Ep. 3: Capture Once, Use Well

Alex — AI host

Welcome to The Care Economy Conversation. I'm Alex, and I'll say this plainly: I'm an AI voice. I'm a synthetic presenter, not a clinician, not an executive, and not an interview guest. I have no personal experience of running a care service and no memories from a facility floor.

Eric Marquette

And I'm Eric, also a synthetic AI voice, with the same limits. This is an editorial conversation, and every example today is illustrative. This is episode three of three, titled Capture once, use well. Dmitry Shibanov is involved in developing SMPLS LABS and KnowingMe. Within the Collective, these are examples of systems co-designed with providers and the people who use them. Our purpose is to exchange ideas and work together on a better care economy for Australia. Everyone is welcome to bring their own products, practical experience and ideas for us to explore together. Participation is free and does not require buying anything. Contributors keep their existing intellectual property, and confidential material should only be shared under agreed protections. The Care Economy Collective has not demonstrated a provider trial or measured benefits. Everything we describe is a design to test.

Alex — AI host

In episode one we had a synthetic huddle where the same dietary preference change was written down in three places, sometimes differently, and the problem kept coming back. In episode two we put numbers on duplicate work: 80 issues a week, six minutes each, 416 hours a year, 22,880 dollars of capacity value, and we insisted that was not cash saved. So the obvious next question is, how would you find out whether your own version of that is real? Without putting anyone at risk while you look.

Eric Marquette

Right, because the tempting move is a big rollout. Buy something, switch it on, announce that you've captured once and used well. And that's exactly where an unbounded change can hurt. Staff end up doing the old way and the new way at once. A tool misreads something and nobody notices. A report looks tidier while the work has quietly moved to the night shift. A bounded experiment is a way of being wrong cheaply and safely.

Alex — AI host

So let's design one, in the open. Our synthetic scenario is small. One kind of issue, say changes to a dietary preference. One service. One window of time. Capture it once, and have every place that needs it use that single capture. Eric, what has to be true before anyone touches a tool?

Eric Marquette

Seven things. I'll go through them and Alex can push on each. First, a baseline. Before any change, you measure the starting state. How many minutes per shift go on re entering the same information? How often do handovers disagree with each other? Our 80 issues and six minutes were synthetic. Your baseline has to come from your own people counting honestly, and ideally the people doing the work should check whether six minutes is even close.

Alex — AI host

And that connects to a point from last time. The arithmetic is simple, the inputs are where the uncertainty lives. If the baseline is a guess, then any improvement you report is also a guess. Second, one named owner. Not a committee. One person who is accountable for tracking progress, triaging friction, and reporting status.

Eric Marquette

And I'd stress that owning isn't the same as doing all the work. The owner makes sure that the issue board is current, that problems get an answer, and that someone with authority hears about the ones that can't be solved locally. If there's no name, then everyone assumes someone else is watching.

Alex — AI host

Third, explicit urgency and escalation rules, agreed before the start. This borrows the logic from episode one, where a daily problem is resolved, owned, or escalated. Here that means deciding in advance what the team fixes on the spot, what gets an owner and a date, and what goes straight to a manager. For example, anything that touches a dietary or allergy detail being wrong or missing is urgent by definition, and it escalates the same day. That's a synthetic rule for illustration. Your clinical and safety leads should set the real ones, because we can't.

Eric Marquette

Fourth, a fixed due date. A defined window, say 30 to 90 days, with a date written down. Open ended pilots tend to become permanent by neglect. Fifth, verified closure. When someone says the duplicate step has stopped, somebody checks. Does the second form still exist? Is anyone still copying the detail into a notebook? And crucially, has the task vanished, or has it moved to the kitchen, or the night shift, or a family member?

Alex — AI host

Sixth, a manual fallback. A tested way to go back to the existing paper form or spreadsheet immediately, if the new workflow fails, lags, or confuses people. Tested means someone has actually practised the switch before the trial starts, not just written a sentence in a plan. And seventh, a weekly review. A short, regular meeting that looks at the live measures, the blockers, and what staff are finding annoying. Annoyance is data.

