What agentic AI inside a city tourism CMS really automates

As of November 2026, a destination site with agentic AI in its CMS still leaves the same three decisions with a person: what gets published, whose listing goes live, and who answers the enquiry. The rest — drafting, tagging, routing, reporting — is where ai automation for business work actually pays, and the split between those two groups is what this article sets out.

Campaign Brief reported on November 9 2026 that The Brand Agency has launched a rebuilt destination marketing platform for the City of Perth, headlined by a re-imagined visitperth.com and a self-service portal for local operators. The platform runs on Optimizely CMS 13, with Digital Asset Management, the Content Marketing Platform and Optimizely's Agentic Platform built in, so generative and agentic AI sit directly inside content and campaign workflows. The Lord Mayor's statement ties the rebuild to the Capital City Plan 2036.

That description is a supplier announcement, not an audit. Read it as a list of capabilities the City now owns and has to operate. The question a marketing lead in another city or a multi-operator tourism business should ask is narrower: of the work this deployment touches, which parts can run without a person, which parts cannot, and what has to be true before the first group is safe to switch on. It is also the question behind most best lead generation procurement — a platform that captures enquiries but cannot show what happened to them has automated the easy half.

Key takeaways

• The workflow the City described has two halves: operator-submitted content going through a City-managed publishing approval, and structured content feeding the public site. Automation helps most in the first half's routing and the second half's consistency.
• A CMS AI feature automates tasks inside its own box. Anything crossing into ticketing, booking, CRM or a partner's system is an integration project with its own owner and failure path.
• Automated routing is measurable; automated judgement is not. Approval rate, rework reasons and time-to-publish are the indicators that tell you whether an agent is helping.
• Operator-side response time, not page traffic, is usually where a destination platform leaks enquiries.
• The people question is the one a course buyer is really asking: the skill is defining a task, its inputs and its stop points, not operating one tool.

What the City of Perth platform actually changed

The announcement names a clear structure, and it is worth taking apart because most tourism bodies are heading toward the same one.

There are two integrated systems. One faces local retailers, hospitality venues and event organisers, who submit and manage their own listings and content. The other faces visitors, and carries the events, dining, retail and cultural calendar. Content flows from the first into the second through a publishing workflow the City controls. The agency's technology lead framed the architecture as exactly that split: a frictionless portal for businesses on one side, rich structured content on the other, with multi-site flexibility and a complex event taxonomy handled inside the CMS.

Why the taxonomy matters more than the interface: an event listing is not a page. It is a record with a date, a venue, a category, a price band and an operator, and it has to appear in several places — a calendar, a district guide, a themed collection — without being retyped. The occasion-based curation the agency describes depends entirely on those records being structured the same way when they arrive. If operators submit free text, curation cannot be automated; if they submit to a schema, most of the curation is a query.

That is the first real automation decision, and it is made at the form, not at the AI layer. A second one follows it: an operator submitting a listing is an early-stage enquiry, and ai automation for business planning works better when intake, review and outcome are treated as one pipeline rather than three tools.

Where the automation line sits

I separate the steps by what happens when the automation is wrong, not by how clever the technology is. A wrong draft costs an edit. A wrong publish costs a listing, a relationship with an operator, or a public correction.

Two rules fall out of the table.

The first is that automation should stop at the point where the cost of being wrong becomes irreversible. Drafting is reversible. Publishing is not. That is why the City's publishing workflow is the right place to keep a person, and also why the field validation on the operator form is the right place to spend automation effort — a listing rejected at submission never reaches an editor.

The second is that every automated step needs a defined owner for its failures. An agent that routes a listing to the wrong editor has not failed loudly, it has failed quietly. Someone has to be responsible for the queue, and the queue needs an age metric, not just a count. Where the intake feeds a commercial funnel, the same discipline is what separates best lead generation from a form that fills a spreadsheet.

Unisoft ai 課程: what an automation course has to teach you

People searching for an Unisoft ai 課程 are asking a version of the same question the City's editors will face: what do I need to know to run this? The honest answer is that the tool names in any course syllabus will be out of date before the certificate is framed. What does not go out of date is the skill of writing down a task precisely enough that someone else — or something else — can execute it and be checked.

Four things a course should make you do, and which you can test on your own platform this week:

1. Write one process as a sequence of steps, each with a stated input, a stated output and a stated stopping point.
2. Name the point in that sequence where a person must look before the next step runs, and say what they are looking for.
3. Define one number that would show the process is working — time from submission to publish, or share of listings rejected on first pass — and the time range it is measured over.
4. Describe what happens when a step fails partway: retry, alert a named person, or stop and hold the record.

If a syllabus cannot produce those four outputs by the end, it is teaching software operation. The difference shows up on the first day the vendor updates the product. That is why ai automation for business projects usually begin with a written process map rather than a tool choice.

