Frequently Asked Questions

Yes, with caveats to explain to anyone reading the results.

Prior Work treats n < 200 as a "small sample". For small samples the model blends your data with patterns from the research literature, and the smaller the sample, the more the results lean on that broader picture. With 20 respondents you are getting a research-based view lightly adjusted by your data.

It is still useful for spotting which factors deserve a closer look in a focus group, and for tracking the direction of change over time. It is not a reliable basis for claiming statistical significance or for comparing sub-groups within the 20.

Focus groups are one way to consult workers. You can use another suitable approach, such as individual conversations or a team discussion.

Survey results help identify concerns, but they do not establish why those concerns arose. Ask workers about the conditions behind the findings and whether proposed changes would help. A sector starting point also needs this local context.

The Focus Groups page has a facilitation guide and session notes template. Choose a format in which workers can raise concerns safely, and record what needs to change.

Response rate affects both how representative and how reliable the results are, but no single threshold makes them useless.

As a guide, 40% of a 100-person workforce (n = 40) gives more useful data than 40% of a 20-person workforce (n = 8). For the model, absolute numbers matter more than percentages.

Below 50%, non-response is often a signal in itself. Workers who distrust management, fear identification or have disengaged are under-represented, which tilts the results towards the more engaged part of the workforce.

Non-response can also change the apparent relationship between two factors, and in the worst case reverse it, when whether someone answered depends on the very things you are measuring. If the workers hit hardest by emotional demand are the least likely to reply, those who do reply are mostly the ones who were coping, so a demand that drives harm can look mild or even protective. Anyone who has already left because of these conditions is invisible to the survey.

No statistical adjustment fixes this, so be wary of any tool that claims one. What helps is practical: report your response rate alongside the findings, check whether the people who did not respond cluster in particular teams, shifts or roles (a finding in itself), and give the most weight in consultation to the groups you heard least from. If a group you have other reasons to worry about looks surprisingly well, silence is a more likely explanation than good health.

No. Prior Work is a psychosocial risk management tool: it identifies and addresses work-related hazards at the organisational level, and informs what the organisation should change. An EAP supports individual employees with personal or work-related difficulties, including while those changes happen and when circumstances are beyond the organisation's control. A complete psychosocial health strategy needs both.

For a mid-sized organisation (50-500 employees), a realistic first cycle is 4-6 months from the leadership brief to first action plan, and 18-24 months for the full cycle including pulse re-measurement.

The slow parts are survey fieldwork (2-4 weeks to allow responses), focus groups (scheduling and facilitation) and implementation, the longest phase, because interventions need time to embed before re-measuring means anything. The analysis in Prior Work is quick by comparison: a model build and report draft usually take an afternoon.

Prior Work uses a structured statistical model, a type of machine learning, that learns patterns in your survey data and lets you ask it specific questions. It is closer to a precise, inspectable calculator for workplace risk than to a chatbot or generative AI.

Why it suits this problem:

• •You can see inside it. The model is a visible map of factors and how they relate, so you can trace why a result came out as it did.
• •It's reproducible. The same data gives the same results every time, with no drift between runs.
• •It is explicit about uncertainty. It shows where the evidence is strong and where it is thin, which matters for the small and medium samples most organisations have.
• •It lets you ask "what if" questions. Act on Role Clarity and the model estimates which outcomes shift, and by how much.
• •It doesn't need anyone else's data to work. The model fits what you upload and answers questions about your workplace from its own data.

L (low), M (moderate) and H (high) are the three states used for every factor. For demands, high = more hazardous; for resources, high = more protective.

Cut-points are set by dividing the scale into equal thirds by default:

• Prior Work 5-point survey: L = mean ≤ 2.33, M = 2.34-3.67, H = > 3.67 (equal thirds of the 1-5 scale)
• Bring-your-own instrument: adjust the cut-points to equal thirds of that scale's range
• K10: items are summed to a 10-50 total, then collapsed from the ABS severity bands into three - L = 10-19, M = 20-29, H = 30-50

Equal thirds is a defensible default. Organisations with existing norms or clinical reference standards may prefer other thresholds. The Survey page has the Excel formulas.

With a small sample there are too few responses in each combination of answers to estimate relationships reliably from your data alone. If you follow the recommendations when uploading, Prior Work combines the signal from your workplace with what is already known from research in the field.

