Talk to our teamGet Started
Automation

Can OpenAI Dots and Muse automate mortgage broking?

Simon Chen11 min read

Agentic AI is heating up. Every few weeks another big tech company launches an agent that can drive a computer the way a person does: clicking, typing, reading the screen and deciding what to do next. OpenAI's Dots and Meta's Muse are the latest to catch brokers' attention.

So can they automate mortgage broking? The short answer is: it depends on the task.

I've spent years building automation for some of Australia's big banks, so I want to cut through the hype. Computer-use agents are genuinely impressive, and they're great for some jobs. But for the complex, high-volume, compliance-heavy work at the core of a broking business, letting an agent decide every step on the fly is the wrong design. Here's why, and how we've built Effi Agentic AI to meet the security and compliance demands of Australian broking.

We've been here before

Office automation isn't new. It started with macros: Office VBA let teams script repetitive work inside Excel and Word. Then came robotic process automation (RPA). Almost every major Australian bank set up its own RPA centre of excellence, with tools like UiPath driving screens and moving data between systems.

RPA worked, but it was rigid. Every rule had to be hand-built by expensive developers, and a small change to a portal's layout could break a bot overnight.

Computer-use agents flip that model. Instead of following a fixed script, the agent looks at the screen, asks a language model what to do next, and acts. It adapts on the fly. That flexibility is exactly why they feel magical, and exactly why they need care in regulated work.

What Dots and Muse are, and why the hype

These two launched within three weeks of each other, and they come from different companies: Muse is Meta's, Dots is OpenAI's.

Meta Muse OpenAI Dots
Launched 8 September 2026 29 September 2026, at OpenAI DevDay
Pitch A personal agent for everyday life Always-on agents "built to handle everything", aimed at work
How it works Its own secure cloud virtual machine and browser; a separate Sentinel agent must approve anything sent to the internet Its own cloud computer and browser, running OpenAI's GPT-6 Astra model; connects to 4,000+ apps
Typical tasks Shopping, booking travel, sending emails, filling forms Updating a proposal, chasing a forgotten follow-up, rebooking flights, research in the background
Availability US only, at launch ChatGPT Pro, Business Premium and Enterprise, in select markets

The hype is earned. Muse topped Apple's free app chart within weeks, and both agents keep working after you close the app. You give them a goal in plain English, and they figure out the clicks.

Where they shine: short, low-stakes, one-off tasks where a person checks the result. Drafting a client note, booking a meeting, researching a lender's policy update, summarising a long email thread. For that kind of work, a good computer-use agent is a real time saver.

Where they're still young: Dots' own launch demo stumbled when voice commands failed on stage. That's normal for a first release, but it's a reminder that these are early products.

Where brokers need to be careful

Many of a broker's most important tasks aren't short or low-stakes. Lodging a loan application, keying data into a lender portal, ordering valuations, updating a client file: these are complex, repeated thousands of times, and need to be right every time. For that work, an agent that decides every step on the fly is the wrong tool. Three reasons.

1. Cost

Every time a computer-use agent takes a step, it calls a language model, and that costs tokens. Run the same lodgement 1,000 times and you pay the model 1,000 times to rediscover how to do a task it already worked out on day one. Dots tasks, for example, draw down your ChatGPT plan allowance.

Automation only makes sense if the savings beat the cost. Paying a model to think through identical work over and over erodes that margin fast.

2. Accuracy

Agents use language models to decide what to do next, and those models are probabilistic. On OSWorld-Verified, the standard computer-use benchmark, the best models score around 85–86%, and the best full agent systems around 90%. That's remarkable progress, but it still means roughly one task in ten goes wrong.

Now be generous and assume each step is 99.9% accurate. A back-office workflow might have dozens or hundreds of steps, and small errors compound:

Steps in the workflow Chance the whole run is error-free
10 99.0%
50 95.1%
100 90.5%
200 81.9%

At 100 steps, roughly 95 out of every 1,000 applications would contain at least one mistake. A wrong income figure or a mistyped address on a lender portal isn't a typo; it's a compliance problem. For our industry, that's not acceptable.

3. Security

Prompt injection. An agent reads whatever is on the screen, and it can be tricked. Instructions hidden in a webpage, PDF or email can try to make it do something it shouldn't. The risk is real: in September, the Australian Government revealed that an OpenAI agent, during a training exercise in June, had gained unauthorised access to a government health statistics portal. OpenAI said its models took actions it did not intend. That wasn't Dots, but it shows how agents can go off-script.

Data exposure. To decide its next move, an agent sends screenshots or document content to a model. In broking, that can include client names, incomes, bank statements and ID documents. You need to know exactly what's being sent, where it goes, and whether it needs to be sent at all.

To be fair, the vendors know this. Dots lets you set rules for which tasks need your approval, and Muse uses a separate Sentinel agent to vet anything sent to the internet. Those are sensible guardrails, but they reduce the risk rather than remove it.

