1 of 42

HUB47 · AI TRANSFORMATION FOR BETTER BUSINESS · WORKSHOP 3

Power Up Your Business With AI

Automation and AI Agents for Entrepreneurs

Anh Nguyen · Founder, Applify Lab · Dubai

2 of 42

2

AGENDA

Five parts, two demos

01

Why agents get things wrong

and why context, not the model, decides it

02

The three foundations

Knowledge Base · Skills · Tools

03

Four ways to build

from the subscription you already pay for, to custom code

04

WhatsApp, specifically

what Meta actually allows in 2026

05

Multi-agent and the “one-person company”

what's real and what isn't

Goal: Understand AI Automation and implement AI in your business after this session

3 of 42

01

Why agents get things wrong

It isn't lying to you. It's guessing — because that's what it was rewarded for.

4 of 42

4

WHY IT MAKES THINGS UP

Your AI is a student who never leaves an answer blank

Think about how it was trained. Like an exam where a wrong answer and a blank answer score exactly the same: zero.

So guessing always wins. A blank gets you nothing. A guess sometimes gets you a point. Do that a few million times and you get something that never says “I don't know.”

It isn't lying to you. It has simply never been rewarded for admitting it doesn't know.

And it's fixable. The same researchers showed a model can be built to say “I don't know.” Most just aren't asked to.

The difference between a useful assistant and an embarrassing one is usually permission to say “I don't know.”

Kalai, Nachum, Vempala & Zhang, “Why Language Models Hallucinate,” OpenAI, Sept 2025

5 of 42

5

WHAT SCALE FIXES, AND WHAT IT NEVER WILL

Cats and birthdays

Patterns → solved by scale

Spelling. Grammar. Sentence structure. These follow consistent patterns, so errors shrink as models get bigger. Nothing you do is needed here.

Arbitrary facts → never solved

A pet's birthday can't be predicted from a photo — OpenAI's own analogy. No amount of training data fixes it, because there is no pattern to learn.

Your Tuesday price, your refund window, this patient's slot, that client's terms — all birthdays.

The rule: arbitrary facts must be GIVEN to the model, never RECALLED by it. That is what the rest of this talk is about.

OpenAI, Sept 2025

6 of 42

6

CONSEQUENCES

None of these were hard questions

Air Canada, Feb 2024

Its chatbot invented a bereavement-refund policy. The airline argued in tribunal that the chatbot was “a separate legal entity” responsible for its own answers. It lost.

Deloitte Australia, 2025

Fabricated references inside a A$440,000 government report. Partial refund, corrected version published.

Courts worldwide

Over 1,500 decisions by mid-2026 where a judge found a party filed AI-invented material.

Every one of these was a gap in context, filled with the most plausible-sounding thing available.

Moffatt v. Air Canada 2024 BCCRT 149 · Fortune, Oct 2025 · AI Hallucination Cases database

7 of 42

7

THE OBVIOUS FIX THAT ISN'T

More information is not better information

The instinct is to give it everything. Upload the whole drive and let it work the answer out.

Tested across 18 leading AI systems: the same question, answered from one page of the right material, beat the same question buried in three hundred pages that contained it.

One nearly-right document is enough to pull the answer off course. A few of them make it worse.

A bigger memory doesn't fix this. It moves the point where things start to slip. It doesn't remove it.

Giving it more to read is not the same as giving it the right thing to read.

Chroma Research, “Context Rot,” Jul 2025 · Anthropic Engineering, Sept 2025

8 of 42

8

DIAGNOSIS

Five ways your context fails

1

Missing

The answer simply isn't there — so it fills the gap

2

Conflicting

Two documents disagree and nothing says which one wins

3

Too much

You dumped the whole drive in and buried the answer

4

Stale

Last year's price list. Last quarter's policy. Nobody told it.

5

No exit

No permitted way to say “I don't know” or “let me get a human”

All five are engineering decisions. All five are yours to make — not the model vendor's.

9 of 42

You are not buying intelligence. Your competitor can buy the same intelligence tomorrow morning, for the same twenty dollars.

The whole of the next hour is about the other half of that sentence

You are supplying the context — your procedures, your prices, your customers, your tone. That part nobody can buy. The industry phrase for doing it well is “context engineering”: giving the model the smallest set of genuinely useful information, and nothing else.

10 of 42

02

The three foundations

Knowledge Base · Skills · Tools. Miss one and you have a demo, not a system.

11 of 42

11

THE ANATOMY

Every working agent is three things you supply

1

Knowledge Base

what it knows

Your prices, policies, services, past work. Lots of documents — the agent has to find the right one and quote it back.

