Generative AI Consulting

Sometimes, and a lot of the value of a consultation is hearing when it is not. Generative AI is extraordinary at open-ended language work: drafting, summarising, extracting meaning from messy text, answering questions over documents, translating between formats. It is a poor and expensive choice for problems that are really about structured data, exact calculation, or repeatable rules — a forecasting problem is usually classical machine learning or plain statistics, a routing problem is often just logic, and a numeric total should never be produced by a language model when a database query is right every time. The failure pattern we see most is a business reaching for a large language model because it is the thing in the news, when the actual problem is deterministic and would be better, cheaper and more reliable solved without one. We will tell you which category you are in before you spend anything.

You do not eliminate it — you engineer around it, because hallucination is not a bug that gets patched, it is a property of how these models work. They generate statistically plausible text, and a confident fabrication is extremely plausible text. The mitigations are architectural, not a matter of a cleverer prompt. Retrieval-augmented generation grounds answers in your actual documents rather than the model’s memory, so the system quotes a real source instead of improvising. Evaluation harnesses catch regressions before they reach users. Confidence thresholds route uncertain outputs to a human. Citations let a reviewer verify a claim in seconds. And for anything consequential, a person approves before it commits. Any vendor who tells you their model does not hallucinate is either misinformed or selling; the honest question is not whether it happens but what the system does when it does.

Usually a mix, and the decision is per use case rather than a single company-wide bet. For general staff productivity, an enterprise seat of an established assistant — with the data controls configured properly — is almost always the right answer, and building your own version of it would be a waste of money. For a workflow that depends on your own documents, data and systems, retrieval and orchestration wired into your environment beats a generic product that cannot see your context. For something genuinely core to your business or your differentiation, a custom build can be worth it. Where firms go wrong is at the extremes: building a bespoke chatbot to do what a $30-a-month product does perfectly well, or trying to run a mission-critical, regulated workflow on a consumer tool with no data processing agreement. We are vendor-independent and take no reseller margin, so the recommendation follows the use case, not a partner tier.

Evaluation is how you know the system works and keeps working, and it is the part most rushed generative-AI projects skip entirely. A demo proves the thing can work once; an evaluation harness proves it works across the range of real inputs, and tells you the day it stops. It means assembling a representative set of real cases with known good answers, scoring the model’s output against them systematically, and re-running that suite whenever anything changes — a new prompt, a new model version, a provider quietly updating the model underneath you. Without it, you are flying blind: quality drifts, nobody notices, and you find out from a customer. With it, model upgrades become a measured decision rather than a gamble. It is unglamorous and it is the single strongest predictor of whether a generative-AI system survives contact with production.

It is the right question to ask first, and the answer depends entirely on the setup rather than on “AI” in the abstract. Every generative-AI system has a data path: where prompts go, whether the provider retains them, whether inputs are used to train their models, which jurisdiction the infrastructure sits in, and who at the vendor can see the content. Those facts belong in a data processing agreement, not in a sales pitch, and they differ enormously between a consumer chatbot and a properly configured enterprise or cloud deployment. For Australian businesses the relevant frame is the Privacy Act 1988 and the Australian Privacy Principles, plus any sector rules you operate under. We are not your lawyers and we do not certify your compliance — what we do is establish the actual data path so the people who make that call are deciding on the facts rather than the marketing page.

It starts with a free consultation about the problem, not the technology, where a frequent outcome is us narrowing an over-ambitious idea down to the one use case worth doing first. If there is a real opportunity, the AI Opportunity Audit at around $3,000 maps the candidate use cases, the data behind them, the constraints on each and a ranked shortlist — and you keep it whatever you decide next. From there we build: use-case selection, model and build-versus-buy calls, retrieval on your own data, a real evaluation harness, guardrails, and the thing running in production rather than a slide about it. The audit typically earns its fee at around 20 or more staff. A smaller team usually needs one well-configured off-the-shelf tool and some training, and we will say so plainly.