cutting data enrichment costs by 97% with a custom AI tool

long story short

Outbound sales at scale is a data problem before it's a pipeline problem. The moment your team starts working on the contact lists for a new outreach campaign, the stakes around data quality become very real. Every message you send should feel like it was written for that one person, which means you need to be confident they still work where you think they work, hold the title you think they hold, and that your email will actually reach them. At scale, verifying and enriching that information manually becomes too slow, and sometimes even impossible. The question shifts from "who do we reach out to" to "how do we know enough about each contact to make that outreach worth anything?

For a long time, the answer was purpose-built enrichment platforms: import your CSV, let the tool fill in the blanks, export the result. Useful in theory, but expensive in practice. Especially when the same underlying work – finding a company website, pulling a LinkedIn description, categorizing a contact by industry segment – is something a well-prompted AI model can do for a fraction of the cost.

Our CTO saw this gap firsthand while working on our sales processes optimization. The existing enrichment tools worked, but they were burning through the credits available in our subscription plan at a rate that didn't scale. So he built something better – an AI-powered enrichment tool tailored specifically to how the Rebbix sales and operations team works.

the problem

Commercial enrichment tools operate on a model built for an older era of data work. Their value proposition rests on access to proprietary databases: you provide a name or a LinkedIn URL, they return an email, a phone number, a job title. That data had to be bought, maintained, and sold back to users at a premium, typically through monthly subscription tiers with credit limits that reset on a schedule or can be added on demand.

The issue is that the underlying logic of that model has shifted. 

"The rules of the game have changed. The data about your leads has always been publicly available, so the problem was never access – it was time. Googling each lead manually to find their current role, company website, or LinkedIn profile would take forever. That's exactly what these platforms built their business on: charging you for convenience. The question is whether you create tooling that lets you use that cheaply and at scale, or keep paying platform rates for something you could do yourself with the help of AI." - Taras Kunch, CTO at Rebbix

So, how do you run AI-assisted enrichment across thousands of rows without writing a new script every time?

what we built

Our CTO answered that question by building a web-based smart enrichment tool that uses LLMs to find and structure any available information about a lead into a clean, exportable table. A user imports a CSV with required contacts, the tool renders it as an editable table, and from there, we can add new columns powered by AI prompts.

The interface is intentionally simple. A user writes a prompt that references any columns in their table as inputs and runs it across their dataset. The model fills in values row by row, in the background, on the server, so closing the browser tab doesn't interrupt anything.

Core capabilities include:

• AI-powered column generation – lets a user write a natural-language prompt referencing any existing columns (first name, company name, LinkedIn URL, etc.) and generate new values across the entire dataset. Supports both OpenAI and Anthropic models.
• Prompt enhancement – a built-in "Enhance" function rewrites user-submitted prompts to improve model compatibility and output consistency, reducing hallucinations without requiring prompt engineering expertise.
• Data merging – allows a user to import a second CSV and merge it into the existing table by matching on a shared column ID. This is particularly useful when working with CRM exports that split contact and company data across separate files (a known limitation in similar out-of-the-box tools).
• Targeted prompt re-runs – lets a user apply a prompt to a subset of rows first (e.g., 10 rows) before committing to a full run, making it easy to catch prompt issues before they propagate across thousands of records.
• Column formatting – column types (text, URL, etc.) can be set to control how values render and behave. URL-typed columns open as clickable links directly from the table.
• Filtered exports – a user can hide columns or filter rows to prepare a clean output view, then download only what's needed. Useful for feeding specific subsets directly into outreach tools or CRM import flows.
• Change highlighting – cells updated in the most recent processing round are visually flagged, making it easy to spot what changed without reviewing the full dataset.

‍

example of our workflow

The primary use case of this tool at Rebbix is outbound campaign preparation. A typical workflow looks like this:

‍

1. We export a raw contact list from Sales Navigator, HubSpot etc. – name, current company, position, LinkedIn URL, email.
2. Import it into the enrichment tool.
3. Add a column to find the company's website (inferring from the email domain when available, falling back to a web search using the company name).
4. Add a column to generate a 100-200-character company description using the website's About section or LinkedIn company page.
5. Add a column to categorize each contact into one of Rebbix's internal industry segments, using the description and company name as input.
6. Filter and export the contacts that match relevant segments.
7. Load the result into the outreach tool.

‍

The tool also supports data hygiene tasks. For example, checking whether a lead's LinkedIn profile still shows them at the same company listed in the CRM, and flagging records where that's no longer the case.

The tool is model-agnostic by design. GPT-family models include built-in web search through the API, making them the default choice for tasks that require live lookups. Anthropic models are available for tasks where web access isn't necessary – structured categorization, format normalization, or logic-based inference from data already in the table.

beyond sales: other applications

This smart enrichment tool was built to solve a specific task in Rebbix's sales operations, but the underlying logic – running AI-powered processing across structured tabular data, row by row, at low cost – applies across a wide range of use cases.

Recruitment and talent sourcing
Recruiters working with large candidate lists can use the tool to pull current role and company information from LinkedIn profiles, generate brief candidate summaries, or score applicants against a set of predefined criteria without manual research per profile.

Market and competitive research
Given a list of companies (competitors, potential partners, industry players), the tool can generate brief descriptions, assign market segment tags, find website domains, and flag companies that match specific criteria, turning a raw list into a structured research dataset.

Marketplace and catalog operations
For platforms managing large product or vendor catalogs, the tool can enrich listings with standardized descriptions, assign categories based on product names or existing metadata, validate URLs, and flag incomplete entries.

Event and conference preparation
Ahead of industry events, teams can take an attendee or speaker list and enrich it with company context, identify which contacts match target profiles, and prioritize outreach.

Data normalization across sources
Any situation where data comes in from multiple sources with inconsistent formatting (names in different languages, varied date formats, mixed casing) benefits from the same prompt-based normalization approach, without needing a developer to write and maintain a script for each case.

tools and technologies

• Node.js
• Firebase Hosting + Firestore
• GPT models (OpenAI API, with web search)
• Anthropic models (Claude API)
• React (frontend interface)

‍

Interested in automating your repetitive processes with AI?

‍