The business OS · Research & analysis

Ask the data.
It runs the tabs.

An analyst that reads your SPSS file, runs the actual statistical functions — cross-tabs with significance, key drivers, TURF, segmentation, rim weighting — and answers in a sentence. It does not estimate the number. It computes it, and shows you what it ran.

One module, 5 products. Sold separately almost everywhere else as a tabulation and statistics package, a data-validation tool, an open-end coding tool, a focus-group analysis tool and an insights-dashboard portal. Fieldwork is not one of them — bring the data you already collected.

Ask the data

Which drivers move overall satisfaction most, and does the metro/non-metro gap actually hold?

search_questionsmatched 6 satisfaction items in the datamap

get_variable_detailsvalue labels, missing codes, weight variable

execute_functiondriver_regression · standardised betas, VIF checked

execute_functioncross_tab · satisfaction × region, significance at 95%

execute_queryverifying the base before answering

Two drivers carry most of it — resolution time and first-contact accuracy — and they are not collinear, so they are separate levers rather than one measured twice. The metro gap is significant; the tier-2 gap in the same table is not, on an effective base too small to call. Every figure came from a function call, on the effective base, not from the model

The same agent builds the dashboard — below

What it covers

From the data file to the deck.

The unglamorous middle of a research project — the part between receiving the data and having something to present.

Your data fileAn SPSS .sav goes in and its variables, value labels and missing codes come with it. Waves are held as waves, so a tracker compares like with like instead of you re-cutting every round.

Appending the next waveA new wave is dry-run against the existing datamap first, so you see what changed — new variables, relabelled codes, a file that looks wrong — before anything merges. The commit re-checks a fingerprint of that datamap and refuses if it moved in between.

The questionnaire, parsedThe instrument is read into blocks and questions and mapped to the data, including across waves — so a question that was re-worded in round three still lines up with round one. It is then reconstructed from what was parsed and compared back to your document, so a misread question shows up as a coverage gap rather than as a table that quietly never appears.

Twenty-five validation checksSpeeders, flatliners, patterns, ranges, skip and gate routing, punch validity, grids, ranks, constant sums, duplicates, outliers, dates, codes, PII and open-end quality — written in a script with loops and conditionals, so a fifty-question grid is one loop rather than fifty pasted blocks.

Cross-tabs with banners and netsThe table you would have built in a tab package: banner points, nets and groupings, base rows, and the counts and percentages agreeing with each other.

Custom tablesA pivot engine underneath, so a table is something you specify rather than something you pick from a menu. Any variable on either axis, with the filters you want.

Significance that respects the weightZ-tests and t-tests with p-values, standard errors and confidence intervals — computed on the EFFECTIVE base, so weighting cannot quietly inflate how confident the letters look.

Rim weightingIterative raking to your targets, run to convergence, with the weights kept on the file rather than applied invisibly at report time.

TURF, drivers and segmentsUnduplicated reach and frequency for portfolio decisions — searched exhaustively where the item set is small enough and greedily where it is not. Key-driver regression reports standardised betas, p-values, R² and adjusted R², with VIF checks so collinear inputs are flagged instead of quietly inflating a driver. Plus cluster segmentation, a correlation matrix and trend decomposition for trackers.

Derived variablesCompute new variables from existing ones, group codes into nets, and control how missing codes are treated — in the tool, not in a spreadsheet afterwards.

Open-end codingAn explicit codeframe, coded responses with a confidence figure, a job you can cancel, and an audit log of every call the model made. Not a black box that returns a chart.

Open-ends in the tablesCoded verbatims behave like any other variable — cross-tabbed and frequency-counted alongside the closed questions, not stranded in a separate deck.

Focus groupsRecordings uploaded, transcript assembled, speakers separated and the moderator identified, themes and sentiment drawn out with the verbatims kept. Exports to .sav so the qual sits beside the quant.

Desk researchPoint it at a question and it builds a sourced secondary-research dashboard, keeps the history, and lets you edit what it produced.

