Decision Trace

AI TRANSPARENCYTRUST

Purpose Create a plain-English “decision story” that shows how an AI system reaches its verdict, who can question it, and what happens next. When people see that story, they trust the output and know where to step in.
Intervention type Transparency & trust scaffolding
Lead Roll-out owner (product, ops, or change lead)
Time 90 min mixed in a group

Expected outcomes

  • User: Understands, in two minutes, how the AI made today’s decision.

  • Team: Knows exactly where a human can review or override.

  • Business: Has a one-page trace map ready for auditors and customers.

  • Organisation: Sets a repeatable pattern for every new model.

What to bring to the session
  • Decision Storyboard sheet (build in the steps below)

  • Trust-Gap Cue Cards (see below)

  • Decision Lane Questions (see below)

  • Ownership Matrix

Steps

1 | Setup

Step 1: Choose one high-stakes decision
A key activity or process within your team.
E.g., “Approve or decline a loan application.”

Step 2: Gather the right voices
At minimum:
• Data or model owner
• Process owner (ops)
• Risk / compliance rep
• One frontline user (e.g., underwriter)

2 | Build the decision story

Step 3: Fill the six lanes
Move left → right. One sticky per fact.

  1. Source Data: Where does the raw info live?

  2. Prep: What do we tidy or change before input into the AI?

  3. Model Decision Rule: How does the system/AI decide?

  4. Machine Output: What is produced?

  5. Human Check: Who can review or change it?

  6. End Impact: What happens with the project, to the customer or business?

Use green stickies for facts, yellow for assumptions, red for “we don’t know”.
Pull questions from the Decision Lane Questions (in resources below) if the group stalls. Or if they struggle to navigate the more technical side.

3 | Spot trust gaps

Step 4: Place Trust-Gap icons
Put a ⚑ on any box that shows:
Surprise: Someone just learned this step exists.
Silence: Data moves teams with no notification.
No-override: Humans cannot step in even if output looks wrong.

Each ⚑ becomes a potential adoption barrier.

4 | Prioritise and assign

Step 5: Rate each trust gap
Score 1-5 for impact (customer harm, legal risk) and 1-5 for frequency.
Top three scores move to the action plan phase below.

Step 6: Name an owner and date
Add one Accountable name per gap, plus a finish date.

Example Decision Storyboard

Marketing team uses ChatGPT to draft a blog post that positions a new product feature.

Resources

Decision Lane Questions
Card 1

Ask:
Who: Data owner
What: “Where does the very first record come from?”
Why: Reveals Source Data

Card 2

Ask:
Who: Data owner
What: “Who controls access to that source?”
Why: Reveals Source Data

Card 3

Ask:
Who: Data engineer
What: “What do we delete or clean straight away?”
Why: Reveals Prep

Card 4

Ask:
Who: Data engineer
What: “If you skip this step, what breaks?”
Why: Reveals Prep

Card 5

Ask:
Who: Data scientist
What: “Explain the decision rule as if I’m an intern.”
Why: Reveals Model Rule

Card 6

Ask:
Who: Data scientist
What: “Which single setting changes the result most?”
Why: Reveals Model Rule

Card 7

Ask:
Who: Product owner
What: “What exact score or label comes out?”
Why: Reveals Machine Output

Card 8

Ask:
Who: Product owner
What: “What does a bad output look like?”
Why: Reveals Machine Output

Card 9

Ask:
Who: Ops lead
What: “Who can press pause before action happens?”
Why: Reveals Human Check

Card 10

Ask:
Who: Ops lead
What: “How often do they actually step in?”
Why: Reveals Human Check

Card 11

Ask:
Who: Process owner
What: “What customer touchpoint fires next?”
Why: Reveals Real-World Impact

Card 12

Ask:
Who: Process owner
What: “How fast does that step happen?”
Why: Reveals Real-World Impact

Card 13

Ask:
Who: Compliance
What: “Which regulation applies here?”
Why: Reveals Risk (Overall)

Card 14

Ask:
Who: Compliance
What: “When is the next mandatory audit?”
Why: Reveals Risk (Overall)

Card 15

Ask:
Who: Any
What: “Where do users complain first?”
Why: Reveals Improvement

Card 16

Ask:
Who: Any
What: “What small win could double trust?”

Trust Gap Cues

Card: Surprise

What to flag: “I didn’t know this step existed.”
E.g. Stakeholder learns something new mid-map

Card: Silence

What to flag: Data or decisions move between teams with no notice
E.g. Break in communication, e.g., emails like “Who pushed that?”

Card: No-override

Label: No-override
What to flag: Humans have zero chance to intervene
E.g. Loss of trust, users create work-arounds or abandon the tool

Other methods within the trust block