2

Skills

how you do it

The way you do a job, written down. One right way, followed every time — the same thing you'd teach a new hire.

3

Tools

what it can reach

Your calendar, your customer list, today's prices, what's in stock. The things that change by the hour and must be looked up, not remembered.

The model is the same one your competitor can buy. These three are not.

12 of 42

12

FOUNDATION 1 — KNOWLEDGE BASE

Where does each thing belong? One test.

If it looks like this…

…it belongs here

Example

There are lots of them, and it has to pick the right one

Knowledge base

Service descriptions, policies, FAQs, past proposals

One right way to do it, every single time

Skill

How we qualify a lead. How we handle a no-show. Our tone in Arabic.

It changes by the hour, and something else already tracks it

A live lookup

Today's price. This week's availability. Where this lead got to.

Anything with a per-customer or per-day value must never live in a document.

13 of 42

13

FOUNDATION 1 — THE HOMEWORK

The Knowledge Base Blueprint

1

General business info

Name, hours, address, contact channels, industry

2

Customer profiles

Who you serve best — and the tone to use with each group

3

Services & ecosystem

Core offerings, how they connect, what to recommend alongside

4

Packages & pricing

Tiers, starting prices, payment methods, cancellation terms

5

Brand & positioning

Mission, jargon to translate, pain points, USP, the action to drive

6

Operating guardrails

Allowed outbound hours, timezone rules, when a human takes over

The full question list as worksheet will be sent out after the workshop.

14 of 42

14

FOUNDATION 2 — SKILLS

A Skill is your SOP, written down, loaded on demand

It is a folder with one plain-English file in it. Not code. Anthropic shipped the format in October 2025; it became an open standard that December.

Two required fields: a name, and a description of what it does and when to use it. A vague description means the skill never triggers.

Loaded only when it's needed: the agent always sees the description, reads the rest only when the task matches. Everything else stays out of the way.

44 agent products now read the same format — including OpenAI, Google and Microsoft's tools. A standard, not one vendor's feature.

Analogy: “like putting together an onboarding guide for a new hire.”

Anthropic Engineering, 16 Oct 2025 · agentskills.io, accessed 9 Aug 2026

15 of 42

15

GETTING THE DISTINCTION STRAIGHT

Prompt · Skill · Knowledge Base

A prompt

One-off. Dies with the chat window. Quality depends entirely on who typed it and how much of a hurry they were in. Not a business asset.

A skill

A reusable procedure. Versioned, reviewable, improvable. Same output whoever runs it, on a Monday or a Thursday. This is the asset.

A knowledge base

The facts the procedure reads from. Separate on purpose: you update the price list without rewriting the procedure.

A prompt is a conversation. A skill is a process.

Which is why the question is never “which AI tool should I buy?” It's “which of my processes is written down?”

16 of 42

16

FOUNDATION 2 — CRAFT

What makes one work, and why they fail

WHAT MAKES ONE WORK

Write down the surprises. The things a new hire gets wrong on day one.

Show an example of the output. Better than describing it.

Be strict where it matters. Loose everywhere else.

WHY THEY FAIL

Too vague. The AI never realises it should use it.

Too long. It can't find the part that matters.

Written by an AI. It doesn't know your business. Looks finished, says nothing.

The test for every line: would it get this wrong without it? If no — cut it.

agentskills.io — best practices, accessed 9 Aug 2026

17 of 42

17

DEMO · REAL-ESTATE RESEARCH

One of these had the skill

WITHOUT the Skill

WITH the Skill

Left — no skill. Prose. Not wrong, but generic.

Right — with the skill. Ask question about your situation. Traffic lights, five checks, the yield you were told against the yield you'd get, stress tests, red flags.

It shows its working. Sources cited, and three items flagged as unverified.

Not a smarter AI.

The same AI, given a written procedure — what to check, in what order, and what the answer should look like.

18 of 42

18

WHAT A SKILL ACTUALLY LOOKS LIKE

Example 1: this is the whole file

name

grill-me

description

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions “grill me”.

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer.

Ask the questions one at a time.

If a question can be answered by exploring the files, explore them instead of asking me.

19 of 42

19

WHAT A SKILL ACTUALLY LOOKS LIKE

Example 2: the skill behind this deck

one folder, shown in two halves

Same idea, more parts. A folder instead of one file: the instructions, plus the templates, brand assets and scripts they point at.

It builds every deck we send. Fonts, colours, layouts, the checks that catch a broken slide. Including this one.

Nobody starts here. This began as a single page. The rest got added the fifth time I explained the same thing.

