How the AI works
Why a model at all
A rule engine can detect that a column called “Date” is a date. It cannot detect that a column called “Signed” or “Week of” or “FY Q2” is also a date, and it cannot decide whether the right chart for that date paired with a column called “MRR” is a line, a bar, or a table. That judgement lives in having read thousands of similar decisions, not in a regular expression.
The second part is harder. Writing “revenue was up 18% in March, driven by the enterprise segment which added three accounts” from a table of numbers, in two sentences, in plain language, is a writing problem. Fixed templates produce sentences that are technically true and completely useless. That is the part customers pay for and it is the part only a language model does.
So the model is the engine. Everything else in this product is plumbing around it: parsing a spreadsheet a person made by hand, running the aggregations in code, and refusing to show you a number the model made up.
The pipeline
Before anything else, a fast cheap model looks at the column names and a sample of values to classify the dataset: sales data, survey responses, time-series metrics, inventory, and so on. This triage takes under a second and sets the context for everything that follows. It is the cheapest step because it does not need to be deep, only fast and directional.
The column profile, which is the types, ranges and cardinalities but not your raw rows, goes to a balanced model with a system prompt that names the chart types Plotwell can draw and the rules for choosing between them. The model is told to pick the views that answer a question a reasonable person would ask of this data, and to skip the ones that would just fill space.
The aggregated output, the actual sums and groups computed from your rows in code, goes to a model with a prompt that asks for two to four sentences per chart. Every figure the model writes is checked against the computed output before you see it. A figure that does not appear in the data is dropped and the count of dropped figures is reported. The model is told to say nothing if there is nothing to say.
A short title and a one-paragraph summary go at the top of every dashboard. The title names the dataset and the period. The summary names the one or two things that matter most. This is a separate routed step because the voice is different from the takeaways and the prompt is different from the chart selection.
Routing
This is not a diagram somebody drew. It is rendered from the same routing table the pipeline reads, so if it is wrong here it is wrong in production.
| Step | Model | Provider | Tier | USD per M tokens |
|---|---|---|---|---|
| Read the columns and choose the charts that fit | gpt-4.1-mini | openai | balanced | $1.60 |
| Write the plain-language takeaways | gpt-4.1-mini | openai | balanced | $1.60 |
| Name the dashboard and write its summary | gpt-4.1-mini | openai | balanced | $1.60 |
| Work out what kind of dataset this is | gpt-4.1-nano | openai | fast | $0.40 |
Wide files, where the column profile is large enough that the right chart is not obvious, escalate to the frontier model. The threshold is configurable by environment variable. Files up to 50,000 rows are supported on Pro and Team, and the escalation point adjusts for files at that scale.
If the routed model fails or returns nothing, the call walks a chain of candidates from other tiers and other labs before giving up. One provider having a bad afternoon should not cost a user their dashboard, and it does not.
Every number in a takeaway is checked against the aggregated output before it is shown. A figure that does not appear in the computed data is dropped and counted. This is the only part of the pipeline the model does not get a vote on.
Candidates
Every model below sits behind the same interface. Adding a lab is one case in one file, which is the whole point of building it this way.
openai · 1,000,000 token context
openai · 1,000,000 token context
openai · 1,000,000 token context
anthropic · 200,000 token context
google · 1,000,000 token context
meta · 128,000 token context
In this deployment the OpenAI and Anthropic adapters are written and the OpenAI one is live. The rest are declared with their real model names and switch on when their key is present. No model is trained on customer data, by Plotwell or by our providers.
The grounding guarantee
Charts are not drawn by the model. The model chooses which columns to chart and which chart type fits. The aggregation runs in code against your real rows. The model never produces a number from memory or training data.
Takeaways are checked before you see them. Every figure a takeaway cites is verified against the computed aggregation output. A figure that does not appear in that output is dropped, not corrected. Plotwell would rather say less than say something wrong.
Ungrounded figures are counted and reported. If a model run drops two figures from a set of takeaways, the count is shown so you know the takeaways you received are the ones that passed, not the full raw output.
Files up to 50,000 rows are supported on Pro and Team. The column profile Plotwell sends to the model is compact by design, which is what keeps large files fast and keeps your raw rows from traveling unnecessarily.