Your AI data scientist — and it shows its work.
InsightAI finds what's driving your costs, revenue and delays, tells you how sure it is, and puts a dollar range on it when the evidence supports one — then shows every step in a plain-English audit. Real statistical tests, not just a chatbot. Your files stay on your computer.
Evaluating it for a team? Talk to us and we'll walk you through it on your own data.
Illustrative example
Most "AI for your data" tools sound confident. That's the problem.
They'll happily turn a coincidence into a "trend" or a fluke into a "finding" — and you'd never know. When you're making real decisions about money, suppliers and schedules, confidently wrong is worse than no answer at all.
InsightAI is built the opposite way. Every finding gets one of four ratings — Confirmed, Robust, Suggestive or a hypothesis to watch — based on how well the evidence holds up, and the written answer is checked so it doesn't claim a stronger rating than the finding got. The report shows every test behind each finding and its result, so you can see for yourself.
A good tool confirms a figure really is in your data. That's necessary — but it isn't enough.
Is the pattern real, or a fluke of small data, the luckiest of many comparisons, or a column that leaks the answer? InsightAI tests the conclusion — not just the number behind it.
Three steps from your exports to a decision.
Bring the data you already have
Exports from your database, ERP or BI tool as CSV or .xlsx, or the spreadsheets you already keep — even years of exports whose columns changed along the way. InsightAI reconciles them, cleans them, and logs every decision. No pipeline to build, no data team required. Direct database connections?
Ask in plain English
"What's driving our costs?" "What kinds of jobs tend to run over budget?" "What will next quarter look like?" No SQL. No code.
Get an answer — and the evidence
A clear answer with a confidence rating, a dollar range when the evidence supports one, and a "show me why" trail behind every finding. Figures in the answer are color-coded: green matches a figure the calculations produced or used, amber is rounded from one, and red is one it couldn't trace — often just a looser rounding — so check it before you quote it.
Four things built in from day one.
Validated, not just verified
Findings are tested against chance, sized with a range, and checked for whether there's enough data behind them — then rated Confirmed, Robust, Suggestive, or a hypothesis to watch.
Actionable
Finds the driver, puts a dollar range on it from cost and volume assumptions it spells out, and lays out the options — like a data scientist on your team, not a search box for your spreadsheet.
Transparent
Every finding traces to a plain-English "show me why" report, built by fixed code from the recorded steps, so the tests behind it and their results are there to check. Glass-box models only.
Private by design
Your files stay on your computer. The AI works from the analysis — names, statistics and results — and every result it's given is in the audit you can download.
| Aspect | Typical AI data tools | InsightAI |
|---|---|---|
| What it checks | That numbers match the source data | The number — and whether the conclusion holds up statistically |
| How sure it is | Usually unstated | Every finding rated: Confirmed, Robust, Suggestive or Hypothesis |
| Models & forecasts | Can fit a model or a forecast; how well it's tested varies | Driver discovery, forecasting and glass-box models, each tested against chance or a simple baseline |
| Output | Charts, queries, a chat answer | An answer and charts, a rating, the evidence, and a dollar range when your costs support one |
| Built for | Often data teams and analysts | The people who make the call — not just the data team |
| Audit trail | Technical computation traces | A plain-English "show me why", built by fixed code from the recorded steps |
| Your data | Often uploaded to the vendor's cloud | Files stay on your computer; the AI gets analysis results, not your files |
Every answer comes with the receipt.
Open "show me why" and you get the whole chain in plain English: where the data came from, every cleaning decision, how each conclusion was tested, and how sure the system is — including why it held a finding back.
Every step it runs is recorded by code as it happens, with its test results, and the report is laid out by fixed code from that record. Each finding's summary restates figures from those steps, and you can check them against the step itself in the same report. It's the difference between trust me and here's how I got there — something you can walk your board, your finance team or your quality lead through.
- Confirmed solid enough to act on
- Robust holds up
- Suggestive worth a look
- Hypothesis watch it
A rating says how sure InsightAI is that a pattern is real in your data, not what causes it. When only a few groups were compared, the report adds a caveat that something tied to that group could be behind it.
unit_cost — read it as currency amounts
Strongest: Supplier C, +$4.10 per unit
Whole search tested with 2,000 random reshuffles — it survived
Range [$3.48, $4.78] excludes zero · stable in 100% of resamples
Confidence based on 1,212 purchases behind the effect
Estimated impact: $1.9M/year (range $1.6M–$2.2M)
assumes an affected volume of 463,000 units a year
correlational, not causal
Caveat: only 6 suppliers compared, so the gap is real in this data but could come from something tied to Supplier C (its product mix, a plant, a region). Worth checking before acting on the cause.
