Capability ยท Agentic AI

AI agents that do real work, with a human signing off.

Agentic AI means software that plans steps, calls tools and completes a task, not just answers a question. I build agents for real processes: discovery with the people who do the work, deterministic checks first and the model where judgement is needed, a written evaluation bar before release, and a human sign-off wherever money or a regulator is involved.

I do this work as AI & Automation Lead at Dexterous Group in Melbourne. This page shows the capability; it is not an offer of personal services.

How I approach it

The method, step by step.

  1. Discovery with the people doing the work

    One person at a time, on live files. Each task broken down to the decisions inside it, with how often it happens and how long it takes.

  2. Deterministic first

    Rules and checks do what rules can do. The model handles the judgement in between. The model is never the control.

  3. A graduation bar

    An agent graduates on real cases: a set accuracy, a reviewer acceptance rate and a written specification, agreed before the build.

  4. Run it like a team

    Every agent has a job description, limits, a log of what it did and someone it reports to.

Work

Where I have done it.

Anonymised: no client, employer product or confidential detail.

Problem
Quarterly compliance preparation took hours of senior review in an accounting practice.
What I did
Discovery with 18 staff (one reconciliation broke into 22 activities), then deterministic checks with an LLM for judgement and human sign-off.
Result
71 evaluation tests; it caught a wages-payable error the manual checklist missed.
Stack
Python, openpyxl, Xero API and MCP, Claude, n8n
Problem
One person running several businesses needed help with sales, content, finance and IT.
What I did
Built a personal AI operating system: 29 agents with roles, schedules, approvals and an audit trail, on local models first.
Result
Daily work across content, finance, IT and security handled by agents, with the owner approving what matters.
Stack
Python, TypeScript, Ollama, Claude, MCP, Postgres
Tools and standards

What I work with.

  • Claude and OpenAI models
  • Local models on Ollama
  • MCP
  • n8n
  • LangGraph
  • Python and TypeScript
  • Evaluation harnesses
Questions

Frequently asked.

What is agentic AI?

Agentic AI is AI that can plan a task, use tools and systems, and take steps towards a goal, rather than only producing text. In business it is used for processes such as invoice handling, reconciliations, support and reporting, usually with a person approving the important steps.

What is the difference between an AI agent and automation?

Automation follows fixed rules. An AI agent can handle variation and judgement, such as reading an unusual document or deciding which rule applies. The safest designs use automation for everything rule-based and the agent only for the judgement.

How do you know an AI agent is ready for production?

Agree a bar before building: a set of real cases, the accuracy required, how often reviewers must accept its work, and who signs off. If it does not meet the bar on real cases, it is not ready.

Can Amjid Ali build AI agents for my business?

Amjid builds agents in his role as AI & Automation Lead at Dexterous Group; he does not take personal consulting work. For projects, contact Dexterous Group.

Read more

Insights on this.

Other capabilities

Also on this site.

For project work, contact Dexterous Group. For speaking, media or a peer conversation, email mail@amjid.au or message me on LinkedIn.