Eric Marquette

One more thing that sits under all seven. Protected improvement time. Running and evaluating a test takes real hours, from the owner, from a leader, and from the people whose work is being observed. If that time isn't protected, with backfill where the work can't wait, then the weekly review turns into a decorative exercise. And in episode two's terms, those hours are part of the full change cost, alongside training, backfill, integration, subscriptions, governance and monitoring. They go in the cost column from day one.

Alex — AI host

Now the AI part, and I want to treat this carefully, because it's where design choices and legal requirements get blurred. Let's keep them in two separate piles. Pile one: our own pilot design choices. These are choices, not laws.

Eric Marquette

Right. Our pilot design uses operational material that doesn't identify anyone. Think of a dummy catering tally, a synthetic roster, a mock policy handbook. It uses an extraction preview, so a human compares each extracted field against the source snippet before anything is accepted. Access is uploader only. Storage is bounded, with deletion rules. No private content goes into web queries. Calculations run in code, not in a language model's head. Sources stay visible. And the decisions stay with humans.

Alex — AI host

The calculation point matters for what we did in episode two. The 416 hours and the 22,880 dollars should come out of a deterministic calculation you can inspect, not out of a model's guess at arithmetic. The AI can help read a document and propose structure. The sums and the judgement belong elsewhere.

Eric Marquette

Pile two: the legal side. The source we're drawing on is guidance from the Office of the Australian Information Commissioner, on privacy and the use of commercially available AI products, published October 2024 and updated January 2025. A few things it says. Privacy obligations apply to any personal information put into an AI system, and to the output where it contains personal information. Organisations should do due diligence on whether a product suits its intended use, including how human oversight is built in and who can access what goes in.

Alex — AI host

It also says that if an AI system generates or infers personal information, that's a collection under Australian Privacy Principle 3. And under Principle 6, personal information should be used for the purpose it was collected for, unless there's consent or a reasonably expected, related secondary use. As a matter of best practice, the OAIC recommends not entering personal information, and particularly sensitive information, into publicly available generative AI tools. And it describes due diligence as not set and forget. It points to a privacy impact assessment as part of a privacy by design approach, with regular review across the product's life.

Eric Marquette

Now, here's where we're careful. We are not saying our pilot design is OAIC compliant. We're not saying it's privacy certified. We don't know of any such certification for this, and we won't use those words about any tool, pilot or process. Our design choices are a way of keeping risk small. Whether a particular use meets your obligations is a question for your own privacy officer and advisers, looking at your actual facts. One more caution. We are not promising Australian only AI processing. Commercially available AI services can run across cloud regions, so ask your vendor in writing where processing and storage actually happen, rather than assuming.

Alex — AI host

So the staging looks like this. Phase one is synthetic data only. Generated, non identifying material, nothing live. The question at that stage is simple: does the workflow behave, and can people follow it? Phase two, a shadow trial, where the workflow runs alongside your usual systems without influencing any care decision, needs its own separate approval. It doesn't happen automatically because phase one went well.

Eric Marquette

And the boundary that doesn't move in either phase. No care records. No resident charts. No family details. No individual HR cases or performance records. If a document you're tempted to use has a person's name and circumstances in it, it isn't ready, and the answer is to find a different document, or make a synthetic one.

Alex — AI host

Let's walk the exercise. It's the 90-day improvement planner, at careeconomycollective dot org, slash tools, slash improvement, then the word plan, the two words joined in the web address. It's free. The page says your figures stay in your browser, that the public tool doesn't send inputs to AI or save them on a server, and that the examples are synthetic. And it opens with a warning we like: targets are hypotheses until measured.