Start with the one-pager. Turn it into a folder once you've explained the same thing three times. Continuously improve it

20 of 42

20

FOUNDATION 3 — TOOLS

The hands: how the agent reaches your systems

MCP is a standard plug. One connector shape, so an AI can reach your calendar, customer list or accounting system instead of you copy-pasting.

Nobody owns it. Anthropic built it, then handed it to the Linux Foundation — the same neutral home as the software behind most of the internet.

Over 10,000 of these connectors already exist, and ChatGPT, Claude, Microsoft Copilot and Gemini all accept them.

One question for your developer: the standard changed in July 2026. If something was built for you before then, ask whether it still needs updating.

Skills are the instructions. Tools are the hands. Neither is much use without the other.

modelcontextprotocol.io · Linux Foundation press release, 9 Dec 2025

21 of 42

DEMO

One instruction. No hands on the keyboard.

I asked for a list of prospects. It opened the browser, searched LinkedIn, read the profiles, and came back with the research done.

22 of 42

22

FOUNDATION 3 — GOVERNANCE

Six rules before you connect anything

1

Read-only first

Let it read before you let it write. That removes most of the damage, not all of it.

2

Give it the narrowest key

Access to one thing, for one job. Never the master key to everything.

3

Approve changes properly

You have to see exactly what it's about to do, not a summary of it. That's how real attacks got through.

4

Check what you plug in

What you approved on Monday can quietly change on Tuesday. Re-check anything you didn't build.

5

Connect fewer things at once

The more it can see in one session, the more ways there are for one of them to mislead it.

6

Treat customer input as hostile

A ticket, a message, a CV, a review — all text a stranger wrote. Agents follow instructions hidden in it.

“Untrusted content” isn't hackers. It's a ticket, a review, a CV — anything a stranger can type.

23 of 42

03

Four ways to build

Same three foundations at every level. Only the machinery around them changes.

24 of 42

24

THE LANDSCAPE

Four routes, honestly compared

Route

Who it's for

Typical cost

Where it breaks

1 · Your existing subscription

Claude Cowork · ChatGPT Work

You, on your own work

$20–125

per seat / month

One person's identity. No SLA, no retries.

2 · Drag-and-drop tools

n8n · Make · Zapier

Work that must run without you

$20–667 / month

plus build time

Hard to fix, and it fails quietly.

3 · Run-it-yourself agents

OpenClaw · Hermes Agent

Tinkerers and technical founders

Free, plus what the AI costs

and a server

Security. You are the support team.

4 · Custom code

Mastra · LangGraph · Agent SDKs

Products, regulated work, promises you signed

Build + ongoing

upkeep

Too early, nine times out of ten.

Nobody moves up a level because the tool is better. They move up because they hit a wall they can name.

25 of 42

25

ROUTE 1

The subscription you already pay for

Cowork inside Claude, Work inside ChatGPT. Built for everyday office work rather than coding — and included in plans you may already be paying for.

You give it a job, not a question. “Draft the partner announcement.” “Turn this contract into a template.” It goes away and does it.

It carries on when you don't. Work continues after you close the laptop, and it stacks up what's ready for you to look at.

It stops and asks. Approve it, answer its question, or send it back. Nothing goes out without you.

Where the majority of AI automation use cases can be accomplished.

26 of 42

26

ROUTE 2

n8n and the drag-and-drop tools

The trigger to move here: “I want it to run without me.”

This one runs every morning. It checks what people are searching for, picks a topic, researches it, writes the post, publishes to three channels, and logs the result.

Eight steps, no code. Every box is something you drag in and fill out.

Where it breaks: big files, no proper history of changes, and no way to test before you ship.

Require a bit of a learning curve.

27 of 42

27

ROUTE 3 — OPENCLAW · HERMES AGENT

An agent that lives in your chat app

WhatsApp

Telegram

Slack

Your gateway

one always-on process, on a machine you control

The model you pick

any provider, or one on your own box

Your tools

files · calendar · browser · posting

One always-on program you host yourself. A small server, or a spare laptop that never sleeps.

It answers inside apps you already use. WhatsApp, Telegram, Slack — same thread, same phone.

You pick the brain and the hands. Any model, and only the tools you choose to connect.

What people run on it: scheduling posts, chasing invoices, a morning summary, turning a photo into a finished listing.

It's yours — and so is the security. Fine for your own work. Don't put a client's data behind it.

28 of 42

28

ROUTE 3 — A REAL ONE

One message, and the post is scheduled

Sent from Telegram, on a Saturday. A photo of the meeting flyer and one line: “schedule an Instagram post for tomorrow morning, add a caption.”

