AI dashboard: built and kept current from your own data
An AI dashboard is a dashboard that AI builds from a plain-language description and keeps current from your own data, with a written note on what changed and alerts when a number moves outside its usual range. An AI dashboard generator can produce a first version in a few minutes. Keeping it accurate takes the same work as any dashboard, starting with a short list of numbers that each have a written definition and an owner. This guide covers how the generators work, KPI examples for businesses that make, move or sell goods, management reporting, what makes a good dashboard, and refresh and alerts.
What an AI dashboard is
A dashboard puts the few numbers a team watches on one screen. The UK government’s Analysis Function, which publishes guidance for government analysts, describes a dashboard as a visual tool that “is updated regularly with minimal effort” and presents data “without lots of commentary or analysis.”
AI helps in two places. It builds the first version from a plain-language description, which is what an AI dashboard generator does. After launch, it keeps the dashboard useful: a written note says what changed and why, and alerts tell the owner of a number when it moves outside its usual range. The written note fills a gap that the same guidance names, since dashboards “are not good at explaining what’s happening in the data.”
How dashboards get built today
Without AI, a dashboard is built one of two ways. Someone draws charts in a spreadsheet and updates them by hand from ERP exports, or a consultant builds a dashboard in a BI tool connected to the data. The first goes stale when the person who updates it is away. The second needs changes whenever the business changes, and the UK guidance warns that dashboards “require resource to maintain and improve.”
Both routes leave the reader to open the dashboard, scan every chart and work out what changed since last week.
AI dashboard generators
An AI dashboard generator turns a description into a working dashboard connected to your data. Amazon Quick Sight’s generate analysis feature is a clear example. You describe the dashboard you want and pick up to three datasets. Quick Sight shows “an interactive plan” of sheets and visuals to review, then builds them with filter controls and calculated fields such as year-over-year growth. AWS says generation takes 2 to 5 minutes, and it can start from an image of an existing dashboard, including one from another BI tool. Microsoft’s Copilot in Power BI creates and edits report pages from prompts in a similar way.
“Build a weekly operations dashboard for our plant: on-time delivery and late orders by customer, scrap rate by line, open order backlog in weeks, and gross margin by product family against the same week last year. Count sales as invoiced, net of credit notes. Show when the data was last refreshed.”
Generators are good at the first draft: choosing chart types, laying out pages and adding standard calculations. They cannot know your definitions unless you write them down, which is why the example prompt says how to count sales. A generated “revenue” tile might include freight charges or leave out credit notes, and it looks the same either way. Check each tile’s calculation against a figure you trust before anyone relies on it.
A generator works from the datasets you give it and nothing else. Point one at a raw ERP export, and it sees column names such as QTYSHP and guesses what they mean. Give it a prepared dataset with plain column names and the definitions alongside, and it has less to guess.
KPI examples for businesses that make, move or sell goods
A key performance indicator (KPI) is a number tied to a goal. The table lists useful ones by type of business, with the calculation and the data each needs. Pick the five to eight that match the decisions your team makes each week.
| KPI | Business | How it is calculated | Data it needs |
|---|---|---|---|
| On-time delivery | Manufacturer, distributor | Orders shipped on or before the promised date ÷ orders shipped | Promise dates on sales orders, ship dates |
| Backlog in weeks | Manufacturer | Value of open orders ÷ average weekly shipments | Open orders, shipment history |
| Scrap rate | Manufacturer | Quantity scrapped ÷ quantity produced | Production and scrap records |
| Fill rate | Distributor | Order lines shipped complete on the first shipment ÷ order lines | Order lines, shipments |
| Inventory turns | Distributor, manufacturer | Cost of goods sold over 12 months ÷ average inventory value | Cost of sales, monthly inventory snapshots |
| Dead stock | Distributor, building-supply chain | Value of items with no sales in the last 180 days | Item records, inventory, sales history |
| Yield | Food producer | Finished-goods weight ÷ raw-material weight | Production batches |
| Shelf life at shipment | Food producer | Days left before the best-before date on the day a lot ships | Lots, shipments |
| Gross margin by product line | Any | (Revenue − cost of goods sold) ÷ revenue | Invoices, credit notes, costs |
Each KPI needs an owner, the person who acts when it moves, and a comparison such as a target, last month or last year. Show each number beside its comparison, so the reader can see at once whether it is on track. Margin analysis covers cutting the margin figure by product, customer and mix, and AI for inventory management covers the stock figures.
Management reporting
Management reporting is the set of reports an owner and managers use to run the business, as distinct from the financial statements prepared for lenders and tax authorities. A monthly management pack for a business that makes or sells goods might cover:
- Revenue and gross margin against budget and last year, by product line and by customer group.
- Cash, receivables by age and payables coming due.
- Inventory value, turns and dead stock.
- Operations: on-time delivery, backlog and, in a plant, scrap.