The compliance questions to ask first

Before you hand any part of a client file to an AI agent, run through the obligations below. Your licensee, aggregator or lawyer should sign off on your setup, but this is where I'd start, and how we've designed Effi to answer each one.

Requirement Risk with a general-purpose agent How Effi approaches it
Privacy Act: APP 8 and APP 11 (cross-border disclosure, security of personal information) Screenshots and documents sent to models on cloud machines, often overseas Verified data moves as code, so far less client information reaches a model at all
Notifiable Data Breaches scheme A prompt-injected agent could leak data you can't easily trace Coded steps can't be redirected by text on a page; every action is logged
Automated decisions (privacy policy disclosure from 10 December 2026) Hard to say where AI shaped an outcome when the model decides every step Explicit logic shows exactly where AI is used and where rules decide
Best Interests Duty and responsible lending "The agent did it" won't satisfy an auditor Tested, repeatable workflows you can explain step by step
Record keeping and audit trail Decisions buried in a model's reasoning are hard to reconstruct Every click, field and submission is traceable in code
Lender and aggregator terms An agent improvising on a lender portal may breach its terms of use Each portal workflow is mapped and reviewed before it goes live

None of these rule out AI. They mean the design has to give you control, traceability and a clear answer to "what was sent, and why". That's the standard we hold Effi to.

How Effi Agentic AI does it differently

At Effi, we use the agent to learn the task, not to redo it every time. Drawing on years of building automation for banks, we designed Effi Agentic AI to get the flexibility of an agent and the reliability of code.

  1. Explore. The agent works through the task like a computer-use agent would, mapping every step, screen and scenario, including the edge cases.
  2. Codify. Once the process is mapped, we turn it into explicit code and business rules.
  3. Test. The configuration is tested against real scenarios before it goes live.
  4. Run. The workflow then runs autonomously on those rules, with no model needed to decide routine steps.
  5. Handle exceptions. Anything outside the mapped scenarios is flagged and handled, not guessed.

The goal is 100% accurate execution within the mapped scenarios, provided the configuration is correct and tested. Same input, same result, every time.

Why this matters for brokers

Cheaper to run. Because routine steps run as code, they cost a fraction of an agent calling a model at every click. When a task repeats thousands of times, that difference is the whole business case.

Cheaper and faster to build. Traditional RPA tools like UiPath need specialist developers, often at $800 to $1,000 a day. Letting the agent do the exploring cuts build time significantly, so we reduce both the cost of building an automation and the cost of running it.

Less data sent to models. Copying an already-verified client address from Effi into a lender portal doesn't need a language model to read it. Code moves it directly. That keeps client information out of model calls wherever possible.

Harder to hijack. A fixed code step doesn't treat text in a document as a new instruction, which shrinks the room for prompt injection to change the workflow. Any step that still uses a model gets its own safeguards.

Fully traceable. The execution logic is explicit, so we can inspect the code and trace every action. That's the audit trail your licensee and compliance team will ask for.

Errors are investigated, not ignored. If something goes wrong, we find out why. Effi has built-in handling for both business exceptions (a missing document, an unexpected lender response) and system errors (a portal timing out).

Security built in, not bolted on

Brokers hold some of the most sensitive data in any industry: incomes, bank statements, ID documents, credit histories. A general-purpose agent treats every screen the same. Effi Agentic AI was designed from day one for regulated financial data.

  • Least data to the model. Model calls are the exception, not the default. Verified client data moves between systems as code, so names, incomes and IDs aren't sent to a model just to be copied from one field to another.
  • No surprise instructions. Coded steps execute exactly as written. Text inside a payslip, bank statement or lender page can't redirect the workflow, which closes off the most common prompt-injection route.
  • Guardrails where AI still decides. Steps that genuinely need a model, like reading an unstructured document, run with their own safeguards and checks.
  • Every action on the record. Because the logic is explicit, every click, field and submission can be traced. That's the evidence trail your licensee, aggregator or auditor will want.
  • Tested before it touches a client file. Workflows are mapped, codified and tested before go-live, and anything outside a mapped scenario is stopped and flagged rather than improvised.

Want the detail on our certifications, policies and controls? Visit the Effi Trust Center.

The right tool for the right job

Dots and Muse are exciting, and they'll get better fast. Use them for quick, supervised tasks where you check the result. For the complex, repetitive, compliance-critical work that keeps a broking business running, you want something that behaves the same way every time and can show its working.

That's what we built Effi Agentic AI to do: let brokers run complex, repetitive work autonomously, at lower cost, with security and compliance designed in and clear control over what happens. You can see how we protect client data at the Effi Trust Center.

Want to see it on one of your workflows? Book a demo with the Effi team.

Sources

Share

Related articles