The deck from Put AI to Work, a free live session on Zomra on 6 October 2026, for business users rather than developers. It covers the AI people already use, what an AI agent actually is, where plain automation is the better tool, and three questions that tell you which one a task needs.
Why most companies are not getting paid back yet
- 88% of organisations use AI regularly in at least one business function, but only 37% see any impact on profit from it (McKinsey, The State of AI in 2026, August 2026).
- Only 5% of US leaders say their business processes are highly prepared for AI agents (Deloitte AI agent readiness survey, August 2026).
- Electricity paid off only once factories were redesigned around it, decades after the first power stations (Paul A. David, "The Dynamo and the Computer", 1990). AI needs the work redesigned around it too.
Warm-up: the AI you already use
Before ChatGPT, AI did one narrow job at a time: recommendations, spam filters, fraud alerts. An LLM is different: it predicts the next word, token by token, and reports, code and strategy documents all come out of that loop. It reads and bills in tokens, and holds a limited context window. When a long chat starts going wrong, start a new one with a short summary.
Prompting in four parts
The same ask, written as Role, Task, Context and Format. Context is the part most people skip.
- Role
- Who it should be. Sets the tone and the expertise.
- Task
- One clear verb, one clear job.
- Context
- What you know and it does not.
- Format
- Language, length, tone, layout.
Most of us still carry AI's output by hand: we prompt, copy, paste into email or Excel, then send and follow up. AI did one step. We did the other four.
Part 1: AI agents
Chat answers you and you do the work. An agent does the work and you decide. An agent is an LLM with tools, memory and a loop: think, act, observe, decide, and repeat until the goal is met.
LLM plus harness equals an agent. The model is rented and the same for everyone; you can swap it any time. The harness is yours: the tools, skills, memory and loop you wrap around the model, and it stays when you change models.
- Tools via MCP
- The Model Context Protocol is the open standard that connects an agent to your apps. Think USB-C for AI: one plug per app. It lets an agent draft the reply, create the file or update the deal, not just search. In Claude and ChatGPT, MCP servers appear as connectors.
- Skills
- A folder with a SKILL.md that tells the agent how your team does a job, plus the files and scripts it needs. The agent reads every skill's name and description, opens the one that matches the request, and follows its steps. Same prompt three times? Make it a skill.
- Harness engineering
- The skill shift from prompt engineering (how you ask) to context engineering (what it knows) to harness engineering: what it can do, how it does it, and how it is kept safe with permissions, guardrails, human approval, test cases and logs.
Few people actually use an agent
- Among people who already use AI, 72% are aware of AI agents, 41% have tried one and 24% use one regularly (Menlo Ventures and Morning Consult, State of Consumer AI 2026).
- An agent thinks at every step, and that has a price: every step is a model call, a job takes seconds to minutes, and the same input can take a different path each run. For work that is the same every time you do not need a thinker. You need a rail.
Part 2: automation
All three run in n8n, one canvas for automations and agents, with 500+ integrations, any model, approvals that pause for a human yes, and self-hosting on your own servers. Zapier, Make and Power Automate work on the same idea.
- Automation
- Rules decide every step. Same path every run. Example: a daily sales report at 8:00.
- AI automation
- Steps in order with AI in one of them, exactly where it is needed. Example: an invoice PDF into the ERP.
- AI agent
- AI decides every step, picks the tools and loops until done. Example: a customer email, answered.
Demo 1: an invoice, end to end
The invoice lands in Drive, an AI step reads supplier, number, date and total, a rule checks for a duplicate, and the row is posted or flagged for review. One AI step, everything else is rules.
Demo 2: a refund, routed to a person
Personal data is masked first, AI drafts a reply from the refund policy, a second model reviews the draft, and a person approves before anything goes out.
Demo 3: a data analyst agent
Ask in plain English, and the agent reads the database, writes and runs read-only SQL, then draws the chart and explains it. Nobody drew those steps; it picked its own tools.
Part 3: three questions that pick the tool
Same steps every time?
Yes: plain automation. Let it run.
Is the input messy?
Yes: add an AI step to a fixed flow. If the next step depends on what it finds, it is an agent.
Is a mistake expensive?
Yes: add a human approval before it acts. A 50,000 EGP discount is a person's call.