Technology
Agents
An agent here is a specialist that finishes a particular kind of work — not a persona, and not a wrapper around the same model with a different opening line.
There are thirteen of them, and most of the time you never pick one. Your request is read, the specialist that can actually finish it is chosen, and the reason is shown. Where two of them plausibly fit, you are offered the choice rather than guessed at.
They are not sealed off from each other either. Chart AI can send out for data, the Financial agent runs Python underneath, and any of them can become a step in a workflow that runs on a schedule with nobody watching.
You describe the work, not the tool
Nobody should have to know that a waterfall is Chart AI and a process map is QA. You say what you want; the routing is the platform's problem.
It says why, and why not
The near miss is the useful part. Being told a request went to QA rather than Chart AI because there was no measure to plot teaches you what to ask next time.
Ambiguity becomes a menu
Where more than one specialist genuinely fits, you are given the options rather than a silent decision — and one of them is usually recommended.
It learns the habit
When a desk reaches for the same specialist repeatedly, that preference is recorded and the asking stops — and it changes back when the habit does.
The specialists
Thirteen, and the line between them
Each card carries what the agent is for and — more usefully — what it is not, because the boundary is what tells you which one you actually wanted.
QA
Finds, explains and presents. Vector search across attached documents with citations, summaries on demand, web and market lookups, chat-history search, table lookups by value. Also renders structural diagrams — flowcharts, process maps, sequence and ER diagrams, org charts, mind maps, decision trees.
Not for charts of numbers, building new files, joining tables or calculating.Financial
Calculations, models and workbooks — either against an attached spreadsheet (formulas, totals, reconciliation, scenario and what-if, nested financial XML) or building one from scratch when all you have is a PDF: cash-flow models, leverage and coverage trajectories, debt schedules, sensitivity tables.
Not for looking a value up, joining tables without arithmetic, or writing prose.Table AI
Transforms tables across several files — stacking rows, joining on a key, or a simple model template — using semantic column mapping against your organisation's own dictionary. The output is a new combined workbook.
Not for displaying values from a table you already have, or for calculating.Chart AI
Builds a chart of numbers: bar, line, pie, scatter, bubble, radar, area, waterfall, funnel, histogram, or a choropleth map of a measure by region. Recommends the chart type, lets you pick the axes, and can pre-aggregate or pull external data first.
Not for structural diagrams — a flowchart or org chart with no measure goes to QA.Statistical AI
Statistics proper: correlation, regression, distributions, hypothesis tests, clustering, time series. Writes the Python, runs it, and returns a multi-sheet workbook with live formulas and embedded charts.
Not for plain sums, joining data, or a single chart.Portfolio AI
Aggregates portfolio and fund performance across many files into a summary sheet whose cells carry togglable SUM and AVG formulas over the source sheets, with those sheets kept alongside.
Not for a general union of data, statistics, or charting.Grid AI
An extraction grid over mixed documents and tables — one row per file, one column per field you name, and every cell carrying prose plus a typed citation: a page number for a document, an A1 range for a table. Follow-up questions the grid can answer are answered from the grid.
Not for a single document, transforming data, or calculating.Document AI
Writes a new structured document. It plans first — you approve the outline and the word budget — then section writers run in parallel, pulling from your files and the web, and assemble a document with tables, charts and citations. Also summarises, extracts and translates existing ones.
Not for retrieving a fact, or for transforming data.News AI
Builds a news widget from the news sources you have connected, with market-sentiment and trending commentary written for someone who has two minutes.
Not for a general news lookup with no source attached — that is QA and web search.Code Gen
Whole runnable projects rather than snippets: plans, writes, executes and repairs a multi-file application in a sandbox. Either a command-line program that runs to completion in Python, C++, Go, Rust, Java, Scala, Ruby, R or COBOL, or a React web app — optionally with a Python backend — served live in the chat.
Not for a repeatable multi-step process; that is a workflow.Python
The plan, generate, validate, fix and execute loop. Writes Python against your data, checks it before it runs, patches the line that broke rather than starting over, and executes it in a sandbox — on its own for analysis, and underneath the Financial agent when a model needs real computation.
Not for a whole application; that is Code Gen.SQL
Queries a database directly — for the questions that are a query rather than a document. It writes the statement, runs it against the source you have connected, and returns the result as something you can keep working on.
Not for a spreadsheet you have attached; that is Table AI or Financial.Workflow
Builds, edits and runs a multi-step automation: a trigger, data loaders, agent tasks, conditional branches, review gates and widget outputs, on demand or on a schedule. It proposes the plan and you approve it before anything is built. Also explores the organisation's knowledge graph.
Not for writing software, a one-off document, or a single analysis right now.Beyond the router
The ones you reach for directly
Not everything arrives through a chat message. These have their own entry points — a step inside a workflow, a comparison you set up deliberately, a microphone — rather than being chosen for you by the router.
Together
They call each other
A specialist that cannot ask for help is a dead end. Chart AI pulls external data through QA when the numbers are not in the file. The Financial agent runs the Python loop underneath when a model needs real computation rather than a formula. Grid AI hands a question back to QA when it goes beyond the columns it extracted.
The gates are shared too. Document AI has you approve the outline before it writes; the Workflow agent has you approve the plan before it builds; anything that runs code shows you the specification first. And every one of them is screened by the same guardrails, and lands in the same record.
One runs inside another
The Python loop is a full agent on its own and a component of the Financial agent — the same code, plan-checked and sandboxed either way.
Approval before the expensive part
Outlines, workflow plans and code specifications are put in front of a person before the work that costs time begins.
No agent is a side door
Every one of them reaches sources through your grants, is screened by your guardrails, and is written into the graph as it runs.
Turn weeks of expert work into minutes
See how Scalata fits your team, your data, and your controls.