It worked the rest out on its own. Read the flyer, wrote the caption, uploaded the image, converted the time to Dubai, and queued the post.

Then it reported back. Account, post ID, status, scheduled time and the exact caption it used.

No dashboard, no laptop — on the same phone you already answer messages on.

29 of 42

29

ROUTE 4

Custom code — when it is unavoidable

WHERE THE OTHER ROUTES STOP

Route 1 stops at “it's you”. One login, one job at a time. If it fails at 2am nobody is told, and no customer can sit behind your account.

Route 2 stops at branching. Fixed boxes in a fixed order. No version history, and no way to test a change before it goes live.

Route 3 stops at security. Your keys, your server, your support rota — and nothing separating one customer's data from another's.

Not one of those says the tool isn't clever enough. They are the same shape: it has to run as the business, not as a person.

SO WHERE IS CODE UNAVOIDABLE?

You sell the agent, not just use it. Many customers on one system, each one's data walled off from the next, and an uptime promise in a contract.

The data isn't allowed to leave. Patient records, case files, banking. Where it is processed is set by a regulator, not by preference.

Every case takes a different path. A claim, a dispute, an application. The agent decides what happens next — a decision tree, not a diagram.

Money or safety moves at the end. A payout, a prescription, a contract term. Every step recorded, replayable and reversible.

Microsoft, in its own agent framework docs: “If you can write a function to handle the task, do that instead of using an AI agent.”

30 of 42

04

WhatsApp, specifically

The channel that matters most here — and the one with the most rules.

31 of 42

32

OPTION A OR OPTION B

WhatsApp is just a doorway

1

Knowledge Base

Services, prices, policies, hours — in both languages, authored not translated.

2

Skills

How you qualify. How you book. How you handle a no-show or a complaint.

3

Tools

Calendar, CRM write-back, payment link, handover. The parts that change hourly.

Option A — Meta runs it

Customer

WhatsApp

Meta

Business Agent

Meta hosts the agent. You configure it through their API.

Option B — you run it

Customer

WhatsApp

Cloud API

Your

agent

Your

systems

Meta just carries the message. The logic and the data are yours.

32 of 42

33

META BUSINESS AGENT, UNDER THE HOOD

Meta named the same three things

In this talk

In Meta's own API

What it holds

Knowledge base

Agent Knowledge —

Business Info · FAQs · Files · Websites

Hours, locations, policies · Q&A pairs · uploaded files · your crawled website

Skills

Agent Skills

“System instructions that shape how it responds.” Meta's words, not mine.

Tools

Connectors + Connector Tools

“An external API the agent can call, so it can do more than answer questions.”

Guardrails & handoff

Settings · Allowlist · Thread Control

Persona, language, handoff policy; who it may talk to; when a human takes over

Measuring it

Agent Test + Agent Eval

The with-and-without discipline from part two, shipped as an endpoint

The minimum to switch an agent on is Onboarding plus Settings. Knowledge and connectors are optional — which is exactly how you end up with a bad agent.

Meta for Developers — Meta Business Agent Platform, “Get started”, updated 3 Aug 2026

33 of 42

34

OPTION A — WHEN NOBODY FILLS IT IN

The same agent, without the setup

one conversation, left then right

The first half went well. It asked the right questions — property type, building, paint colour, furnished or empty. Straight from the knowledge base.

Then the customer typed “Human”. Three times. The agent: “I'm not able to help with that. Is there something else I can help you with?”

Nothing was wrong with the AI. Category six of the blueprint was empty. Nobody told it when to hand over, or to whom.

Meta requires a route to a human. It doesn't write one for you — and this is what missing it looks like.

34 of 42

35

OPTION B — YOUR OWN AGENT

The architecture, end to end

Customer

WhatsApp message

Cloud API

Meta webhook.

No provider required.

Your agent

Skill + knowledge base

+ guardrails

Your systems

Calendar · CRM ·

prices · payment link

Reply

…or handover

to a human

You do not need a solution provider. Direct Cloud API signup is open. New accounts start at 250 unique recipients per 24 hours and ladder up.

A provider buys you three things: credit terms, faster verification, and throughput above 80 messages a second. The inbox and the AI are software, not provider status.

Meta has billed in AED since April 2026, so a UAE business can be invoiced directly in dirhams.

The one extra box is the whole difference: it's where your knowledge base, your skill and your tools live.

Meta for Developers — Cloud API onboarding, messaging limits and billing, accessed 9 Aug 2026

35 of 42

05

Multi-agent and the one-person company

Where this is going — and how much of it is true today.

