AI GOVERNANCE
A capability you cannot move is one you are only renting. Price the exit before you celebrate the launch.
The Write-Off: A ROSE by any other vendor
THE WRITE-OFF · ISSUE 006
A Tasmanian council lost an AI phone agent that was handling about a third of its calls when the supplier stopped supporting it, and now faces an unbudgeted rebuild of up to $97,800. The AI did its job. Nobody had planned for the day it stopped.
TODAY’S STORY
Pulse Tasmania
·
20 November 2026
Devonport City Council, in north-west Tasmania, put its AI agent ROSE on the phones on 12 December 2025. It was soon handling about 35% of customer calls: checking bin collection days, answering rates questions, booking callbacks and routing people to the right team. The vendor, Cassie AI, promoted Devonport as the first Tasmanian council to use its technology.
On 4 September 2026, the supplier, FormsExpress/Cassie AI, stopped supporting the service. According to council papers reported by Pulse Tasmania, calls to customer service staff rose by just over 40% once ROSE disappeared. The reason the support ended has not been reported.
On Monday 23 November, councillors will vote on a $64,900 proposal to rebuild ROSE on xAI’s Grok, with Microsoft Azure hosting the council’s own systems. The money is not in the budget. A single consultant, the same one who helped build the original, would do the work at $1,200 a day over an estimated 37 to 81.5 days, so anywhere from $44,400 to $97,800. A Microsoft-based rebuild was costed at $77,400. The council expects running costs to fall from $21,570 a year to around $5,230, with Grok’s voice service priced at 8 US cents a minute against the old supplier’s $1.
THE VERDICT
Devonport did plenty right. It started small, kept humans available and measured the load ROSE carried. Where it slipped was ownership. The capability lived on a supplier’s platform, so when the supplier stepped away the capability left with it, and the bill for getting it back arrived about eleven weeks later as a day-rate estimate whose top end is more than double its bottom. Any software business running a single-vendor agent in front of customers carries the same exposure, usually without a council meeting to make it public.
The service ran on the supplier’s platform, and the council papers describe a stop followed by a gap of more than two months before a replacement reached the agenda. Whatever the contract said about notice, the council was left without a working fallback, a portable build or a priced plan B. Dependency risk was accepted by default rather than by decision.
The council knew ROSE was taking around a third of its calls. What appears not to have been modelled was the reverse case: the cost of absorbing those calls overnight. A 40% rise in human call volume is real staff time, longer queues and irritated residents, and none of it shows up in the rebuild quote.
Much of ROSE’s work was looking things up. Bin days, rates information and routing are deterministic questions with deterministic answers. At $1 a minute, the old platform cost $21,570 a year to answer them conversationally. The proposed version cuts that sharply, which is welcome, and also suggests the original price was never tested against the best non-AI alternative, such as a lookup page, an SMS service or a simple phone menu.
The rebuild moves control in-house, which is the right instinct. It also rests on one consultant, on a day rate, with the proposal itself conceding that “the council carries the variance”. Swapping a single platform vendor for a single model vendor and a single contractor lowers the bill without fully lowering the dependency.
Resilience by design means treating supplier exit as a planned event, like a lease ending, and paying for it upfront while you still have leverage. The goal is a service that can lose its model, its vendor or its builder and degrade gracefully instead of vanishing.
Write the exit before the go-live. Require notice periods, export of prompts, knowledge base, call flows and transcripts, and paid transition support, so a supplier’s departure becomes a project with a timetable.
Keep a deterministic floor. Serve bin days, balances and opening hours from a simple lookup any channel can call. If the AI goes, customers lose the conversation and keep the answer.
Chart the load the AI carries. Track deflected calls and human call volume on a process behaviour chart. You will see the value while it lasts, spot drift early and know the staffing cost of losing it.
Own the knowledge, rent the model. Keep the knowledge base, test scripts and integration logic in your own repository, so changing model provider is configuration work instead of a rebuild.
Cap the rebuild. Buy fixed-price milestones with acceptance tests, and put the variance on the party who controls the estimate.
Devonport deserves credit for being early and for publishing its numbers, which hands everyone else the lesson for free. Monday’s vote is the chance to buy the second ROSE with the exit plan the first one lacked. For PE-backed software firms with agents in front of customers, the cheaper move is to ask the same question this quarter, before a supplier answers it for you.
THE QUESTION FOR MONDAY’S LEADERSHIP MEETING
If the honest answer is “we’d find out on Monday”, the board has its homework.
NEWTHISTLE NUGGET
AI GOVERNANCE
A capability you cannot move is one you are only renting. Price the exit before you celebrate the launch.
The Write-Off: A ROSE by any other vendor
SOURCES
Devonport City Council considers $65,000 rebuild of AI phone agent on Elon Musk’s Grok, Pulse Tasmania, 20 November 2026
Devonport City Council case study, Cassie AI (vendor page, background only), accessed 21 November 2026
Council Meetings, Devonport City Council, accessed 21 November 2026