There is a great deal of noise around artificial intelligence. Much of it is speculative. Some of it is evangelical. Most of it is either wildly optimistic or deeply alarmist. What is often missing, amidst these swings between opposites, is any practical detail. For me, the question is straightforward: how do you implement AI across my art galleries?
This article is the first in a case study, hopefully answering that question. Part-One deals with the groundwork. The rest will look at the implementation itself: what worked, what failed, what colleagues embraced, what they resisted, and whether the efficiencies promised by AI proved to be real.
The starting point was not glamorous. Like many galleries, my colleagues and I have spent years operating across a sprawling network of spreadsheets. Exhibition lists, stock registers, artist records, shipping notes, catalogue requests and pricing sheets had accumulated over time into a dense operational web. The system functioned, but largely because my colleagues invested substantial time navigating it.
This is where the practical story of AI begins. Artificial intelligence is only as useful as the information it can access. If the underlying data is fragmented, inconsistent and buried, the machine will not impose order. It will simply process the confusion more quickly. That is not the same. For the old computer saying applies with some real force: “garbage in, garbage out”.
Before automations, before agents, before any of the fashionable language surrounding AI, the gallery first needed to consolidate its information into one structured environment. Every artwork, every artist, every exhibition, every image and related document, every shipping quote, detail and contact all sitting in a single coherent system. One source of truth.
The practical exercise unfolded in a series of straightforward steps.
First, we identified every spreadsheet used across the business. Not only the main stock registers, but exhibition sheets, shipping logs, catalogue request spreadsheets, personal sales trackers and various ad hoc documents that had become embedded in day-to-day operations.
Second, we examined how the gallery actually works. How an exhibition is uploaded. How an editioned photograph is recorded for sale. And, how an artwork moves from an artist to a gallery, to an art fair, to a collector and, finally, to installation. Where images are stored. Where PDF attachments such as consignment agreements and condition reports are kept.
Third, we designed a single database to hold this information in a consistent format. One Ring to Rule Them All.
Fourth, we planned the migration of all existing data under one guiding principle: “capture everything, lose nothing”. Records are cleaned where possible, but nothing of value is discarded. Ambiguous information is flagged for human review.
Fifth, a colleague friendly web application is built over the database so the team can search, edit and update records through one interface, rather than continually reverting to spreadsheets.
Only after these steps does AI become genuinely useful — to my galleries.
In our case, we interviewed a number of people working in the AI and the automation space before engaging two young guys from Sydney University. Luca and Will, were not selected because they offered the grandest vision of artificial intelligence. They were selected because they approached the task in a disciplined and practical manner.
Their proposal was refreshingly straightforward: replace our sprawling spreadsheet-based inventory system with a single structured database and a colleague-facing application that reflected the way the gallery actually operates.
From the outset, clear guardrails were imposed. The engagement was fixed in scope and fixed in price. The objective was to build the underlying data architecture, not to embark on an open-ended and expensive exploration of hypothetical possibilities. All infrastructure, code and documentation were to be owned by the gallery from day one — fully transferable to another developer if required. In short, the gallery retained control.
The implementation was modular. Systems such as Salesforce, Xero, website integrations, shipping workflows and more advanced AI agents were deliberately excluded from the first stage. They can and most definitely will be added later, but only once the core inventory system is stable and functioning. The existing spreadsheets remained the operational system during the build, with colleagues continuing to work as normal until the new platform is fully tested and ready.
The technical team was required to test the application against real workflows and real staff behaviour. The software must adapt to the gallery, not the gallery to the software. And the migration principle was clear from the beginning: “capture everything, lose nothing”.
These guardrails were critical. They kept the project grounded and reduced the risk of technology overreach. Most importantly, they ensured that the gallery remained in control of its own systems and data.
This is not an exercise in replacing judgement. It is an exercise in removing friction. A gallery runs on knowledge: where a work is located, where it has been shown, whether it has sold, what price history it carries, which collectors have responded to similar works, and what documentation sits behind it. Historically, that knowledge has often been spread across people, emails, spreadsheets and memory. Once it is structured properly, it becomes available immediately, both to staff and to machines.
That is the inflection point. Artificial intelligence can then interrogate the data directly. It can compare works, trace exhibition histories, surface patterns and answer operational questions in seconds. The inventory register ceases to be a passive archive and becomes a system that allows the business to reason across its own information.
The technology, however, is not the hard part. The discipline is. Duplicated artist names, incomplete records, incorrect stock statuses and years of ad hoc notes become immediately visible. Humans are remarkably good at working around disorder. Machines are not.
Implementation is therefore as much cultural as technical. A badly designed database is simply a new form of friction. If staff find the system cumbersome, they will revert to old habits, and nothing will change. Success depends on testing the system against real behaviour: how exhibitions are uploaded, how stock moves between galleries, how images are attached, and how information is updated in the midst of daily interruptions. The software must adapt to the gallery, not the gallery to the software.
This is where the broader significance becomes clear. Historically, inventory systems were designed to record information. Increasingly, they are becoming intelligence systems. They do not simply store the past; they allow patterns to emerge, relationships to surface and weaknesses to become visible.
The practical implications are considerable. Exhibition planning becomes more analytical. Client engagement becomes more precise. Administrative tasks that once consumed hours can be reduced to minutes.
Speed is the story. Not speed for its own sake, but speed as the result of clarity. The businesses that organise their information in a form machine can read will move faster, decide better and serve their clients more effectively. Those that do not will increasingly feel heavy and slow.
The central lesson is straightforward. Before artificial intelligence can be of any practical use, a business must first put its own house in order AND THEN PLACE AI ACROSS ALL. The One Ring to Rule them All.
The Part Two report — on what happened when our underlying information database went live — is for the future