Ask the dataAn agent, not a search box. It finds the right question in the datamap, calls the statistical functions — cross-tabs, drivers, TURF, segmentation — reads the results and answers. Where no function fits it writes a query, which is checked at the syntax-tree level and capped on time, rows and memory before it runs, then written to an audit log. Every figure traces back to the call that produced it.

A dashboard it builds itselfPoint it at the file and it works in seven passes — reconnaissance, core metrics, cross-tabs with significance, advanced analytics, a verification pass over its own numbers, assembly, then market context. The mandatory passes are mandatory; it cannot quietly skip the cross-tabs and hand you four pie charts.

Run the same analysis next waveEvery tool call the agent made is kept in order as a recipe, stored beside the dashboard. Replay it mechanically to move only the numbers, smartly to have the narrative rewritten around them, or fully to start again — which is what makes a tracker comparable round to round.

Publish or embedA dashboard goes out on a share link with view and visitor tracking, or embeds in a client portal — with its own key you can rotate and its own session-level analytics. The same report the reader sees is the one you built.

Straight answers

The things people ask on the first call.

Does the AI make the numbers up?

It cannot. It answers by calling the same statistical functions the tabulation engine runs — cross-tabs, driver regression, TURF, segmentation, weighting — and reading what comes back. Where no function fits it writes a query, which is validated at the syntax-tree level and capped on execution time, rows and memory before it runs. Every figure in an answer traces to the call that produced it.

Can it actually build a dashboard, or just chat?

It builds the dashboard. Pointed at a file it works in seven passes — reconnaissance, core metrics, cross-tabs with significance, advanced analytics, a verification pass over its own numbers, assembly, then market context. The cross-tab and advanced-analytics passes are required, so it cannot skip the hard part and hand you four pie charts.

Can we re-run the same analysis on the next wave?

Yes, and this is the part most AI analysis cannot do. Every tool call is recorded in order as a recipe stored beside the dashboard. Replay it mechanically to move only the numbers, smartly to have the narrative rewritten around them, or fully to analyse the wave from scratch. A tracker rebuilt by hand each round is a tracker whose rounds are not comparable.

How is significance calculated on weighted data?

On the effective base, not the row count. Weighted means, variances and standard errors feed the test, because a weighted sample of a thousand does not carry a thousand rows of information and a test that pretends otherwise will report differences that are not there.

Can it clean the data, or only analyse it?

Twenty-five checks run over the file — speeders, flatliners, patterns, ranges, skip and gate routing, punch validity, grids, ranks, constant sums, duplicates, outliers, dates, codes, PII and open-end quality. They are written in a script with loops and conditionals, so a large grid is one loop rather than fifty pasted blocks, and a custom rule can run in the same pass.

Whose AI keys does it use, and can we see the cost?

Your tenant’s own keys, held in the platform key store and used at the moment of the call. Coding runs report their cost per job rather than aggregating into a monthly surprise, and the model calls are kept in an audit log. On client work that is a cost of sale, so it belongs on the job.

Can a client ask it questions directly?

Yes. A dashboard or the agent itself can be embedded in a client portal with its own key you can rotate and its own session-level analytics, so you can see what was asked. Dashboards also publish behind a share token with view and visitor tracking.

Does it handle qualitative work?

Focus group recordings are transcribed with speakers separated, assembled into a transcript and turned into reports — with a .sav export so the qualitative sits beside the quantitative, and a public link when a client needs to read it without an account.

Ask it the question you would have tabbed for.

Faster is only useful if you can still show your working when the client asks.

This is one module of the business OS, not a point tool. Research on its own does what is usually sold as five: a tabulation and statistics package, a data-validation tool, an open-end coding tool, a focus-group analysis tool and an insights-dashboard portal. The same seat opens Meetings, Mail, Chat, Drive, CRM, HR, Accounts, Projects, Procurement, Learning and Research — eleven modules covering about twenty-eight products' worth, on one login, one permission model and one ledger. You are not buying a survey tool; you are putting the company on one system.

See pricing Book a walkthrough →