The smaller the sample, the more weight the model puts on the research-based starting point; the larger it is, the more the model leans on your own data. You don't need to set this. The balance follows the number of responses you upload, and Data Analytics shows which applies.

With a small sample the estimates are therefore more conservative and less extreme, which guards against overconfident claims about rarely seen combinations of factors.

One important exception. All of this depends on your survey lining up with the measures the research starting point is built from. If you bring your own survey, with your own questions and factor names, there is usually nothing to line it up against, so you are almost certainly relying on your own data alone. That is a reasonable way to work, but with nothing to stabilise the estimates, small-sample results stay close to whatever your respondents happened to say. Read them as a prompt for a conversation rather than a settled finding. The Data Analytics page tells you which situation you are in, and the written report says so in its method note.

Survey rows and models are sent to a calculation server. The application has no database for saving your work, but your browser can keep models, reports and some uploaded rows between visits. For the scope of this statement, hosting-retention limits and clearing instructions, see the Privacy page.

Check your organisation's requirements before uploading. Data sent for calculation can include each respondent's scores as well as workforce averages.

Review server processing, browser storage, document export and optional Anthropic or OpenAI report writing against your privacy policy and agreements. The Privacy page explains each operation and the limits of the available retention information.

There is no single right number. Every worker has an experience of their workplace, so a survey that reaches more people gives a more representative picture. More is usually better, up to the 2,000 responses the free tool analyses in one upload. Larger analyses are available as a paid service: contact patrick@priorwork.au.

Prior Work adjusts how it fits the model to the sample size: smaller samples draw more on patterns from the research literature, larger ones rely more on your own data. Data Analytics shows which applies once you've uploaded.

Demographic columns let you break results down by group - team, role, tenure, work arrangement, location, and so on. Include the groupings that matter for your analysis and action planning.

Prior Work limits its own demographic fields to those with a clear analytical purpose, which reduces re-identification risk. Apply the same discipline to any columns you add: if you wouldn't use it to target an intervention, don't collect it.

Age and other protected characteristics carry re-identification risk in small teams and should generally be excluded unless there's a specific reason to analyse by that characteristic and your sample is large enough to protect anonymity.

Translation is encouraged where needed for representative participation. Excluding workers by language is both a validity and a consultation problem.

The goal is construct equivalence - the translated item should evoke the same psychological construct as the original, which a literal word-for-word rendering often fails to do. For validated scales, back-translation (translate → independent reverse-translate → compare) is the standard check.

If rigorous translation isn't feasible, a plain-language simplified English version beats no translation. Contact patrick@priorwork.au if you're planning a translated deployment and want guidance.

Yes - this is the bring-your-own-survey option. If your existing survey covers any of the 20 Prior Work constructs (or close analogues), code those responses to L / M / H and upload directly. You don't need to cover all constructs; factors without data are handled by the model, not dropped.

The main caveat is benchmarking: the Prior Work model is calibrated against its own survey items and the published research. Using different item wordings or scales makes cross-model comparisons indicative rather than exact - within-organisation tracking over time is unaffected.

The model works with discrete states, so continuous scores are grouped. L / M / H splits each scale into equal thirds, which is transparent, consistent and makes no assumption about how scores are distributed. The cost is some lost detail, and a score right on a cut-point could fall either side, so treat borderline bands as indicative rather than exact.

It also maps directly onto the "low risk / moderate / high risk" language WHS documentation and regulators typically use.

Some practitioners prefer quartile or percentile cuts, which suit benchmarking within a known population. Prior Work uses thirds because they are consistent, portable and need no reference dataset.

Prior Work lets you choose a survey collection platform that fits your workplace and exports CSV. It does not collect responses directly from workers.

Check your chosen provider's privacy terms, access settings, retention and any secondary use of responses. Configure the collection process and tell workers who can see their answers; using an external platform does not remove these responsibilities.

When you upload results, Prior Work sends data to its server for calculations and can save survey rows in your browser.

Report generation uses your selected provider, Anthropic or OpenAI, with your own API key. Prior Work is free to use. Your selected provider bills your account for AI usage; Prior Work receives no payment from either provider for generating reports.

Both Main and Pulse generation include the report, an accuracy check, any necessary corrections, and a consultation guide. Providers charge for text read and generated across those stages, including hidden thinking. A token is a small piece of text. The cup-of-coffee comparison is a rough guide, not a fixed report price.

Each cost is for generating the example report and consultation guide together. Downloads are Word documents (.docx). These examples show the format and writing; they are not findings about your workplace. Your own report costs will vary.

