Skip to content
Marcelo Noriega
Book a call(opens in a new tab)

02 · FinTech · B2B marketplace · operations and AI

6 min read

Taxfyle

Completed

A tax services marketplace connecting clients with CPAs and Enrolled Agents, where I led automation, AI and go-to-market work.

Frameworks

TL;DR

Problem
Sales, Ops and Accounting ran on manual data processing, and vendor contracts had no shared SLA criteria.
What I did
Built 30+ agentic automations, overhauled vendor management, shipped a source-cited RAG assistant and took TXF to market.
Result
$350K+ in vendor savings, 45% less manual processing and 500+ TXF users.

01 · Overview

Five workstreams across operations, AI and go to market in a two-sided tax marketplace.

  • Sales automation, vendor management, a RAG assistant, the TXF launch and web scraping for leads.
  • I worked across Sales, Ops, Accounting, Customer Success and Pro Relations.
Company
Taxfyle, a US tax services marketplace
My role
Business Process Analyst & Product Operations
Timeline
Mar 2024 to Sep 2025 · Miami, USA (remote)
Product
A tax services marketplace connecting clients with CPAs and Enrolled Agents, where I led automation, AI and go-to-market work.
Team
Sales, Ops, Accounting, Customer Success and Pro Relations, with engineering and design

02 · Problem or challenge

Manual work, costly contracts and slow answers limited how much the marketplace could handle.

  • Manual data processing slowed Sales, Ops and Accounting.
  • Vendor contracts had no shared SLA criteria to judge or renegotiate them.
  • EAs and CPAs reviewed and corrected tax documents by hand, which capped how much work they could take on.

03 · Analysis

Error analysis found the real reason people did not trust the assistant.

  • Early RAG answers had no traceability: users could not tell where an answer came from.
  • Customer Success and Pro Relations field questions from about 200 clients and contacts a month.
  • Web scraping identified and qualified potential clients as the first step of the sales flow.

04 · Framework applied

Trust and data control came before autonomy.

  • Error analysis turned reading real failures into one fix: source attribution.
  • A trust ladder sets how much the assistant may do alone: suggest, answer with review, then answer alone.
  • For TXF, positioning framed it as an early mover in automating document review.

Trust ladder

Agency is earned one rung at a time
Text alternative
Level 1, suggest: the assistant drafts and a person decides. Level 2, answer with review: the assistant answers with its source and a person checks it. Level 3, answer alone: only for question types that pass evals consistently.

05 · Solution

Five workstreams, each with its own build and its own impact.

  • Select a workstream to see the problem, what I built, the tools and the result.

Sales automation

Problem

Pipeline stages and prospect contacts were tracked by hand across Sales and Ops.

What I did

  • Built 30+ agentic sales and ops automations with Claude Code, Cursor, n8n and HubSpot.
  • Automated client stage tracking across the sales pipeline in HubSpot.
  • Captured contact information for potential clients to feed engagement sends.
  • Re-architected processes across Sales, Ops and Accounting.

Tools

  • Claude Code
  • Cursor
  • n8n
  • HubSpot

Impact

Agentic automations. Adopted by 4 departments.
30+
Less manual processing. Across Sales, Ops and Accounting.
45%

Sales flow · workstreams a and b

Scraping sources the prospects that the automations stage and engage
Text alternative
Step 1, sourcing through web scraping, workstream b. Step 2, enrichment and contact capture. Step 3, pipeline stage in HubSpot. Step 4, engagement send. Step 5, follow-up. Steps 2 to 5 are the automations in workstream a.

Vendor management

Problem

Vendor contracts had no shared SLA criteria to judge or renegotiate them.

What I did

  • Led the vendor management overhaul, setting SLA criteria and renegotiating contracts.

Impact

Vendor savings. Renegotiated against new SLA criteria.
$350K+

RAG assistant

Problem

Customer Success and Pro Relations answer questions from about 200 clients and contacts a month.

What I did

  • A separate project, not one of the 30+ automations.
  • Answers questions for the Customer Success and Pro Relations teams.
  • Runs on n8n and Ollama, a locally hosted LLM, so company data stays in-house. Answers are validated against the company database.
  • Early answers had no traceability, so I added source attribution: every answer shows where it came from.

