IEEE Computer Society Malaysia Chapter
2026 Webinar Series & Professional Workshops
From grammar checks to autonomous agents
Operationalizing AI for real productivity.
Ts. Dr. Muhammad Faiz bin Mohd Zaki
Director, Pusat Pengurusan Data dan Maklumat, Universiti Malaya
Wednesday 29 July 2026
Everything in this talk descends from one low-key tweet — and it is not yet four years old.
—
days from that tweet to this webinar.
What moved in that window
A text box you typed into became systems that plan, call tools and finish multi-step work unattended. The technology did not slow down. Most of our workflows never moved at all.
One mark per day since that tweet — seven to a column, one column per week
Sam Altman (@sama), 30 November 2022 — the day ChatGPT was released publicly. Counter is computed live from that date.
Malaysian workers are ahead of the global frontier. The organisations around them are not.
Classed as “Frontier Professionals” · Malaysia24%
AI users who say leadership is clearly and consistently aligned on AI32%
AI users producing work they could not have made a year ago69%
What counts as a Frontier Professional
Microsoft classes someone as one only if they do all three: use agents for complex, multi-step work; routinely redesign their own workflows around AI; and take part in structured, repeatable practices that scale beyond themselves.
19%
say they are rewarded for reinventing how work gets done when the effort does not immediately produce results.
The transformation paradox
Microsoft's own term: employees feel pressure to adopt AI quickly, while the metrics, incentives and norms around them still reward the old way of working.
Frontier Professionals, drawn as people — six in twenty-five
Microsoft 2026 Work Trend Index, Malaysia release, 23 June 2026 — trillions of anonymised Microsoft 365 signals plus 2,000 full-time employed and self-employed knowledge workers in Malaysia. Figures for leadership alignment and reward are among AI users. news.microsoft.com/source/asia
And yet the value has not shown up. Nineteen out of twenty pilots report no measurable impact.
95%
of enterprise generative-AI pilots delivered no measurable P&L impact. Only 5% reached production with value that could be measured.
Twenty pilots. One reaches production.
USD 30–40 billion
spent by enterprises on generative AI. The money moved before the workflows did.
89% of nearly 6,000 executives
report no measurable effect of AI on their firm's labour productivity over three years.
Individual, not institutional
MIT's diagnosis: the tools raise personal output but never reach the operations that show up in earnings.
MIT NANDA, “The GenAI Divide: State of AI in Business” (2025) — 300 public deployments, 52 interviews, 153 executive surveys. NBER Working Paper 34836 (2026), Yotzov, Barrero, Bloom et al., “Firm Data on AI” — nearly 6,000 senior executives at US, UK, German and Australian firms. Worth stating plainly: MIT's 95% is contested — it counts only P&L impact measured within six months, so efficiency and quality gains fall outside the definition.
The hours are not lost writing sentences. They are lost in the coordination that surrounds them.
“Work about work” — chasing, coordinating, searching, re-clarifying60%
Skilled work you were hired for~25%
Where the tool lands
A grammar checker improves the wording inside the quarter. It cannot touch the 60% — because that work is not writing. It is remembering, chasing, checking, routing and following up across systems.
That is multi-step work. It needs something that can take steps.
Illustrative — the same three shares laid across one working week
Asana Anatomy of Work Index — over 10,000 knowledge workers globally; Asana reports 60% on “work about work”, about a quarter on skilled work and 13% on strategy. Vendor-run research, so treat the exact split as indicative. Grammarly / The Harris Poll, State of Business Communication (2024): knowledge workers spend 88% of the workweek communicating across channels.
The shift
DelegatedRef 06/11
Agentic AI is a ladder, not a switch. Each rung takes a little more of you out of the loop.
We have spent three years asking AI questions. A colleague is not something you query — it is something you hand work to.
01 Assist
You drive. It edits.
You are present for every step. It improves what you have already produced.
You act at every step
02 Delegate
It proposes. You approve. It executes.
It reads the situation, decides what it can finish, and asks. One instruction, several steps, a finished artefact.
You act once — the approval
03 Unattended
It runs. You read the exceptions.
It wakes on a schedule with no prompt at all, works while nobody is watching, and surfaces only what needs a human.
You act not at all
Record 01 / rung one
AssistedRef 07/11
The documentation agent turns a folder of scattered notes into a draft you edit — not one you write.
The painful thing it removes
Nobody minds writing the document. What they mind is assembling it — finding the last version, reconciling three people's comments, re-typing numbers that already exist in a system, and reformatting to a template someone circulated in 2019.
The agent does the assembly. You do the judgement.
Scattered sources → one draft
What it does on its own
Gathers the sources, extracts what is relevant, drafts to your house format, and flags what it could not find.
What stays with you
Every claim, every number that matters, and the decision to send. It drafts; it does not sign.
Why start here
Lowest risk on the ladder — a bad draft costs you nothing but a delete key.
Record 02 / rung two / live
DelegatedRef 08/11
The secretary agent briefs you each morning, tells you what it can finish itself, and waits for your word.
The whole idea is on the fourth line
An agent that claims it can do everything is a demo. An agent that tells you which task it should not touch is a colleague.
Nothing executes until you say so. It proposes; you instruct; it carries out the multi-step work and hands back a finished artefact. Emails are drafted and saved — never sent.
That single approval step is what makes autonomy safe enough to actually deploy.
Record 03 / rung three
UnattendedRef 09/11
The night auditor works while nobody is watching, and leaves an exception report waiting at 8am.
03:14
Wakes on a schedule. No prompt, no session, no human.
03:16
Reconciles yesterday's figures against the sources that should agree.
03:22
Tests each against thresholds the team defined — not thresholds it invented.
03:25
Writes a short exception report. Silent when everything reconciles.
How you prove it
This is the rung people find hardest to believe — and the proof is never a live demonstration. It is a log with a timestamp on it, from a run that nobody attended.
Direction
DelegatedRef 10/11
Agents pay handsomely inside their capability frontier and cost you outside it — so choose the task before you choose the tool.
Measured effect inside the frontier
+12.2%
more tasks completed
25.1%
faster to finish
+40%
higher rated quality
Same study, same people: outside the frontier, AI made their work worse.
Schematic — the boundary is uneven, and similar-looking tasks fall on either side
Rule one
Pick something that recurs. A task you do weekly is worth automating. A task you did once is worth doing once.
Rule two
Keep the approval step. Propose-then-execute is not a limitation; it is what lets you deploy at all.
Rule three
Measure the hour, not the vibe. If you cannot say what it saved by week four, you are in the 95%.
Dell'Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon & Lakhani, “Navigating the Jagged Technological Frontier”, Organization Science (2026) — preregistered field experiment with 758 Boston Consulting Group consultants, run on GPT-4.
Closing minute
For your imaginationRef 11/11
When execution stops being the bottleneck, the scarce thing becomes knowing what is worth doing.
This asks less of nobody. It asks more of everyone — and it asks it further upstream.
Can use the tools.
Trainable, and we should train everyone. This is the floor, not the goal — and it is where most AI programmes stop.
Can encode how the work actually gets done.
Also trainable — but only for people who already master the domain. You cannot teach an agent a process you have never owned.
Can see the thing nobody has asked for yet.
This is the ceiling, and no course reaches it. It is the one layer that decides how far any of this actually goes.
Towards an intelligent, data-driven university