AI development that starts with a task, not a technology

Which AI service fits which problem

Chatbots answer questions from a defined body of knowledge and hand over to a person when they cannot. They suit high volume, repetitive enquiries. AI agents go further and take actions across systems, such as triaging an enquiry, checking availability and drafting a reply, which raises the stakes because a mistake now has consequences beyond a wrong answer. Voice AI handles inbound calls and after hours enquiries for clinics, trades and service businesses that lose bookings to unanswered phones.

Document AI reads invoices, forms, contracts and clinical paperwork and turns them into structured data, which is often the highest return option because the task is well-defined and currently manual. Generative AI supports content and image production at scale. Custom GPT solutions package your own knowledge into an assistant your team can use. Machine learning applies where you have enough historical data to predict something specific, such as churn or demand, and it needs real data volume rather than enthusiasm.

What an AI engagement looks like

We start by picking one task, measuring how long it currently takes and how often it is done, then building the smallest thing that could handle it. Before anything goes live we build an evaluation set: a few hundred real examples with known correct answers, so we can measure accuracy rather than form an impression from a demo. Demos are easy. Consistent performance on the awkward twenty percent is the whole job.

A first deployment typically runs 3 to 8 weeks. Everything goes out with a confidence threshold and a defined escalation path to a human, because a system that says it does not know is far cheaper than one that invents an answer. We also decide up front what data leaves your environment. If you handle personal or health information covered by the Privacy Act 1988 and the Australian Privacy Principles, that decision needs documenting before a single record is sent to a model provider, not afterwards.

• One task chosen, with its current cost in hours measured before we build
• An evaluation set of real examples with known correct answers
• Confidence thresholds and a human escalation path on every deployment
• A written record of what data goes where, for your privacy obligations
• Model choice kept swappable, so you are not locked to one provider

The expensive mistake: buying capability without a task

The pattern is familiar. Leadership decides the organisation needs AI, a pilot gets funded, something impressive is demonstrated, and eighteen months later nobody uses it because it never attached to work anyone actually does. The failure was in the framing. There was no task, no baseline and no way to tell whether the thing was working.

The second failure is skipping the boring alternative. A surprising share of the problems presented to us as AI problems are solved better and more cheaply by deterministic automation, a corrected process, or writing down the twenty answers customers keep asking for. AI is worth its cost when the input is genuinely unstructured, language or images, and the volume is real. When the input is structured and the rules are knowable, use automation and keep the budget. We will say which one you are looking at before quoting.