For shorter follow-up reports, see the Pulse examples and estimated costs.

Monitor usage and review spending controls in your selected provider's account, such as the Anthropic console. OpenAI users can use their OpenAI billing settings. Budget alerts may not stop spending.

A pulse report explains the comparison with your recorded baseline, distinguishes measured results from estimates, and suggests what to investigate or review next. Its consultation guide helps you discuss the findings with workers. These examples show the format and writing; they are not findings about your workplace.

Estimates use the recorded usage and provider rates checked on 27 September 2026. Each includes the report, consultation guide and accuracy checks for that run. Your provider's bill is authoritative, and your costs will vary. These individual examples do not establish that a higher effort level will produce a better report. Downloads are Word documents (.docx).

See Main report examples and how report billing works for the longer initial assessment.

Report writing uses the provider's API, which has different data policies from its consumer chat app. Anthropic and OpenAI do not use API content to train their models by default; sharing it for training requires your permission.

Prior Work provides the selected service with calculated analysis, the context you choose to add and relevant report drafts. Individual survey response rows are not automatically included in these writing requests.

Data retention is separate from training use: retention periods depend on the model, features and your account terms. See Anthropic's API data handling and OpenAI's API data controls for current policies, or the Prior Work privacy page for browser storage and other data handling.

Because controls are chosen through consultation with workers.

The Australian WHS Codes of Practice require employers to consult workers when identifying and controlling psychosocial hazards, and ISO 45003 recommends the same. There is a practical reason too: the same elevated factor can have different causes in different teams. Role overload could mean understaffing, poor work allocation, unclear scope, or a mismatch between role and work. Each needs a different control, and only the people doing the work can tell you which applies. Prescribing controls without consultation would fail the duty and would likely waste effort.

Instead, Prior Work points you at where to look. On the Action Planning page, the interventions library is aligned to the hierarchy of controls (eliminate / substitute / isolate / engineer / administer) and, with a model loaded, puts the hazards most linked to your priority factors first, so you can match candidate controls to what consultation surfaces. Reports show what is elevated, consultation tells you why, and Action Planning is where you record what you will do, who owns it and when you will review it.

Because Prior Work is built to continue a legacy, and to lift the bar.

The People at Work platform was the free national reference for psychosocial risk assessment in Australia. It has now closed, and Prior Work exists to keep a credible free option available.

Prior Work looks at how hazards interact rather than scoring them one at a time, adjusts for the size and shape of your business, and is built on work stress models with a long research record in Australian workplaces. Findings come through an interactive web format, with optional AI-written reports you run on your own API key.

A rising tide lifts all ships: better free tools push the whole field forward, consultants and clients alike. It can't cook you breakfast, but hey.

The honest limitations to be aware of:

• •No support contract. Prior Work is maintained on a best-efforts basis. Bugs will be fixed when reported and capacity allows.
• •Non-commercial use only. The licence permits use by organisations for their own psychosocial risk management. Reselling Prior Work outputs as part of a paid consulting engagement requires a separate arrangement.
• •AI features require your own API key. Your selected provider charges for report generation. Check the usage notice on the Report page before generating.
• •No data hosting. Prior Work does not run your survey for you. You need a survey platform (MS Forms, Qualtrics, SurveyMonkey, Google Forms, etc.).

For almost everyone, you don't need to. Whatever the size of your upload, Prior Work learns how your factors connect directly from your data and picks a sensible default. This choice only appears if you open Advanced settings and want to steer how that happens.

Different build methods can land on slightly different connection patterns from the same survey, and that is expected: it is like asking two experienced analysts to read your data and seeing where they agree and where they differ. What each option means in practice:

Recommended. The right default for almost every upload: a well-tested search that balances how well the picture fits your data against keeping it simple enough to trust, set up to match the settings our review of the research landed on. It takes longer on a large file, but only once. If you're unsure, use this.

Score-equivalence search. Fast and well founded in theory. It decides more of the connection directions itself instead of reading them from your responses, so use it as a second opinion on the default.

Two-pass hybrid. A more cautious build that first sets aside the connections your data doesn't support, then refines what's left. A little slower. Choose it for a conservative starting point.

Constraint-based only. The fastest on large datasets, and it gives the same answer every time for the same data. It reasons differently from the score-based options and can give a noticeably different picture, which makes it useful as a second opinion.

