AI Product Design & Innovation · 15-minute pitch

NEWS-GUARD.AI

An AI assistant for the business & finance newsroom: from market signal to a verified story file.

● Evidence-first AI assistant Chu Hoài Linh · VTV1 — Business & Finance News Reporter
Photo: Television Studio Control Room, Wikimedia Commons, CC BY 2.0
01 · Executive summary

AI doesn't write the news.
It helps prepare stories — with sources.

Decision authority increases →

AI

finds · organizes · summarizes · flags

Reporter

verifies sources & context

Editor

approves · accountable

Problem

Sources, figures and notes scattered across websites, PDFs, email and chat. Under deadline, cross-checking slows or slips.

Solution

A Research Case with Source Cards, a Claim–Evidence Map, Conflict Check, Story Brief and an Editorial Approval Gate.

Pilot

MVP on public-source stories, 6–8 weeks, human-in-the-loop, measured on time, traceability, conflicts and adoption.

02 · Business context

High speed.
High stakes.

Finance news must be fast — but fast can't replace accurate and accountable.

Diverse sources

Regulators, company filings, macro data, experts, wires, social media.

Complex data

One number changes meaning by period, unit, scope or revision.

Tight deadlines

Research, verification calls, interviews and briefs — all at once.

Reputational stakes

A wrong source or context hits the public and the station's credibility.

The newsroom isn't short of information — it lacks a way to make origin, context and certainty visible.

Photo: Chicago exchanges, Wikimedia Commons, CC BY 2.0
03 · Problem definition

The core problem:
fragmented information

1Scattered sources 2Hard to trace 3Missing context 4Re-ask, re-search, redo 5Slow story prep 6Higher editorial risk 1Scattered sources 2Hard to trace & cross-check 3Missing source & context 4Re-ask, re-search, redo 5Slow story prep 6Higher editorial risk

Reporters aren't careless. It's systemic: nothing links claim – evidence – source – time – verification status.

04 · Design thinking — Empathize

What users told us

PRIMARY · Reporter

“I don't need AI to write for me. I need to know where this number comes from, which period it covers, and whether any source says otherwise.”

SECONDARY · Editor

“I don't need a pretty summary. I need to see the evidence and what's still uncertain.”

SUPPORTING · Research

One place for sources — with metadata, status and an owner.

How might we

…turn scattered information into an evidence-backed, source-traceable story file that helps editors decide on time — without replacing professional judgment?

05 · Journey redesign

From scattered → evidence-first

AS-IS · DISCONNECTED STEPS Ad-hoc search Personal notes Manual summaries Questions at review ↺ rework loop TO-BE · EVIDENCE-FIRST WORKFLOW Signal ResearchCase SourceCard Claim–EvidenceMap ConflictCheck StoryBrief EditorialApproval AS-IS · DISCONNECTED STEPS Ad-hoc search Personal notes Manual summaries Questions at review ↺ rework loop TO-BE · EVIDENCE-FIRST WORKFLOW Signal Research Case Source Card Claim–Evidence Map Conflict Check Story Brief Editorial Approval

New focus: claim – evidence – source – verification status – editorial decision

06 · Solution design

7 modules.
One evidence-backed file.

01 · Collect

Topic Signal Inbox

Log signals, tagged “unverified”

02 · Collect

Trusted Source Library

Official sources first; source tiers

03 · Collect

Source Card

Link, metadata, excerpt, scope, status

04 · Verify

Claim–Evidence Map

Links each claim to its evidence

05 · Verify

Conflict Detector

Flags discrepancies; never picks a winner

06 · Verify

Story Brief Builder

Sourced brief, open points, interview prompts

07 · Human

Editorial Approval Gate

Checklist review; humans approve

Every key AI output shows source, time, excerpt and limits. Too little evidence → it says “insufficient sources to conclude.”

Photo: CERN Server, Wikimedia Commons, CC BY-SA 3.0
07 · Prototype

MVP: 5 core screens

+ Research Case

01 Signal Inbox

Signals tagged “unverified”

02 Research Workspace

Questions, sources, timeline

🔗 official-source… Date · Scope · Unit ✓ Verified

03 Source & Evidence

Link, date, excerpt, verifier

⚠ Figures differ Source A Source B Period ≠ · Unit ≠ · Scope ≠ Reporter must resolve

04 Conflict Check

Period, unit, scope compared

Verified facts Sources complete Open points (2) Interview questions Approve Revise audit log ▸ 14:32 editor

05 Brief & Approval

Editorial decision, logged

Goal: validate the workflow, not launch a full newsroom platform. CMS, transcription, social analytics and auto-scripting come later.

