CLIENT: Camarones Punta Mangle S.A.🇪🇨Ecuador

Brújula - AI Policy Assistant That Cites the Clause and Escalates the Risky Questions

Policy Assistant for an Ecuadorian Shrimp Exporter

Designed and built a clickable internal policy assistant for a Guayaquil shrimp farmer, processor and exporter: staff ask in Spanish or English and get answers built only from company policies, each statement linked to its clause and version, while gift approvals, conflicts and possible breaches go to Compliance, HR or the confidential Speak Up channel instead of being answered.

Scope of Work
Product design, UI build and clickable demo
Industry
LegalTech / AI Policy Assistant
AIEnterprise
Compliance & Certifications
LOPDP (Ecuador)SPDP AI resolution 2026COIP anti-briberyFCPAUK Bribery Act 2010Ecuador Labour CodeASC and BAP

Most of our work is under NDA. Per agreement, we can share the work but not the company details - all names and identifying information shown here have been changed.

Impact

5 Screens
Clickable end to end

Compliance dashboard, escalation queue, escalation detail with the work panel, the employee conversation view and the policy register with its policy drawer.

3 Paths
Answer, escalate or Speak Up

Every question ends as a cited answer, an escalation with a reference and reply target, or a confidential Speak Up handoff with no answer and no named team. The demo conversations show each one.

USD 100
Gift limit routed, not judged

Gifts above USD 100 from one giver in a year, and any meal or hospitality for a public official, go to Compliance for pre-approval. The assistant's draft is kept for the analyst and never sent.

18 Escalations
Realistic demo data

15 policies, 15 cited clauses, 18 escalations across Compliance, HR, Legal and Finance, 5 employee conversations, 16 employees across 7 sites and 10 staff users.

The Problem

Punta Mangle's policies lived as PDFs on SharePoint, several in two or three versions with no clear current one, and a compliance team of four answered the same questions by email and WhatsApp every week. Answers depended on who replied, nobody could show what had been told to whom, plant supervisors often never asked, and attestations collected on paper and in a spreadsheet could not be matched to policy versions when retail buyers asked for evidence in supplier audits.

Our Solution

Brújula answers staff policy questions in Spanish or English only from Punta Mangle's published policies, with every statement linked to the exact clause and version. When a question needs a decision, the assistant is unsure or the question suggests a possible breach, it sends the question to Compliance, HR, Legal or Finance with a reference and reply target, or hands the person to the confidential Speak Up channel, and the staff console gives those teams one queue to work it.

The Challenge: Explaining Policy Without Approving Anything

Punta Mangle has about 2,900 people across a Guayaquil head office, farms in Guayas and El Oro, a hatchery, two processing plants and a Miami sales subsidiary. An assistant for them has to explain the rules clearly while never deciding whether someone's situation is fine.

The Same Question, Different Answers

"Can I accept a basket from the feed supplier at Christmas?" was answered by whichever of four compliance staff replied, by email or WhatsApp, with no record. Plant and farm staff on shared kiosks or phones rarely asked at all.

No Current Version to Cite

Several policies existed in two or three versions on SharePoint. Two large retail buyers began asking for proof that staff had acknowledged the anti-bribery and code of conduct policies, and paper attestations could not be tied to a version.

Questions That Need a Decision

Ecuador's COIP criminalises bribery and can reach companies, and the FCPA or UK Bribery Act may apply through the US subsidiary and UK customers. Punta Mangle applies the stricter standard: facilitation payments banned outright and any gift to a public official pre-approved.

Reports That Must Stay Confidential

Ecuador generally has no comprehensive whistleblower statute for private employers, so protection comes from the company's own Speak Up policy. Questions can also reveal health data covered by the LOPDP, so neither kind may land in review queues or analytics.

Our Approach: From Policy Rules to a Clickable Assistant

01

Spec and Rules Research

Mapped 9 roles, from employee and plant worker to ethics officer and data protection lead, and 36 screens: 27 staff, 8 employee and 1 public Speak Up page. Defined the data model for policies, versions, clauses, conversations, answers with citations, escalations, Speak Up handoffs, escalation rules and attestation campaigns. Wrote down the rules the design follows: the LOPDP, the 2026 data protection superintendency rule on AI, COIP, FCPA, UK Bribery Act, the Labour Code and buyer certification requirements such as ASC and BAP.

02

Clickable Demo on a Production UI Kit

Built 5 key screens on a production dashboard kit: the dashboard with six KPIs, work by role and escalations by reason, the escalation queue with reply-target states, escalation detail with asker card, assistant draft, retrieved clauses and outcome, the employee conversation with source chips and clause preview, and the policy register. Assign, refer to Speak Up, escalate to a person, guided declaration and new policy modals all run on local state.

03

What a Production Build Adds

Microsoft Entra ID sign-in, a Microsoft Teams bot, SAP SuccessFactors for staff and sites, SharePoint for existing policy files, Claude on Amazon Bedrock with pgvector retrieval, document parsing into clauses and a Speak Up provider such as NAVEX EthicsPoint or EQS Integrity Line. Grounding and blocked-phrase checks on every answer, an append-only audit log, working-hours reply targets with the Ecuadorian holiday calendar, and the remaining screens: review queue, canonical answers, test questions, version upload, campaigns, analytics and settings.

Why It Works: The Assistant Explains, People Decide

Every Statement Points to a Clause

Answers are written as "the policy says" with numbered markers. Each source chip shows the policy, version, clause and effective date and opens the clause with the quoted span highlighted. When retrieval is weak, the answer says so and offers Escalate.

Decisions Arrive Complete

Guided declarations collect the giver, relationship, value and whether a public official is involved, so the analyst sees the full case, the rule that fired and the assistant's unsent draft, then records an outcome with conditions and an expiry.

Speak Up Stays Out of Sight

A possible breach gets no answer and no named team: the card offers Open Speak Up, the phone line and the non-retaliation commitment. The conversation shows as Private in the person's history, and the dashboard counts possible breaches without showing content.

Sensitive Questions Show a Label Only

Health, pregnancy and disciplinary questions appear in the queue as a label, not the text, and HR cases are visible only to HR partners scoped to the asker's site. Exports are designed to leave question text out.

Related case studies

View all: LegalTech
Logo
Oceanlab

Enterprise Software Engineering - Human + AI

Languages

Locations

Global Headquarters

21 Merchant Road, 3rd Floor, Singapore 058267

Main corporate office and leadership operations

North America Office

548 Market Street, PMB 12345, San Francisco, CA 94104

Client operations and business development

India Office

402 Shivalik Shilp II, SG Highway, Ahmedabad 380015

Engineering and delivery center

© 2026 - Copyright. All Rights reserved Oceanlab