Tools

  • n8n
  • Ollama
  • Company database

Frameworks

Impact

Contacts served a month. Clients and Pro Relations contacts whose questions it answers.
~200
Saved per person per day. For each internal team member who uses it.
~30 min

The two numbers measure different groups, so they are never multiplied into a total.

RAG pipeline

Every answer is validated and cites its source
Text alternative
Step 1, a question. Step 2, retrieval from internal documents. Step 3, a local LLM through Ollama, orchestrated with n8n. Step 4, validation against the company database. Step 5, an answer with a cited source.

TXF go to market

Launched

Problem

EAs and CPAs review and correct tax data and documents by hand, which caps how much work they can take on.

What I did

  • TXF automates the processing of tax data and documents that EAs (Enrolled Agents) and CPAs would otherwise review and correct by hand, so they can take on more work.
  • Positioned as an early mover in automating this review work for tax professionals.
  • Supply-side go to market for a two-sided marketplace.

Frameworks

Impact

Users. EAs and CPAs in the Taxfyle network.
500+

Web scraping for leads

Problem

The sales flow needed a steady way to identify and qualify potential clients.

What I did

  • Identified and qualified potential clients. This is the sourcing step of the sales flow in (a).

Sales flow · workstreams a and b

Scraping sources the prospects that the automations stage and engage
Text alternative
Step 1, sourcing through web scraping, workstream b. Step 2, enrichment and contact capture. Step 3, pipeline stage in HubSpot. Step 4, engagement send. Step 5, follow-up. Steps 2 to 5 are the automations in workstream a.

06 · Product and project management

Specs written with the builders, quality owned through release.

  • I authored PRDs and acceptance criteria with engineering and design, then owned QA, bug triage and release validation.
  • I led the vendor overhaul: SLA criteria first, then renegotiated contracts.
  • I built the business case and finance planning for an internal invoicing and vendor expense product.

07 · Launch

Two rollouts: TXF to the network, automations to four departments.

  • TXF launched to EAs and CPAs in the Taxfyle network as supply-side go to market for a two-sided marketplace.
  • It was positioned as an early mover in automating tax document review.
  • The 30+ automations were adopted by 4 departments.

08 · Validation

Answers are checked against the source of truth before anyone relies on them.

  • RAG answers are validated against the company database and cite their source.
  • I owned QA, bug triage and release validation, checking each release against its acceptance criteria.

09 · Analytics and impact

$350K+ saved, 45% less manual work, 500+ TXF users.

  • 30+ automations adopted by 4 departments.
  • About 30 minutes saved per internal team member per day with the RAG assistant, which serves about 200 contacts a month.
  • The two RAG numbers measure different groups, so they are never multiplied.

  • Agentic automations · Sales automation

    30+

  • Less manual processing · Sales automation

    45%

  • Vendor savings · Vendor management

    $350K+

  • Contacts served a month · RAG assistant

    ~200

  • Saved per person per day · RAG assistant

    ~30 min

  • Users · TXF go to market

    500+

How I would replicate it at your company

  1. Map the manual steps in one pipeline end to end before automating any of them.
  2. Set SLA criteria first, then renegotiate every vendor against them.
  3. Read real AI failures before tuning prompts; fix the most frequent error type first.
  4. Give AI more autonomy only as evals and users show it can be trusted.

Playbook in action

The playbooks behind this case

  • New Product Development

    TXF launch

    • Opportunity identification and selection: Target: EAs and CPAs in the Taxfyle network, who review tax documents by hand. Supply side first.
    • Concept generation: Concept: automate the tax data and document review that EAs and CPAs did by hand.
    See the new product playbook

  • Feature Development

    RAG assistant and automations

    • Frame: Customer Success and Pro Relations answer questions from about 200 clients and contacts a month.
    • Specify: Authored PRDs and acceptance criteria with engineering and design.
    See the feature playbook

  • Bugs and Incidents

    Bugs and quality

    • Triage: Owned bug triage and release validation.
    • Diagnose: Error analysis showed early answers had no traceability to a source.
    See the bugs and incidents playbook

Let’s build something people actually need.

Tell me who you are building for. A short call is enough to see where I can help.

Book a call(opens in a new tab)Download resume (PDF)