08 · Governance

Human-in-the-loop:
lines AI can't cross

● AI may
  • Search configured sources
  • Classify, extract metadata
  • Summarize with citations
  • Link claims to evidence
  • Flag conflicts & gaps
  • Suggest interview questions
✕ AI may not
  • Declare information true
  • Draw market conclusions
  • Invent quotes or figures
  • Pick a side in a conflict
  • Approve or publish
  • Replace professional judgment
■ Humans must
  • Check origin & context
  • Verify claims, figures, quotes
  • Decide the angle
  • Interview & assess sources
  • Approve the Story Brief
  • Own final accountability
Safeguards mandatory citationverification statussource tieringeditorial gateaudit trailrole-based access
09 · Agile thinking

Start narrow.
Learn fast.

Hypothesis: does an evidence-first workflow make story prep faster and easier to review — without extra operational load?

In the MVP

  • Research Case + Source Card
  • Initial source whitelist
  • Cited summaries
  • Claim–Evidence Map
  • Basic conflict flags
  • Story Brief + approve
  • Minimal audit trail

Not yet

  • Broadcast scripts
  • Auto-publishing
  • Market forecasting
  • Social analysis at scale
  • Image/video verification
  • Deep CMS integration
  • Confidential data
MVP trace + verify Confidential data Auto-publish Marketforecasts Image/videochecks Social at scale CMS Scripts
10 · Delivery plan

Agile & pilot: 6–8 weeks

Sprint 0

Research & governance

Interviews, baseline, source policy, prototype

Sprint 1

Core workspace

Research Case + Source Card

Sprint 2

Evidence & conflict

Cited summary, Claim Map, flags

Sprint 3

Review workflow

Brief, Approval Gate, audit log

Sprint 4

Pilot refinement

Usability, KPIs, retrospective

2 weeks in parallel → controlled use on live stories
Pilot group
Reporters
5–8
Editors
2–3
Research
1–2
Scope

Listed companies, periodic economic data, market news from public sources

Duration

6–8 weeks, incl. 2 weeks in parallel on real stories

11 · KPIs & learning

Faster — without going
soft on quality

01
Brief prep timeCase opened → brief ready for review
02
Traceable-claim rateLink + excerpt + metadata + status
03
Conflicts caught earlyResolved before review / production
04
Revision roundsEditor requests for missing source/context
05
Adoption & trustInterviews, surveys, continued use
PILOT SUCCESS Solid but slow Fast but risky Baseline EFFICIENCY → VERIFICATION QUALITY →

Safety metrics · unsourced AI output · false-positive conflict flags · users mistaking AI output as “verified”

12 · Risk & change management

Risks → mitigations

● Hallucination
Mandatory citations; say so when evidence is thin; random audits
● Stale or weak sources
Source tiers, whitelist, effective dates, outdated warnings
● Over-reliance on AI
Draft labels, verification status, training, approval gate
● Alert fatigue
Severity levels, tuned thresholds, “not useful” feedback
● Data leaks / copyright
RBAC, data classification, no confidential data in pilot
● Low adoption
Co-design, minimal workflow, parallel run, retrospectives
ADKAR change model
Awareness
Desire
Knowledge
Ability
Reinforcement
Photo: Magnifying glass, Wikimedia Commons, Public domain
13 · Value & scale

Expected value

Organization credibility · responsible AI Newsroom standard files · less rework Reporter less copy-paste

Reporters — more time for verification, interviews, analysis, storytelling.

Newsroom — handover and knowledge reuse by topic, source, claim, lesson.

Organization — quality protected by sources, logs, access control, human oversight.

Scale only when the pilot proves
  • KPIs beat baseline, no rise in quality incidents
  • Users understand AI limits; editors confirm it helps review
  • Source governance, security, copyright & an owner are in place
Conclusion

“AI writes the news faster.”
A newsroom that decides better, on clear evidence.

Evidence-first

Every key claim traces back to source, excerpt, time and verification status.

Human-in-the-loop

AI researches and flags; reporters verify; editors approve.

Start small, learn fast

Narrow MVP, controlled pilot, measure and adjust before scaling.

NEWS-GUARD AI helps VTV1 move faster without trading away verification, editorial accountability or public trust.

Thank you.
Questions welcome
Photo: QTQ9 studio camera, Wikimedia Commons, CC BY-SA 3.0
← → · space · F fullscreen