all subject to legal duties, professional obligations, or simple common sense that forbids handing them to an external service for processing.
This is where the choice of AI service becomes decisive. The Workbench's statistical engine is already self-hosted: it runs against your own data, on infrastructure you control, so the data never has to leave to be stored or calculated. But the AI guidance layer — the concierge that interprets the user's request and explains the results — has historically been the weak point, because a typical setup calls a cloud AI service, which means at least some of the data and the questions being asked are sent to a third party's servers.
For an open dataset that may be acceptable. For a dataset of patient outcomes, or citizen records, or unpublished trial results, it is not. The privacy of the whole system is only as strong as its weakest link, and a cloud AI call is exactly such a link.
The Workbench removes this weak link. It can be configured to use a local AI service — such as Ollama — so that the language model runs entirely on the organisation's own hardware. No query, no data, and no result is ever sent outside. The whole pipeline — data, analysis, and the AI guidance that sits on top of it — stays private.
Ollama is a way to run open-weights language models locally, on your own machine, without depending on a cloud provider. When the Workbench is pointed at a local Ollama instance instead of a cloud API, the AI concierge's work — understanding the request, guiding the analysis, explaining the result — is performed by a model running on hardware the organisation controls. Nothing is transmitted, because there is nothing to transmit to.
The practical question has always been whether a locally-run model is good enough to be useful. The answer, for the guidance work the Workbench needs, is that it can be. The concierge's job is not to perform the calculation — that is done by the deterministic engine — but to interpret and explain, and for that a capable open-weights model running locally is entirely adequate.
And the hardware to do it is becoming ordinary. Ollama running on an Apple Mac M5 — a high-specification machine — can provide a reasonable level of AI performance. Apple Silicon is well suited to running local models: its unified memory lets the model use a large amount of memory directly, and its integrated neural-engine hardware accelerates inference without requiring a separate, expensive data-centre GPU. A high-end Mac is a machine an organisation may already have, or can obtain at a fraction of the cost of the specialist hardware that local AI used to demand.
That is the point that changes the equation. Local AI has historically been treated as something only a well-resourced IT department could run. But when a standard high-specification desktop machine can provide a reasonable level of performance, truly private statistical reporting comes within reach of most budgets. The organisation does not need a cloud AI subscription, does not need to send its data off-site, and does not need to accept the privacy trade-off that cloud AI has always implied.
Read together with Who Needs the Statisticians Workbench?, the picture is complete. The Workbench brings sophisticated statistical reporting to the people who could not previously afford it — corporate middle management above all, but also educators, researchers, farmers, government departments, and a wide range of other fields where decisions must be made from data.
And it does so without the one compromise that has traditionally disqualified self-service analytics from the most sensitive kinds of work: the data stays put. With a local AI service such as Ollama, and hardware as ordinary as a high-specification Mac, the flexibility of AI guidance and the reliability of deterministic statistics can both be had entirely in private. For anyone whose data cannot leave the building, that is the difference between being able to use the tool at all and not.