This is also why the City's workflow language is more instructive than any feature list. A portal that lets operators submit content, a workflow that lets the City approve it, and a CMS that keeps the records structured — that description already contains the task definition, the stop point and the review gate. The AI is the part that fills the gaps between them.

Which numbers to watch after go-live

The announcement does not publish performance figures, and I would not trust any that arrived before a full event cycle. What it does give you is a set of steps that produce their own data from the first day.

For the operator side, the useful measures are the share of submitted listings that pass validation on the first attempt, the median time from submission to publication, and the count of listings rejected for the same reason twice. That third one is the only one that tells you whether the guidance improved or the operators simply gave up.

For the visitor side, page views are the least informative number in the set. A destination platform exists to move an enquiry to an operator, so the chain to measure is view to enquiry, enquiry to operator acceptance, and acceptance to booking. If the platform cannot see the third step, it cannot say whether the automation helped anyone — and that gap is where ai automation for business programmes most often stall.

Attribution is where this usually breaks. A visitor reads a district guide on a phone, books by WhatsApp two days later, and pays at the venue. No CMS report will join those three events without a shared identifier carried from the click through to the booking record, and without the operator sending the outcome back. That is a data design decision made before launch, not a report added afterwards.

One caution on the AI features themselves. A generation tool's own logging tells you that a draft was produced, not that it was good. Review pass rate and the reasons for rework are the indicators that survive a vendor's reporting change, because they come from your editors, not from the product. Any best lead generation report should carry the same caveat: it defines the funnel it measured, and its own reporting period.

What Topkee builds above and beside a CMS

For a platform of this shape, Topkee's position is the agent and integration layer rather than the CMS, and it is worth being precise about where that layer starts. Topkee's custom AI agent development packages are scoped by workflow: the trigger, the inputs and outputs, and the human confirmation points are defined before anything is built, and delivery moves through fetch, analysis, human review and handover. The stated design rules are that sources are traceable, execution is permissioned, results are verifiable and delivery is handover-ready — which map one to one onto the review gate the City of Perth described around its operator portal.

In practice the components sit in different places. A visual CMS carries the destination, event and guide content, with templates, multilingual pages, forms and SEO settings; the Travel X template supplies a tourism site structure to start from. YME Chat Agent covers website chat, WhatsApp and Messenger, identifies intent, guides data capture and escalates to a person by stage. YIS Feed turns an article into channel posts and schedules them, drafting by default with scheduling requiring explicit authorisation. Leadbox and TTO close the attribution loop by carrying one identifier from the ad click through to the lead record.

The parts a city platform most often underestimates are outside all of that: ticketing, booking, membership and point-of-interest catalogues are not generated by a site builder, and need back-office configuration or separate build work. Custom agent work is quoted per scenario after an assessment, with a fixed total confirmed only then — scope, the number of system integrations and the review flow drive it, not a licence tier.

Recommended next action

Take the process you are most tempted to automate first — usually listing intake or first-response to an enquiry — and write it out as the four outputs above. Then measure the current baseline for one month before changing anything. If your team already runs paid campaigns alongside the platform, the reporting discipline in Align Google Ads KPIs with Business Goals to Maximize ROAS applies directly: define the metric, fix the time range, then decide. For the parts of campaign setup that can safely move to a reviewed automated flow, AI-Driven Google Ads Automation sets out the pattern.

FAQ

Does an agentic CMS mean we no longer need editors?

No. It moves the editor's work from typing to reviewing. On a destination platform the editor's judgement covers local accuracy, partner relationships and the decision to publish — none of which a CMS feature decides. What changes is the volume; an editor reviewing structured drafts works differently from one writing pages from scratch.

What is the difference between agentic AI and ordinary automation in a CMS?

Ordinary automation fires a fixed rule: when a record arrives, set a field. Agentic work chains steps and decides between them, calling tools as it goes. The second kind needs more definition, because a wrong branch is not caught by a rule. Ask any vendor which branches a step can take and who is told when it takes the wrong one.

How do we stop operators submitting bad listings?

By validating at the form rather than at review. Required fields tied to the event schema, a duplicate check against existing records, and an inline message naming what is missing cut the volume that reaches an editor. Track first-pass validation rate; if it does not rise after the guidance changes, the form is the problem, not the operators.

Should the platform answer enquiries automatically?

It should acknowledge them automatically and answer routine questions from a written source. Anything involving availability, price or terms needs a person, because a wrong commitment becomes the platform's liability and the operator's problem. Route by intent and stage, and keep a named owner for the ones that escalate.

How long before we can tell whether the automation worked?

One full cycle of the business, which for events is usually a season. Before that you are measuring the launch, not the workflow. Fix the indicators and their definitions on day one, and note the date the baseline was taken, so a later report can be read against something.

Sources

1. City of Perth unveils next-generation destination marketing platform via The Brand Agency — Campaign Brief — 2026-11-09. https://campaignbrief.com/the-brand-agency-unveils-next-generation-destination-marketing-platform-for-city-of-perth/