INTERSECTIONS · Service Design ConferenceSingapore · 23–24 July 2026
High-touch service at machine scale

Human
in the Loop

AI backstage.
Humans in the relationship.

Jensen Loke

CEO, Success IT

Jensen Loke · Success IT · Human in the Loop01
Your speakerProduct · service · organisation · software
JLJensen Loke · Singapore
CEO, product builder and community organiser

I design products, workflows and teams that make complex services work better.

Across consumer and corporate digital products, the goal has stayed the same: use technology to solve real business problems and improve outcomes.

Consumer products

DBS digital banking

Corporate platforms

Temasek · Ant Financial · Amazon

Now

CEO, Success IT

Community

1,000+ Agentic Builders

linkedin.com/in/jensenloke · agenticbuilders.sg/community02
Success ITEnterprise software · Singapore & Malaysia
Founded in 2005

Twenty years of software that runs real businesses.

Customized accounting and operational software for automotive, insurance intermediaries, wholesale and trading businesses—plus bespoke client systems.

Years operating20+
Client entities188+
Software products30+
Policies managed250k+
Sources: GBrain and Obsidian company profile; company metrics snapshot dated February 202503
Our operating philosophyDesign the service before adding the agents
Two promises to every B2B customer

Easy, configurable software.
Human, accountable support.

Clients should not have to prompt a bot, supervise it or fight through it to reach someone who owns the outcome.

Why people stay at the front door

B2B requests arrive with history, ambiguity and consequences.

  • Customers often describe symptoms, not the underlying problem.
  • Exceptions require judgment across data, product and relationship context.
  • A support subscription buys accountability—not another queue to navigate.
What AI changes backstage

Management can now engineer the tools that expand staff capability.

Jensen + coding agentsHands-on software engineering from management
→
Zach + InsaPurpose-built tools, context and guardrails
→
Iris + JianzhiMore technical reach without becoming engineers
Success IT operating model: agents expand staff capability while people retain the client relationship.04
Today’s routeFollow one customer issue backstage
Agenda

One issue. Two agents.
Several human decisions.

ACT I

Meet Iris and Jianzhi

Two people · two human loops
ACT II

Zach and the customer map

From shorthand to customer context
ACT III

Insa and configured workrooms

From support knowledge to a web fix
ACT IV

Adoption across the team

Why non-engineers lead agent adoption
ACT V

Boundaries and business impact

What remains human—and what it may save
Narrative: context → case study → evidence → operating principles → economics04
OpenClaw setupConversation outside · controls underneath
The conversation is only the front door

OpenClaw turns a message
into a bounded workflow.

01 · SURFACE

Discord, WhatsApp, email

Staff speak in the tools they already use.

02 · ROLE

Zach, Insa, Laylah

Each agent has a job, memory and operating boundaries.

03 · WORKFLOW

Graphs, tools, checkpoints

Routing, diagnostics and approvals become auditable steps.

04 · AUTHORITY

Success IT people

Humans correct, approve, release, test and communicate.

Sources: OpenClaw agent routing, Zach/Insa workflow plugins and support-to-codeflow runbooks06
The backstage teamFour roles · one service system
The customer never needs to learn their names

Meet the team working behind ours.

Laylah

PA and orchestration

L

Zach

Support intake and coordination

Z

Insa

Investigation and engineering depth

I

Peggi

Finance and bookkeeping

P
Sources: OpenClaw workspace roles and workflow plugin manifests02
The Success IT teamSmall company · blurred role boundaries
Working team size ≈14

people across Singapore and Malaysia

The adoption surprise

Our heaviest agent users are not engineers.

JZ

Jianzhi

BD & support lead · 5+ years
No technical background. Today, the most active Insa user.
JO

Jackson

Business development lead · 5+ years
Learned to code over the past five months.
IO

Iris

Level 1 support · 8 months
Uses Zach for intake and Insa for investigation.
ENG

Engineering team

Deep product knowledge
OpenClaw adoption is uneven—and generally lower per person.
Roles and capability changes supplied by Jensen; approximate headcount should be confirmed before publication06
Meet the humansTwo roles · two forms of leverage
The agents are not the protagonists

Iris carries the relationship.
Jianzhi carries the product history.

Iris

Level 1 customer support · 8 months at Success IT

Works with Zach + Insa
Jianzhi

BD and support lead · Microsoft Access product specialist

Works directly with product-scoped Insa
Human leads: customer-service judgment and legacy-product knowledge, amplified through different agent workflows07
Thirty-second decoderFive terms used in the case study
Plain language first

Enough technical vocabulary
to follow what happens backstage.

SuccessGraphThe customer mapShows the client, their products and related companies.
CodeGraphThe software impact mapShows which parts of the product a change could affect.
Database · DBThe business recordsWhere customer transactions and operational data are stored.
Pull request · PRA proposed code changePrepared for a human to review—never the same as releasing it live.
Product laneA configured workroomA dedicated chat tied to one person, product and set of permissions.
Working rule for the talk: describe what each mechanism does before naming how it is implemented.09
Iris · Level 1 supportUnstructured report → structured work
The first transformation

Iris speaks naturally.
Zach makes the work legible.

