Generative AI inside your workflow, with a human signing the work off
The value is rarely in the first idea. It is in the tenth draft of something repetitive, produced against your rules and your source material, and checked by somebody who knows the difference.
What is generative AI?
Generative AI is the use of models that produce text, images, code or documents, built into a business workflow with brand rules, grounding material and a review step. It suits Australian organisations producing the same kind of output repeatedly, such as product descriptions, proposals or campaign variants, where a person still approves what goes out.
Get a fixed written quote- Typical timeline
- 6 to 14 weeks
- What drives cost
- How much of your own content has to be prepared and grounded, how many workflows are in scope, and your expected usage volume.
- Best for
- Repetitive, structured output where volume is the constraint
- You own
- The prompt library, the brand rules, the outputs and the accounts
- Built with
- Prompt systems, brand controls, grounding sources, review queues
Your handover
The jobs generative AI is genuinely good at
There is a shape to the work that pays off. High volume, structurally similar, grounded in facts you already hold, and reviewed by someone who can spot an error quickly. Four hundred product descriptions built from a specification table. A first pass proposal assembled from your approved method statements and last year's scope wording. Twelve ad variants against one offer. Alt text across a neglected image library. Meeting notes turned into a structured action list. In each case the model is compressing a task that was tedious, not inventing something that required taste.
- 01One production workflow built end-to-end
- 02Prompt library versioned in a repository you control
- 03Brand and claim rules encoded as automated checks
- 04Grounding source structure and update routine
- 05Review queue showing sources beside each draft
- 06Written acceptable use policy for staff
- Pilot measurement of time saved and quality scored
- Cost per output tracked at real volume
- Training session for the reviewers and editors
Generic output ranks poorly, reads as filler and quietly costs you authority
The shape that disappoints is the opposite: low volume, high stakes, genuinely original thinking, or work whose whole purpose is that a specific human wrote it. Generic output ranks poorly, reads as filler and quietly costs you authority. So we treat generation as the second step, never the first. Facts, structure and angle come from your business, then the model does the drafting labour, then a person edits. Teams that already have a working content strategy get far more out of this than teams hoping the tool will supply one.
Generic output ranks poorly, reads as filler and quietly costs you authority.
Brand controls are the actual product
Anyone can prompt a model. The engineering is in the constraints around it. We turn your style guide into machine readable rules: sentence length, banned words, the claims you are not permitted to make, the spelling conventions your editors care about, how products are named, and the specific Australian English forms that a model trained largely on American text will otherwise get wrong every single time. Those rules are applied in the prompt, checked automatically after generation, and anything failing a check goes back or gets flagged rather than reaching a reviewer.
More on brand controls are the actual product
For images the same logic applies to palette, composition, and whether generated imagery is permitted at all in a given context. Plenty of organisations decide that generated photography is fine for internal decks and unacceptable for anything depicting their people, premises or products, and that is a sensible line to draw explicitly rather than case by case. We document these decisions as a written policy your team can point to, which matters more than it sounds when a new staff member starts using an unapproved tool because nobody told them what the rule was.
How the engagement runs
How a generative AI build runs
We build one workflow end-to-end before adding a second. A pilot that produces fifty real outputs, measured against what your team would have produced unaided, tells you more than any vendor demonstration and costs a fraction of a platform commitment.
- 01Select the workflowHighest volume, most repetitive, clearest definition of good
- 02Gather grounding sourcesThe specifications, approved wording and factual material outputs must be built from
- 03Encode the rulesTone, banned claims, Australian English conventions, naming and structure
- 04Build the prompt systemTemplates with variables, versioned in a repository rather than pasted into a chat window
- 05Automated checksRule violations, missing facts, duplicated phrasing and reading level flagged before human review
- 06Review queueOutputs presented with their sources so an editor can verify quickly instead of rewriting
- 07Pilot and measureReal volume, time per output tracked, quality scored by the people who own the standard
- 08Roll out or stopExpand where the numbers hold, and abandon the workflows where they do not
Two decisions on your side that keep the project moving
The measurement matters because the gains are frequently smaller than the excitement suggests, and occasionally larger in places nobody predicted. We have seen drafting time fall by more than half on structured output and barely move on anything requiring judgement. Knowing which is which, in your business rather than in general, is the point of the pilot.
Choose the right level
How much review does each output type need
The review burden is what determines whether a generative workflow saves time or just relocates it. If an editor has to fact check every sentence against a source, you have replaced writing with proofreading, and proofreading someone else's confident prose is slower than most people expect. So we classify output types by risk before building anything, and design the queue around that classification.
Output type
01
Internal summaries and meeting notes
Review needed
Spot check
Why
Errors are cheap and the reader knows the context
02
Product descriptions from a specification table
Review needed
Sampled review with automated checks
Why
Facts come from your data, so failures are usually tone
03
Campaign and ad variants
Review needed
Full human approval
Why
Claims made in advertising are your liability under Australian Consumer Law
04
Proposals and quotes
Review needed
Full human approval, always
Why
Numbers, scope and commitments must be verified by the person accountable
05
Regulated or clinical content
Review needed
Subject matter expert plus compliance sign off
Why
The model has no idea which claims your regulator prohibits
How we work this out during scoping
The table below is roughly how we sort a first engagement. It changes with your risk appetite and your sector. A regulated financial services firm will move several rows down a level, and that is the correct instinct rather than excessive caution.