Stability check. The recommended model, plus a figure showing how often each connection survives when it's rebuilt on different slices of your responses. Nothing is dropped; you learn which parts are firm. Steady means a connection doesn't hinge on who happened to reply. It does not mean the connection is correct. The slowest option.

Whichever you choose, the strength of each connection is estimated from the same data afterwards. Picking a sector model in Advanced settings adds a research-informed starting point for any connections that match that sector's framework.

Overview is a snapshot of where your workforce sits right now, before any change is applied. Nothing on it is a prediction.

Two things people misread:

• •The comparison figures are a reference point, not a target. Where a comparison is shown, it comes from the model you loaded rather than from a national benchmark. Sitting worse than the reference on a factor is a prompt to look, not a finding on its own.
• •The worst-scoring factor is not automatically the one to fix. A factor can look bad and still move very little of what you care about. The Summary tab answers that question; Overview deliberately does not.

The diagram view (the toggle beside the list) shows the same factors arranged by how they relate to each other. It is off by default because most people find the ranked list easier to act on, and the Combined Effects tab is a friendlier way into the same idea.

Summary ranks factors by how often they turn up across the strongest modelled scenarios, rather than by their score in isolation. That is why the ranking can put a mid-scoring factor above your worst one.

One limit worth knowing. A factor tends to appear in many top scenarios partly because it has the largest effect on its own, so the count and the individual effect are not independent signals. It tells you which factors keep earning their place. It cannot tell you whether pairing two of them adds anything, which is what the Combined Effects tab is for.

If the panel tells you the top three deliver less together than their separate figures suggest, that is not an error: the combined figure is worked out by moving all three at once, not by adding up three separate runs. Plan against the combined figure. What the comparison cannot tell you is why the total comes in lower.

The ranking is per outcome. Switching between, say, burnout and work engagement will reorder it, sometimes substantially.

The other tabs answer how much each factor matters on its own. Combined Effects answers what happens when you act on several at once.

It shows two things that are worked out separately. The tag and the figures compare acting on a set together against acting on each member alone. The picture beside them draws how each factor you picked reaches the outcome in the model. The picture is drawn from the model's structure and never reads the figures, so the two can point different ways about the same set without either of them being wrong.

Factors are rarely independent. Improving one often improves how workers perceive another, which means doing both can deliver less than adding their separate figures together. That is not a reason to do less. It is a reason to know what to expect before you commit to a plan, and it is why the Codes of Practice ask you to consider psychosocial hazards collectively rather than one at a time.

The tag on each option:

• •Green. The set is worth at least its members' separate figures added together. The word on the tag says which: “Adds up” for about equal, “Better together” for more.
• •Amber. The set delivers less than its members' separate figures added together. Plan against the combined figure and treat them as one piece of work. The tag does not say why the total comes in lower.
• •Grey. We can't tell either way, usually because the outcome already sits high in its healthy state and there is little room left for a second improvement to register.

Two limits worth holding onto. The survey captures one point in time, so the tab describes how factors move together rather than proving that one causes another. And the estimates assume an intervention actually shifts the factor it targets, which is a matter of design and delivery in your workplace rather than something the analysis can check.

Diagnostics works backwards from the selected outcome category, usually its adverse state. It compares the model baseline with the workplace conditions estimated for that outcome category, giving leads for further investigation. The differences are in percentage points; they do not mean workers have moved between categories.

This is the one tab whose ranking you should not read as a to-do list. It answers “what else is probably true?”, which is a different question from “what should I change?”. A factor can rank high here because it moves together with the outcome while still being a poor thing to act on, and the reverse happens too. For what to change, use Summary or Combined Effects.

It describes a pattern represented by the model, not an individual. If a claim or complaint prompted you to open this tab, treat what you see as a list of questions worth asking in consultation, not as an account of what happened to one person.

Scenario Builder is where you test a plan you are actually considering, rather than the one the model would pick.

Pinning a factor means “we intervene and bring this to its healthy state”. It does not mean “assume this is already true”. That distinction is why pinning a factor here and seeing it listed on Diagnostics give different numbers: one is an action, the other is an observation.

Two practical notes. The projection assumes the intervention succeeds in moving the factor you pinned, so it is a ceiling rather than a forecast, and how far you get depends on the quality of what you actually implement. And a scenario you save here is carried into your report as the scenario the report is written around, so it is worth saving the one you intend to argue for.

Both are marked experimental. They are exploratory rather than settled, and nothing in your report depends on either of them.