Iris

“Customer cannot update the policy number. It must go out today. Screenshot attached.”

→
Zach

Customer · product · priority · actual vs expected · evidence · missing facts

→
Human checkpoint

Iris corrects, approves, clarifies or reroutes.

Synthetic example. Workflow: Zach ticket-create draft → human confirmation → verified creation.04
In practice · Iris + ZachNatural request → draft → approval → ticket
What the staff member sees

Approval is part of the conversation.

Unstructured intakeIris adds the request and evidence in Discord.
Structured draftZach exposes the fields before taking action.
Explicit approvalIris remains accountable for what is logged.
Redacted Discord conversation showing Iris asking Zach to create a ticket, reviewing a structured draft and approving it
Public-safe reconstruction from a real workflow screenshot; customer and ticket identifiers redacted.
Real Success IT workflow · public-safe redaction prepared for conference use10
Zach deep diveMeasured system activity
How good is Zach?

Mature at intake.
Honest about failure.

682ticket-created records in the email-state ledger
456audited deterministic workflow invocations since 29 May
837recorded triage decisions since 23 March
12durable Zach–Insa handoff items: a real but emerging lane
Internal Zach ledgers, read-only extraction on 19 July 2026. Activity may include probes or migrated records.05
Zach deep diveSuccessGraph · internal customer map
Redacted SuccessGraph interface showing a customer relationship blast radius within a larger knowledge graph
Customer graph and blast radius; identifying records redacted.
Context before escalation

It tells Zach
where to look.

Customer

Client + related companies

System

Customer account + platform

Commercial

Products + subscription

Routing

Iris, evidence gate or Insa

Important: SuccessGraph adds context. Iris still decides whether the ticket reflects what the customer meant.

Verified in Zach's SuccessGraph client, ticket playbook and triage engine. Enrichment can degrade gracefully.11
Zach workflow mapMermaid · intake → context → bounded routing
Where the human checkpoints sit

SuccessGraph tells us where to look.
Iris tells us what the customer meant.

Adapted from workflows/support-to-codeflow/workflow-diagram.md and current Zach safeguards.12
Insa deep diveSupport symptom → technical evidence
Redacted CodeGraph interface showing a highlighted dependency blast radius in a large codebase graph
Code dependency blast radius; internal identifiers redacted.
01

Open the right workroom

Product, source code, customer system and permissions

02

Trace the system

Software impact map, screen, service and customer-data dependencies

03

Investigate safely

Read-only, audited diagnostics with durable findings

04

Return evidence

Explain, package a fix, or ask the human for more

Sources: Insa workflow plugin, engineering execution and data-access runbooks06
Insa deep dive35 configured channel profiles · role + product context
Success IT · Product support
# iris-insuranceSupport investigation · read-only diagnostics
# jianzhi-expJianzhi · Expat product development lane
# jackson-workshopJackson · Workshop delivery context
The conversation feels unstructured

The execution behind it is anything but.

01

Resolve who + product

Working role, product and permitted intake mode

02

Load the configured workroom

Source code, customer-data path, safe working copy and review target

03

Stop when uncertain

Ambiguity, protected code or missing approval stops the workflow

Illustrative channel names; structure verified in Insa's live registry and lane resolver. No private channel IDs shown.13
Insa product routerConversation → configured workroom
The chat already knows which product and tools are allowed

One Insa.
Different operating boundaries.

Verified against Insa's product map, live channel registry, lane resolver and workflow tools.15
Jianzhi · BD & Microsoft Access supportProduct knowledge → small web fix
A different human loop

Jianzhi knows the workflow.
Insa bridges the web stack.

Jianzhi brings

Customer and Access-era product knowledge

  • Recognizes the business rule
  • Describes the bug in Discord
  • Does not navigate coding CLIs
Insa’s Discord lane

Scoped to Jianzhi + one product

  • Starts with the correct product context
  • Finds the code and affected components
  • Investigates and makes a small change
The result

A reviewable web-app fix

  • Small bugs and scoped changes
  • A proposed change with verification evidence
  • Humans retain review and release authority
Workflow description supplied by Jensen; product-scoped Insa channels and PR artifacts verified in runtime evidence12
In practice · Jianzhi + InsaProduct judgment → scoped technical execution
A product lane in use

“This report only.”

Human boundaryJianzhi specifies what may—and may not—change.
Agent confirmationInsa restates the exact scope before continuing.
Visible failureTool issues are surfaced without abandoning the task.
Redacted Discord conversation in the Expat product lane showing Jianzhi defining a narrow report change and Insa confirming scope
Public-safe reconstruction from a real Expat product-lane conversation; legal text and identifiers redacted.
Real Success IT product-lane workflow · public-safe redaction prepared for conference use18
Insa deep dive29 May–17 July 2026
How good is Insa?