Accuracy, claims and what your data is used for
A model will produce a product specification that sounds right and is not. It will attribute a capability you do not offer, invent a certification, or state a delivery time nobody agreed to. Published as marketing, that is a misleading representation, and the accountability is yours. Grounding every factual claim in a source you control, and showing the reviewer that source beside the draft, is the only version that works at volume. Copyright and provenance deserve the same care, particularly with generated imagery where training data is opaque and indemnity varies by provider.
Retailers publishing at volume should also read our retail work
Then there is what leaves your building. Drafts routinely contain unreleased pricing, client names and personal information, so provider terms about training on your inputs are worth reading rather than assuming. Enterprise arrangements generally exclude your data from training while consumer tiers frequently do not, which is why staff quietly pasting client material into a personal account is a real privacy exposure under the Privacy Act 1988. Where content includes personal information we keep source material and outputs in Australian infrastructure and document what crosses a border, so your privacy officer can assess it. Retailers publishing at volume should also read our retail work.
When generative AI is the wrong fit
If you publish two considered articles a month, this is not your bottleneck and a good copywriter will serve you better and cost less in total. If your differentiation is expertise, generated content actively erodes it, because the thing readers value is the judgement no model has. If your source material is thin, generation produces confident filler and the fastest way to find out is to read the output aloud to someone in your industry.
Volume economics also cut both ways
Volume economics also cut both ways. Generating at scale means paying per token every time, and long grounding documents multiply that quickly, so there is a real crossover point where a template, a snippet library or a properly structured data feed beats a model outright. We look for that crossover before recommending anything. And if what you actually want is one internal assistant your staff can ask questions of, that is a different build entirely and sits with custom GPT solutions rather than a content pipeline.
How we scope it
Four ways to scope your Generative AI project
We do not publish package prices, because the same brief can be a short build or a long one. These are the shapes the work usually takes. Tell us which one sounds like you and you will get a fixed written quote that spells out exactly what it covers.
Proof of value
One use case, evaluated honestly before it goes near a customer
Fixed written quote, agreed before work starts
- One production workflow built end-to-end
- Prompt library versioned in a repository you control
- Brand and claim rules encoded as automated checks
Production build
In production, with a human approval step and an evaluation set
Fixed written quote, agreed before work starts
- Everything in Proof of value
- Grounding source structure and update routine
- Review queue showing sources beside each draft
- Written acceptable use policy for staff
Embedded platform
Built into the product rather than bolted onto it
Fixed written quote, agreed before work starts
- Everything in Production build
- Pilot measurement of time saved and quality scored
- Cost per output tracked at real volume
- Training session for the reviewers and editors
Model care
Monitoring, evaluation and retraining as the inputs drift
Rolling monthly, quoted in writing
- Evaluation set rerun as the model and the inputs change
- Cost and quality reported monthly, not assumed
- Prompt, tool and guardrail changes as the work shifts
- Rolling, cancel with 30 days notice
These are shapes, not menus. Most quotes end up somewhere between two of them, and we will say so when the honest answer is the smallest one. Describe the problem and we will tell you which it is.
Questions buyers usually ask
Frequently asked questions
How long does a generative AI project take?
Around 6 to 14 weeks for a first workflow taken to production. Encoding the brand rules and assembling grounding sources takes longer than building the prompt system, because it forces decisions your organisation may never have written down. A tightly scoped pilot on a single output type can be running in a few weeks if the source material is already structured.
What will it cost, and what are the ongoing costs?
The build is quoted in writing after scoping, driven by how many workflows, how much rule encoding is required and which systems the output has to flow into. Ongoing you pay the model provider per token, so cost scales with volume and with how much grounding material each request carries. We measure cost per output during the pilot so the comparison against your current process is concrete.
Who owns the prompts and the generated content?
You own the prompt library, the rule sets, the grounding material and the outputs, and the provider accounts sit in your name. Copyright in purely machine generated material is legally unsettled in Australia, which is a reason to keep meaningful human authorship in anything you need to protect. We flag which outputs fall into that category during scoping.
Will Google penalise us for AI assisted content?
Search systems target unhelpful, low value content rather than the tool used to produce it. Mass generated pages with nothing original in them do perform badly, consistently and deservedly. Grounded, edited, genuinely useful content performs on its merits. The practical test is whether a person with real expertise would sign their name to the page, and if not, do not publish it.
Can our staff just use a consumer AI tool instead?
Many already are, which is the problem. Consumer accounts often permit training on inputs, sit outside your privacy controls and leave no record of what was shared. A sanctioned workflow with proper terms, logging and brand rules is usually less about capability and more about removing that exposure while making the good uses easier than the risky ones.
How do you handle advertising claims and compliance?
Prohibited claims are encoded as automated checks, so an output containing one is blocked before a reviewer sees it. Anything customer facing keeps mandatory human approval by the person accountable for the claim. For regulated sectors we add a compliance step in the queue and keep an audit record of who approved each item and when.
Related services
Start with one workflow, measured properly
Tell us the output your team produces most often and how it is checked today. We reply within one business day with a scoped pilot and a fixed written quote.