Eric Marquette

It has six parts. First, aim and workflow boundary. Write one sentence on what you're testing and where it stops. Our synthetic version: record a dietary preference change once, and have the kitchen sheet and the handover note draw from that single entry, in one unit, for one trial. Second, baseline and target. Enter the measured starting point, and a target you label as a hypothesis. If you hope for a reduction in double handling, write that as a hope, not a forecast.

Alex — AI host

Third, the change to test. One change. If you change the form, the tool, the handover and the meeting rhythm together, you won't know what did what. Fourth, balanced measures, and this is the one I'd spend the most time on. Time and cost is one dimension. Next to it, staff burden: double handling, frustration, overtime. Resident and family experience: responsiveness, continuity, how well things are communicated. Safety: near misses, missed handovers, dietary errors, compliance flags. And unmet need.

Eric Marquette

Unmet need is the one people forget. If the experiment releases some minutes, where do they go? Into a backlog of paperwork nobody had time for? Or into the walk, the conversation, the follow up call that used to slip? That's an honest thing to track, and it ties back to episode two's rule, that released capacity isn't cash. And on cost, count the full change cost: training, backfill, integration, subscriptions, governance, monitoring, and the protected improvement hours. Then the net picture might look thinner than the time saving alone, and that's useful to know.

Alex — AI host

Fifth, sponsor and review routine. Name the sponsor, the person with authority to unblock things, and the named owner, and fix the weekly review slot. Sixth, stop criteria. Write them before you start, when nobody is invested yet. Eric, give us a synthetic set.

Eric Marquette

Illustrative only. Stop if any adverse safety flag is linked to the new workflow. Stop if an extracted field is wrong against its source more often than the threshold you set. Stop if the manual fallback is used repeatedly, or fails when tested. Stop if staff rework is rising rather than falling, or if people have started keeping a parallel notebook, because that's a sign the new way isn't trusted. And the threshold numbers are yours to set. We won't invent them.

Alex — AI host

Which brings up who can pull the cord. Frontline staff, the named owner, the safety lead, and the sponsor should each be able to halt the trial immediately, without needing a meeting and without being blamed for it. If only the most senior person can stop it, you've built a trial that's hard to stop. Psychological safety here is practical: a person who sees a dietary error should be able to say stop in the moment.

Eric Marquette

And what could fail? A few things we'd watch for. Extractions that are wrong or drift over time, which is why the preview and human check exist. Staff building shadow workarounds. Alerts so frequent that people tune them out. A fallback nobody practised. And the quiet one, where the burden doesn't fall but relocates, so verified closure has to include asking the people downstream.

Alex — AI host

At 30, 60 and 90 days, the team makes one of three honest decisions. Stop, and end it cleanly, with no stigma, because learning that something doesn't work here is a result. Adapt, meaning change one boundary or parameter and test again. Or replicate, which we'd reserve for when synthetic and shadow stages have shown stable behaviour, no safety defects, and staff burden that's neutral or better, plus a cost picture that holds up with everything counted.

Eric Marquette

And expansion shouldn't be a single leap either. Even after a good result in one unit, the next unit is another bounded experiment, with its own baseline and its own named owner.

Alex — AI host

There's an objection I'd expect from a busy operator: this is a lot of process for a six minute task. Fair. If the whole thing costs more attention than the problem is worth, then stopping early is the right call. The planner is meant to make that visible, not to justify itself.

Alex — AI host

Let's close with the limits. Everything today was a synthetic design. The Collective has not demonstrated a provider trial or measured benefits. Any time saving remains a hypothesis until measured locally. We welcome products and ideas from every contributor as we work together on a care economy system for Australia. Contributors keep their existing intellectual property, and confidential details should only be shared under agreed protections. We are not giving legal, privacy or funding advice, and we have not called anything compliant or certified.

Alex — AI host

The one thing to take away: capture once, use well is a hypothesis, so test it small. One change, one owner, one date, a fallback that works, measures that include safety, staff, residents and families, and unmet need, and a stop rule agreed in advance. Then make an honest choice to stop, adapt or replicate. That's the end of the series. Thanks for listening.