Here is our second part of Art Expert Michael Reid OAM’s guide to Build an AI Ready Gallery on Women Love Tech.
Part One of this case study ended with a simple conclusion: before artificial intelligence can be useful, a business must first put its own house in order. For Michael Reid Galleries, that meant replacing a sprawling web of spreadsheets with one structured source of truth.
The next question was more interesting. Once the foundations were being built, what should AI actually do?
Over recent weeks, our two AI wizards, Luca and Will, have presented the first five operational workflows developed specifically for the galleries to our Executive Team. None was theatrical. There was no robot curator, no machine deciding whether a painting was good, and no attempt to replace the relationship between a gallery and a collector.
The workflows dealt with ordinary gallery friction: building exhibition and collector catalogues, recording client conversations, identifying which collectors to approach, finding enquiries buried in the inbox, and setting up new collector profiles.
That lack of glamour is precisely the point.
Each workflow was approached in the same practical way. Map what we do now. Identify where it stalls. Build the AI-assisted version, run it live, test how it might fail, install it with the colleagues who will actually use it, and retain the underlying code so the gallery is not left with a mysterious black box.
The catalogue generator is the most visible example. A colleague tells the system whether the document is for a single artist, a group exhibition or a client proposal. It gathers the images, builds the InDesign file, and leaves approval and sending with the human. The internal estimate reduces a process modelled at seven hours and 35 minutes to one hour and 20 minutes. That remains an estimate. The more important question is whether the system makes up or alters an artwork’s title, price or medium; drops or duplicates a work; or drifts away from the gallery’s house style. In ten test proposals, none of those failures appeared.
The second workflow records client follow-up. After a call, text message or conversation, a colleague can dictate what occurred. The system records the note against the correct person in Salesforce and creates the follow-up task. It was tested for same-name clashes, vague dates such as “next Thursday”, made-up activity and duplicated tasks.
The larger gain may be behavioural. If recording a conversation becomes easier, more conversations will be recorded. Information that once remained in someone’s head becomes part of the gallery’s shared knowledge.
Collector discovery is the third workflow. A colleague sets the criteria, the system queries Salesforce and ranks collectors using the affinity information already held. It respects do-not-contact settings, identifies the exact artist, and admits when there are fewer suitable matches than requested rather than padding the list. Matching those collectors directly against available inventory depends on the new register connection and remains part of the next phase.
The fourth workflow searches for buried leads. Once a week, it scans the previous seven days of emails, looks for genuine buying intent and produces a ranked report for human action.
Enquiries do not always arrive with the subject line, “I would like to buy this artwork”. They sit inside a long exchange, a website form, a social-media response or a casual question that was not followed up. In testing, the system found the buried enquiries while rejecting phishing attempts and ordinary inbox noise.
The machine finds and ranks. A person decides what to do.
The fifth workflow handles collector onboarding. It captures contact details and context, checks for duplicates, asks for approval before writing to Salesforce, and verifies artist tags rather than guessing. The time saving is not dramatic. The consistency may be.
This first implementation round has clarified what useful AI looks like inside a gallery. It works best where a task is known, necessary and repetitive, but awkward enough that people sometimes postpone it. AI lowers the cost of completing the task. That can change not only the speed of the work, but whether the work happens at all.
The guardrails matter just as much as the automation. No unapproved writing to the client database. No contacting people who have opted out. No made-up or altered artwork facts. No list padded with unsuitable names. And no pretending the system cannot fail.
When it does fail, the instruction is practical: identify the issue, report it and fall back to the manual process.
Nor is implementation simply a matter of switching on AI and walking away. Each workflow requires installation, access to the relevant accounts, a short session with colleagues, monitoring, and a clear route for reporting problems. One workflow is still waiting on the inventory-register connection.
The projected time savings will also need to be measured in daily use, not repeated as fact simply because they appeared in a presentation. We have modelled the hours. We have not yet banked them.
That distinction matters.
AI discussions are full of claims about what might be possible. Implementation is less exciting and far more useful. It involves testing, correcting, retraining, listening to colleagues and occasionally admitting that the manual process still works better.
Part One was about creating order. Part Two is about attaching AI to specific pieces of work.
The early lesson is plain: the value is not in having AI somewhere in the business. It is in choosing the right task, defining it properly, testing the failure points and keeping a human decision at the end.
Speed remains the story.
But speed without accuracy, consent and control is simply a faster way to make a mess.
Check out Michael Reid Gallery Sydney here: https://michaelreid.com.au/sydney/