AI can take over two parts of the pack. A management dashboard refreshed from the ERP and the books replaces the monthly assembly of numbers. AI then drafts the commentary: which lines moved against budget, by how much and what drove each variance, with links to the rows behind each figure. The controller edits the draft and signs off before it goes out. A drafted line might read, for example: “Gross margin for the quarter was 31.2% against a budget of 33.0%. Most of the gap came from the contractor customer group, where freight on small orders rose.” Each figure in that sentence should open the rows it came from. AI financial modeling covers the budget and forecast the pack compares against, and AI in finance covers the month-end work around it.
Turn your monthly pack into a live dashboard
Tell Derik which ERP and accounting system you run and which numbers your managers review each week. He will tell you what it takes to keep them current.
Start a conversationWhat makes a good dashboard
The UK Analysis Function’s guidance on building dashboards says to consider one only when there is a clear user need and data updates “are frequent and can be automated, with users likely to revisit the dashboard.” If the numbers change once a quarter, the guidance points to simpler formats, such as a short written report. When a dashboard is the right format, six rules keep it useful:
- Start from decisions. List the decisions the audience makes each week, and include a number only if it informs one of them.
- Keep it short. Five to eight numbers per audience. The same guidance warns that dashboards “can overwhelm users.”
- Define every number. Show what is counted, from which system and as of when, in a note or on hover.
- Compare. Show each number beside its target, last month or last year.
- Link to the rows. Every number should open the records behind it, so a manager can check a surprise before acting on it.
- Show the refresh time on every page.
Charts follow the usual honesty rules as well: bar charts start at zero, and charts placed side by side for comparison share a scale.
Refresh and alerts
A dashboard that imports its data shows that data as of its last refresh, and the refresh runs on a schedule. Power BI, for example, allows up to 8 scheduled refreshes a day on a Pro licence and up to 48 on Premium or Fabric capacity, and it pauses scheduled refresh after a period of inactivity. Match the schedule to the decision: a morning operations meeting needs an overnight refresh, and a shipping dock may need hourly figures.
Check the load as well as the numbers. In 2020, a file-size limit in the process that loaded lab results into Public Health England’s reporting dashboards left 15,841 positive COVID-19 cases out of the daily figures between September 25 and October 2. A check that compares row counts and control totals with the source after each refresh catches that kind of failure. It should alert the dashboard’s owner before anyone reads the numbers.
Alerts on the numbers themselves come in two kinds:
- Threshold alerts fire when a number crosses a limit you set, such as fill rate below 95%. Power BI’s data alerts work this way on dashboard tiles: they notify you “when data in your dashboards changes beyond the limits you set,” on gauges, KPIs and cards, at most once an hour or once a day.
- Anomaly alerts learn a number’s normal range and fire when a value falls outside it, which suits numbers with seasonal patterns. AI business intelligence covers anomaly detection in more detail.
Send each alert to the owner of the number, with a link to the rows that caused it.
A weekly written note complements the alerts. It says which numbers moved, by how much and what moved with them. The safe way to produce it is to compute the changes with statistics first and let the language model describe them, as AI business intelligence explains. Send it to the attendees before the meeting, with links to the rows behind each figure.
What data an AI dashboard needs
An AI dashboard reads the same data as any reporting: the ERP, the accounting system and the spreadsheets that hold budgets and targets, copied into a reporting database on a schedule. Trend lines also need history, so keep a nightly snapshot of inventory and open orders. Write the definitions down once and point every tile at them. AI reporting covers connecting these sources and checking the answers.
How to start small
- Pick one audience, such as the weekly operations meeting or the owner’s Monday review.
- List the decisions that audience makes and the five to eight numbers behind them.
- Define each number and name its owner.
- Generate a first draft, then check every tile against a figure you trust.
- Add the refresh time, one alert per number and a weekly written summary.
- Review after a month and remove the tiles nobody used.
Risks and limits
- Wrong definitions. A polished chart makes a wrong number more convincing. Check every tile at launch and after any change to its source.
- Stale data that looks current. Show the refresh time, and alert the owner when a load fails or brings in fewer rows than expected.
- Too many numbers and too many alerts. Both teach people to stop looking. Remove what nobody acts on.
- Access. A dashboard shows whatever its data allows. Apply each viewer’s permissions, and review who can see cost and margin figures.
- Where the data goes. Written summaries and generated dashboards send your data through an AI model. Data sovereignty and private AI for business cover where the data and the model can run.
How ThriveAI helps
ThriveAI is an AI engineering company in Ottawa that builds private AI systems on a company’s own data, for businesses that make, move or sell physical goods. It brings what your ERP, email, drawings and spreadsheets hold into one database that belongs to your company. Your team asks it questions in plain language, and each number links to the document or ERP record it came from. When the data cannot answer a question, the system says what is missing instead of guessing.
ThriveAI’s systems read your ERP as it is, including older versions installed on your own server, and every connection only reads data. The platform is designed to keep each client’s data on its own server in Canada. You choose a model on that server or a hosted model under a written zero data retention agreement, and a hosted model may process requests outside Canada. Derik Lawlis, the founder, leads every project and stays close to the build. About ThriveAI covers the company.