It does the hard analysis — and keeps itself honest.
Beyond charts and summaries, InsightAI finds what's driving your outcomes, forecasts what's coming, and builds predictive models you can actually read.
The difference is the guardrails: it cross-checks its own work the way a careful analyst would — and says "I can't be sure" when it can't be sure.
Driver discovery
Pinpoints the supplier, product line, site or route that stands out on your outcome, tests it against chance, and puts a dollar range on the gap.
Predictive models you can read
Finds what predicts an outcome your data records — a job that ran over budget, a delivery marked late — and shows which factors matter most. When the simplest model wins, you also see each factor's direction with error bars; for a yes/no outcome you may get plain IF/THEN rules. If nothing beats a simple guess, it says so.
Forecasting
Projects forward with honest ranges. Given enough daily history, it can split out your weekly rhythm, your yearly season (with two years of data) and big holidays such as Thanksgiving and Christmas — and it tells you when your history is too short to forecast reliably.
Anomalies and shifts
Flags outliers and sudden shifts in your history — when they started and how big they were — as leads to check, not conclusions.
It won't over-claim
Screens for columns that leak the answer, the luckiest-of-many-comparisons trap, and samples too small to trust — and says so plainly.
Your files stay on your computer. The AI works from the analysis, not your files.
InsightAI does the number-crunching on your own machine, and the AI directs it from the results.
What reaches the AI (Anthropic's Claude, on your own API key) is your questions, file, sheet and column names, statistics and calculation results. Some results include individual values, such as a maximum or minimum, the category labels in a column it examines (a product line or supplier — or people's names, if a column holds them), up to ten flagged outliers, and the period-by-period values behind a trend or a forecast. If your data has one row per day or month, those values are your data. Every result the AI received is in the audit, which you can download while the analysis is open.
Your API key is kept in Windows Credential Manager and sent only to Anthropic. The only other thing the app sends is a sign-in and license check, which carries none of your data. Full details
The whole analysis inside a sealed sandbox with no network access at all. It's built and proven in testing; it isn't in the installer yet. We're also exploring a fully on-device AI option for the most sensitive environments.
It traced a $1.3M-a-year problem to a source the weld logs never recorded.
Rebuilt from a real investigation, on synthetic data: eleven years of weld records, and the recall invoices that followed, in 250 messy files — weld logs in three different layouts, plus monthly shift schedules. Misalignment jumped in mid-2009, and nothing in the weld logs said why. Here is what InsightAI did.
- Reconciled 111 weld logs across three record layouts into one table, even though they never said who did the work. → 135,000 welds
- Ruled out the machine settings first: voltage, current and force showed no real relationship to misalignment. → not the machines
- Joined the welds to the shift schedule and found the source only showed up when the data was sliced by who was on shift. → one schedule assignment stands out
- Tested the whole search against 2,000 random reshuffles; the gap beat every one. → survived · Confirmed
- Checked it against the recall records, an independent source, which pointed the same way. → agrees
- Put a dollar range on it from the average recall cost in the invoices and how often recalls happen. → ~$1.3M a year ($1.1M–$1.5M)
InsightAI points to where to look; your quality team decides what it means.
Synthetic data, generated to recreate a real investigation. Figures are from InsightAI's analysis engine run on this 250-file archive, assuming about $225,000 per recalled section (the average in its recall invoices). A live chat run's dollar figure depends on the cost and volume assumptions it's given.
Built for the people who make the call.
Whether you don't have a data team, or you're waiting in line for yours.
Operations, quality & finance leaders
Get a defensible answer about costs, delays, defects or margin without joining the analytics queue — and a plain-English trail your leadership can follow.
Owners & operators
Find what's driving your numbers and make confident decisions — without hiring a data team or learning SQL.
Consultants & advisors
Deliver auditable analysis to clients, faster. The plain-English trail lets any stakeholder see exactly how you got there.
Common questions it answers
- What's driving our costs, delays or defects?
- Which suppliers, product lines or sites matter most?
- What predicts a late delivery or an over-budget job?
- What will next quarter look like?
- Where are the anomalies and sudden shifts?
- What is this problem costing us?
- Can I walk my board through the evidence?
Pricing, coming soon.
Pricing isn't set yet. We'll agree it directly with the teams in early access.
Be first in the door.
