AI Isn't the Magic Fix for Bad Processes
· 3 min read · Nic Keating
Many "AI deployment" briefs are integration projects in disguise. Before layering an agent on top of a broken workflow, ask whether an upstream fix would make the task disappear entirely — and why the boring process improvements so rarely get done.
My daughter was only three days old, and I had already recorded our details in three separate forms.
Somewhere, a paper admissions form contains a typo I cannot recall or correct; it will remain in a filing cabinet for the next decade. While we drive AI adoption across multiple organisations, I notice the easy process fixes, the low-hanging fruit, that don't need AI.
The Unglamorous Backlog: Higher Value Than Most AI Pilots?
New Zealand has executed some of this work effectively. SmartStart allows you to register a birth and apply for an IRD number via a single online form. It is efficient.
Duplication persists after leaving the hospital; Plunket collects the same data during the first home visit. Subsequently, you must complete another PDF, often more cumbersome than paper, to register them with a GP.
The actual integration connecting these systems is a physical book, the Well Child Tamariki Ora book, which parents carry between appointments for providers to update by hand. We have effectively outsourced the system integration challenge to mothers. This admin burden — keeping systems that don't talk to each other in check — falls on the primary caregiver. Every duplicated form is a cost transfer from the institution straight to the household.
Each of these organisations is competent in isolation, yet they remain disconnected. The gaps between them are unowned, which is why nothing changes.
Why the Form Survives
This is not an AI problem. The maternity workflow does not require a Large Language Model; it requires integration. That integration is fifteen years overdue. No AI agent layered on top will be effective, either, because there is no clean state to act from. My daughter's name exists in five separate databases, none of them authoritative, and three of them are slightly different.
Many "AI" or "agent deployment" briefs I have received over the past twelve months have been integration projects in disguise. Clients seek intelligence, but the fundamental requirement is de-duplication. These issues are technically simple, following the "collect once, use many" principle, but politically expensive because performance is rarely rewarded for 'boring' process improvements.
The Choice: Buying an Ambulance or Building a Fence?
In many AI engagements, I observe use cases where AI is tasked with compensating for process failures that organisations have chosen not to solve: re-keying data between silos, summarising identical weekly status reports because upstream ownership is missing, or converting files because a supplier provides the incorrect format. While AI has countless legitimate use cases, wallpapering over process issues is not one of them.
This pattern made me reflect on my own practice. I realised I had similar issues in my own workflows, and I occasionally catch myself when using AI to automate a process that is fundamentally broken. So I had the idea to add a small instruction to every AI I configure. It contributes more to my productivity than any feature I have deployed this year.
Before you complete a task I assign, perform two checks.
1. Is there an upstream process improvement that would render this task unnecessary? If the work is repetitive, such as converting files, formatting reports, re-keying data, or summarising recurring documents, identify what should be fixed at the source so the request becomes obsolete.
2. Am I asking you to act as 'the ambulance at the bottom of the cliff'? If the true issue is structural — a broken form, a missing integration, or disconnected systems — state this clearly. Suggest the 'fence' that should be built instead and the appropriate fix. Complete the task, but explicitly flag the underlying gap.
The first time I applied this to my Copilot configuration, it challenged a request to convert ten supplier PDFs into a standard report. The supplier was using PDFs because their internal form was a static download. The solution was not a superior conversion script; it was a web form.
Much of the integration backlog is hidden here. AI is exceptionally useful for identifying these gaps if given permission to challenge the brief. The default setting is compliance; instead, ask it to pursue excellence. AI's best use is not in automation, but in demanding organisational excellence.