36 of 42

37

ORCHESTRATION

When one agent becomes several

The shapes have names, and vendors will use them at you: one agent managing others · a chain where each hands to the next · several working at once · one writing while another checks · one that stops and asks you.

The upside is real: Anthropic's research system beat a single agent by 90.2% — and used roughly fifteen times the tokens of a chat. Same paragraph, both numbers.

The downside is also real: Berkeley tested seven popular multi-agent frameworks and found failure rates between 41% and 87%.

The field changed its mind in public: Cognition published “Don't Build Multi-Agents” in June 2025, then “What's Actually Working” in April 2026.

One agent holds the pen. The others read, review and advise.

Anthropic Engineering, Jun 2025 · Cemri et al., arXiv:2503.13657 · Cognition, Jun 2025 & Apr 2026

37 of 42

39

ECONOMICS

Not every step needs the smartest model

Model

Cost to read

$ per million words

Cost to write

$ per million words

What it's for

Claude Opus 5

10.00

50.00

Hardest reasoning you have

GPT-5.6 Sol Pro

5.00

30.00

Frontier work, long chains

Claude Sonnet 5

2.00

10.00

Everyday drafting

Gemini 3.5 Flash Lite

0.30

2.50

High-volume summarising

GPT-5.6 Luna

0.10

0.60

Classify, extract, route

DeepSeek V4 Flash

0.09

0.18

Bulk tagging and cleanup

Top to bottom of that table is roughly a hundred times the price. Same job, wildly different bill — and most steps belong near the bottom.

OpenRouter list prices, 9 Aug 2026. Priced per “token” in the industry — roughly 750 words per 1,000 tokens; the table converts it to words.

38 of 42

40

THE IDEA WORTH STEALING

The “one-person company”

1

What people mean by it

A business that reaches serious scale without hiring — because software, and now agents, absorb the work a department used to do. Altman put a billion-dollar version of it in a bet in 2023.

2

Why it isn't real yet

He revised it to “two or three people” in late 2025. The company he later called proof has two staff, an FDA warning letter and a class action. Real revenue, and a mess nobody had time to catch.

3

Why it still matters to you

Treat it as a design principle, not a destination: before you add a person or a process, ask what would have to be true to not need one. Then automate that.

Solo businesses adopt AI at about half the rate of businesses with employees — 15% versus 26%. In a room like this, that gap is the opportunity.

Fortune, Feb 2024 · Conversations with Tyler, Nov 2025 · JPMorgan Chase Institute, 2026 (bank payments data)

39 of 42

41

THE CONSTRAINT THAT DOESN'T GO AWAY

Generating is cheap. Checking is not.

AI improves fastest at things that are easy to check — maths, code that either runs or doesn't. It improves slowest where the answer needs judgement. That's most of your business.

41% of knowledge workers report AI output that “looks finished but lacks substance” — about two hours of rework each time.

A randomised trial of experienced developers found they were 19% slower using AI while believing they were 20% faster. Small study — the confidence gap is the point.

Delegate → Review → Own. A named human is accountable, whatever produced the output.

Budget review time explicitly. If you don't, quality slips quietly and you'll blame the model.

Wei, Jul 2025 · BetterUp Labs & Stanford via HBR, Sept 2025 · METR randomised trial, Jul 2025

40 of 42

42

RECAP

Three things, if you remember nothing else

1

It's the context, not the model

Missing, conflicting, stale, buried in noise, or no permitted way to say “I don't know.” None of it is solved by switching vendor.

2

Automate a process you already have

If you can't do it manually and describe it, you cannot automate it. The written procedure is the asset — the tool is interchangeable.

3

You still own the review

Never auto-execute money, contracts, external sends or deletions. Design for wrong, not for perfect.

Everything else in this talk was implementation detail.

41 of 42

43

WHAT TO DO MONDAY

One page. That's the whole assignment.

1

Pick one task you do twice a week

Repetitive, rules exist, low damage if it's wrong once.

2

Write it down like training a new hire

Inputs, steps, edge cases, and what “good” looks like.

3

Add your gotchas

The things that defy an outsider's assumption. This is the valuable half.

4

Write two real examples

And the exact output you wanted. Real ones, not invented ones.

5

Run it with the page and without

That comparison is your baseline. Now you can evaluate any vendor honestly.

That single page is the input to every tool in this talk — the subscription, the workflow, and the custom build.

42 of 42

Questions

Anh Nguyen · Founder, Applify Lab info@applifylab.com · www.applifylab.com · +971 553110959