• •Budget Lab is a thought experiment about how a fixed budget might split across factors. It knows nothing about what anything costs in your organisation, so treat the split as a way of thinking rather than a recommendation.
• •Calibration compares prediction error with the outcome rate from the training responses, using the same held-out cases. It needs the original fitting information, independently specified survey preparation and uploaded responses. When the model's relationships were selected using those responses, the check covers probability fitting only. Scored counts and omitted answers are shown; confidence bands and universal quality grades are unavailable.

If you are working through the process for the first time, skip both. You lose nothing.

A bow-tie is a one-pager that shows how a hazard turns into harm, and where you have put something in the way. It reads left to right: the conditions that could set the harm off, the controls meant to stop them, the moment things get away from you in the middle, then the controls that limit the damage and the harm that follows if they don't.

Decades of research show that psychosocial factors interact and rarely act in isolation, which is part of why the format suits them: the conditions that set harm off are usually several at once, and so are the things standing in their way.

It earns its place when you need people who don't read statistics to see the logic of a risk and weigh in on it. A ranked list tells you emotional demand matters; a bow-tie shows a manager which specific thing they own, what it is supposed to prevent, and what happens if it slips. It is also the format most regulators and boards already recognise.

It is not a good fit for showing how much a factor matters, or how factors add up together. A bow-tie has no numbers in it. Use the analysis for the size of things and the bow-tie for the shape of them.

A worked example for one hazard. Note that burnout sits on the right as a consequence, not in the middle as the top event, and that one outcome runs on into another.

Work outwards from the middle. It takes a couple of hours for one hazard, and it is much better done with the people exposed to it than at a desk.

1. Name the hazard, then the top event. The top event is the moment you have lost control but nobody has been harmed yet. This is the step people most often get wrong: burnout is not a top event, it is what happens after one. "Sustained harmful exposure to emotional demand" is a top event. If you can still act to stop harm, you are at the top event; if the harm has landed, you are on the right-hand side.

2. List the threats. Each is a specific condition that could drive you to the top event, not a restatement of it. "Understaffing against ratio" is a threat; "high workload" is usually too vague to control. Three to six is normal. One line each.

3. Put a preventive control on each line. Every threat gets its own barrier, and a barrier only counts if it would actually stop that threat on its own. If you find yourself writing the same control on every line, it is too general to be a control.

4. Do the right-hand side, which is usually the thin one. Consequences are what the top event leads to, and recovery controls are what limits them: early check-ins, a workload reset, a genuine return-to-work path. Most first drafts are strong on prevention and almost empty on recovery, which is worth noticing, because recovery controls are what carry you when prevention fails.

5. Add escalation factors last. See the next question - this is the step that makes a psychosocial bow-tie worth more than a tidy list.

Then give every control a named owner and a way of checking it is still working. A control nobody owns is a hope.

A control is specific, owned, and checkable. "Training" is not a control. "Quarterly rotation off high-acuity work, rostered by the team leader, checked against the roster each quarter" is. If you cannot say who does it and how you would know it had stopped, it will not survive a busy month.

Psychosocial controls skew heavily towards rules, training and procedures, which sit near the bottom of the hierarchy of control and depend on people having the time and standing to apply them. Push for the ones built into how work is designed and rostered wherever you can, and be honest on the page about which kind you have.

Escalation factors are the part worth slowing down for. An escalation factor does not cause the hazard. It stops one of your controls from working. Low job control does not create understaffing, but it does stop a relief-staffing rule being used, because the person who needs it cannot make that call. Low supervisor support does not create emotional demand, but it does hollow out a debrief.

This matters more for psychosocial risk than almost anywhere else, because so many psychosocial controls are things a person has to choose to do. On the diagram they attach to the control they weaken, not to the threat. In practice they are often the highest-value thing on the page: fixing one escalation factor can restore several controls at once.

The tab does the first pass for you. It takes one outcome as the top event and arranges what the analysis already found around it: which factors lead into that outcome, what it leads on to, and which pairs of factors amplify each other most. It also suggests example controls from the WHS Model Code of Practice.

What it cannot do is know your organisation. The controls it offers are examples, not your controls; owners and verification are left blank on purpose, because they are the parts that make a bow-tie real and they have to come from you. Treat the tab's output as a starting draft to take into a room with the people doing the work.

From there you can copy it into whatever format you use, or generate the Prior Work written report, which includes a controls appendix per outcome.