8,159 audited operations.
Not one vanity “accuracy” score.

2,142read-only checks of customer data
1,721software impact-map checks
1,165readiness-gate operations
441code-change packages prepared for review
Internal Insa audit ledger. Counts measure workflow activity, not correctness, fixes, or hours saved.07
AdoptionOne agent · different jobs
How much does the team use Insa?

The pattern follows the work.

PersonAudited operationsPrimary patternRelative activity
JianzhiBD & support · nontechnical2,538Workroom setup · safety checks · change preparation
IrisLevel 1 support · 8 months527Customer-data checks · investigation · support context
TracyEngineer262Issues · product investigation
JacksonBD lead · coding for 5 months209Multi-product delivery · investigation
JustinEngineerNot cleanly attributableBroader adoption signal only
Insa audit ledger matched via current channel registry; counts are operations, not personal productivity.08
Capability expansionAdoption is not an engineering-only story
The shift happened outside engineering

They did not start as developers.

Jianzhi

BD and support lead · no technical background

245

proposed code changes recorded in Jianzhi’s workrooms

Jackson

Business development lead · learned to code in five months

12

proposed code changes associated with Jackson’s workrooms

Artifact records in staff-scoped channels; includes lifecycle and promotion work, not a feature count.09
Human in the loopThe stop conditions matter
The trust story is not perfect automation

The system knows when it does not know.

01

Evidence gate

Missing screen path, sample record or expected behavior returns to Iris.

02

Authority gate

Changes to live customer systems and software releases remain human-controlled.

03

Reality gate

A human tests the live system and Iris decides what the customer should hear.

Verified in handoff graph tests, data-access policy and testing/closure runbooks.10
Bounded autonomyCapability is not authority
The operating contract

Agents carry context.
Humans carry responsibility.

Zach + Insa

  • Structure incomplete information
  • Load product and customer context
  • Trace dependencies and diagnose
  • Prepare reviewable actions
  • Maintain durable audit trails

Iris + engineers + Jensen

  • Understand the customer’s real need
  • Correct and approve the record
  • Review code and data changes
  • Release and verify the live system
  • Own the relationship and outcome
Success IT operating model; customer communication remains on the human support surface.11
Founder in the loopFrom fixing incidents to improving the system
My job changed too

I no longer ask only:
“How do I fix this?”

The old escalation

Jensen or an engineer takes over the issue and fixes it directly.

The customer is helped once.

The new escalation

Why did Zach or Insa fail to help the staff member?

Model · context · tool contract · routing · permissions · infrastructure. Fix the layer, then replay the incident.

Operating philosophy supplied by Jensen; system-layer RCA pattern verified in OpenClaw incident reports and handoff safeguards.18
Business impactCapacity without simple headcount equivalence
5
Fewer engineering roles since 2025

The business kept operating while the team became smaller.

Agent-assisted workflows are part of that capacity story—but headcount change alone does not prove causation.

Conference-safe wording: “We operate with five fewer engineering roles than in 2025 while expanding the depth available to support and engineering staff.”
Headcount change supplied by Jensen. Savings attribution requires payroll timing and before/after workload evidence.12
Capacity modelIllustrative time assumptions · validate internally
An illustrative, testable base case

How much human time could the agents return?

Edit the assumptions live. These are working estimates—not measured productivity claims.

Illustrative time model. Validate with a sample of real tickets, investigations and code-change tasks.13
Evidence standardSay only what the data supports
The honest impact statement

We can prove adoption.
We still need to prove causation.

We can claim now

  • Zach operates a mature, audited intake lane.
  • Insa is used extensively for diagnostics and engineering workflows.
  • Iris, Jianzhi, Jackson and Tracy use Insa in role-specific ways.
  • Critical production authority remains with people.
  • Success IT operates with five fewer engineering roles than in 2025.

Measure before claiming

  • Exact hours saved per ticket or investigation.
  • Resolution-time improvement.
  • Reduction in reopen or escalation rates.
  • How much headcount reduction is attributable to agents.
  • Net savings after infrastructure, model and management costs.
Recommended next evidence: 10-ticket time study, pre/post comparison and payroll-cost reconciliation.14
Continue at the intersectionAgentic Builders Collective · Singapore
Builders helping builders

Come learn agentic coding with us.

A Singapore community for developers, designers, operators and tinkerers building software with AI.

1,000+active members
DailyWhatsApp discussion and troubleshooting
Monthlymeetups, demos and workshops
QR code to join the Agentic Builders Collective WhatsApp group Scan to join the WhatsApp group agenticbuilders.sg
agenticbuilders.sg/about · Community count as of July 202631
End · QuestionsHuman in the Loop
The AI is not the relationship. Our people are.

Questions?

Thank you.
Let’s continue the conversation.

QR code to connect with Jensen Loke on LinkedIn Connect with Jensen linkedin.com/in/jensenloke
Jensen Loke · CEO, Success IT · Co-organiser, Agentic Builders Collective32
1 / 15