An AI assistant that turns natural-language requests into decision-ready dashboards

A London-based B2B SaaS company partnered with EffectiveSoft to build a machine learning (ML)-based investment analytics platform for aviation assets.

The solution supports fleet evaluation, depreciation, and investment analysis, as well as forecasting and aircraft data management. It equips users with dashboards that provide up-to-date data and help them analyze large volumes of information through multiple graph types. Users work in a customizable workspace to compare scenarios and validate investment assumptions.

The platform handled complex investment analysis, but configuring dashboards took significant time: every chart and table that depended on multiple parameters had to be set manually. This made the platform cumbersome in daily use, delayed reviews, increased routine tasks, and slowed investment decisions.

As a solution, we first considered using business intelligence (BI) tools, but the platform required deeper customization than standard BI could support. Therefore, we proposed an AI assistant that helps users create custom dashboards from natural-language requests. Along with our partner, we agreed to validate the approach with a pilot.

Since charts and dashboards in the platform are defined by JSON settings, we used a Large Language Model (LLM) to generate these configurations automatically.

How it works in practice

The assistant takes a natural-language prompt and translates it into the JSON settings the platform expects. The platform then uses those settings to retrieve the required data and render the requested chart, table, or dashboard in the user’s workspace.

A key concern was that an LLM could generate probabilistic responses to valuation queries that are not reliable enough for investment reporting. To address this, we omitted the AI layer from calculations. The model and supporting agents either generate JSON configurations or instruct the platform to run a predefined backend function.

Under the hood, the assistant currently relies on two agents. One generates a JSON file needed to build charts, tables, and dashboards. It is triggered by a request like “build a dashboard to compare three aircraft by market value trend, depreciation, and forecast over five years.”

The other detects user intent. For example, it determines when a user’s request is about retrieving specific data (e.g., “Show me the valuation of the A320 with tail number G-XXXX as of Jan 2025”) and routes the request to a predefined backend function; the agent itself does not perform calculations.

This separation ensures the assistant consistently chooses the right action for each request.

To keep the generated output consistent, we described each JSON parameter and its purpose and added a set of few-shot examples to the prompt, so the assistant follows the same configuration pattern for similar requests. For example, when a user asks to “compare three aircraft by market value trend, depreciation, and a five-year forecast,” the assistant generates the same dashboard structure every time (the same chart types, required fields, and filter logic), instead of assembling a different layout on each run.

We also implemented two validation levels. The first level checks that all required fields are present and not empty, while the second performs logical validation to prevent mutually exclusive or incompatible attribute values (e.g., manufacturer: Boeing; aircraft family: Airbus A320).

Dashboards that used to take hours of manual setup are now assembled in minutes. Client usage analytics showed a reduction of around 60% in dashboard configuration time for complex, multi-asset scenarios. With the core functionality in place, the next step is to extend the assistant’s capability, enabling users to return with follow-up inputs, adjust dashboards iteratively, and work across multiple projects within a single workspace.