Early-access members get priority onboarding and a direct line into the roadmap. During the alpha, AI usage is billed by Anthropic to your own API key — yours or your company's.
Get early accessEvaluating it for a team? Email info@transparentai.tech and we'll walk you through it on your own data, on your own computer.
Questions we get a lot.
Is it only for small businesses?
No. It works the same way for a 20-person contractor and for an operations team inside a large manufacturer. What matters is whether you have a data scientist on hand when a question comes up — most teams don't, or they're waiting in line for one.
What kinds of files does it work with?
CSV and Excel files (.xlsx, and older Excel 97–2003 .xls workbooks) — including piles of messy exports whose column names changed over the years. InsightAI reconciles and cleans them and logs every decision. Workbooks with several sheets are read up to the first 20,000 rows of each sheet; for bigger tables, export them as CSV and they're read in full on your computer. Very large tables take longer to analyze and are limited by your computer's memory; if yours run to millions of rows, email us first.
It analyzes the columns already in your files. It doesn't yet combine columns into a new measure for you (margin from price and cost, say, or days late from a promised and a delivered date), so add that column to your spreadsheet or export first.
Can it connect to our database?
Not yet. Today it works from exports: pull a table from your database, ERP or BI tool as CSV or .xlsx and drop it in.
Read-only database connections are on our roadmap, designed around two rules: they'll use a read-only account that your database administrator controls, and the AI will never write the query itself. It will say what it needs, and fixed code will build and run the query.
How is this different from asking ChatGPT or Copilot?
General-purpose assistants like ChatGPT and Copilot can analyze a file you give them, but they aren't built around checking whether a finding holds up. InsightAI is. Claude reads your question and explains the answer, and fixed code does the math: it tests findings against chance, puts a range on its estimates, and gives each finding a rating on a four-step scale, with the tests behind it in the report. The number-crunching runs on your computer.
Does my data leave my computer?
Your files don't. The AI (Anthropic's Claude) receives your questions, file and column names, statistics and calculation results — and some results include individual values, such as a maximum, a category label it needs to name a result, or the values behind a trend or forecast. Every result it received is in the audit, which you can download while the analysis is open; it isn't kept after you close the app. Apart from that, the app only signs you in and checks your license, which carries none of your data. Details are on our privacy page.
Which AI does it use, and what does that cost?
Anthropic's Claude, through an Anthropic API key — yours, or one your company issues from its own Anthropic account. An API key is a password-like code that lets InsightAI use Claude on that account: you create it on Anthropic's website, paste it into InsightAI once, and Anthropic bills the account directly for the AI usage. We never see the key or the bill. The key is stored in Windows Credential Manager and sent only to Anthropic, and we'll walk you through it when you get your invite.
It needs a key from Anthropic's own API: Claude through AWS Bedrock, Google Vertex or Microsoft Azure isn't supported today, and neither are Copilot, ChatGPT or other models. If your company's IT policy requires a particular provider, tell us — we want to hear it.
What does "validated" actually mean?
The finding was tested, not just calculated. For a finding like "Supplier C costs more", InsightAI tests the whole search at once against 2,000 random reshuffles — a check designed so the luckiest of many comparisons doesn't pass for a finding — puts a range on the effect, checks that it's stable, and checks there's enough data behind it. Predictive models are cross-validated against a simple baseline and screened for columns that leak the answer. The "show me why" report shows each test and its result, including when there was too little data to confirm anything.
What won't it do?
It's built not to claim cause and effect: its findings are patterns in your history, and the "show me why" report marks its dollar estimates as correlational, not causal. It supports decisions rather than making them. If your data identifies individual employees, it can single one out; treat that as evidence for a person to review, not a rating, and check with your data-protection lead (in Germany, also your works council) before analyzing it.
Its predictive models use numeric inputs; categories such as carrier or supplier are tested by driver discovery instead. And today it doesn't save a model to score new records, connect to live systems, run on a Mac, or use AI models other than Claude.
Do I need a data background to use it?
No. You ask questions in plain English and get answers in plain English — along with an audit you can hand to someone technical if they want to check the method. The "show me why" trail is written to be readable by a business owner, not just a statistician.
When will it be available?
It's in private alpha on Windows now, by invitation. Join the list and we'll reach out about an invite. macOS is on the roadmap.
Be first to try it.
Join the list for InsightAI — priority onboarding and a direct line into the roadmap.
To join, email info@transparentai.tech.
Need your IT or security team to review it? We'll walk them through what the AI receives and what stays on your computer — info@transparentai.tech.