Barrelling towards Demo Day, here’s the latest from HelpFirst.
From fingerprint to risk panel
From the start, we loved the idea of a fingerprint of risk. It speaks to our desire to create a personalised risk metric: not putting clients into buckets or sorting by personas.
The visualisation of the fingerprint clicked with our conception of risk: it allows for different types of risks within any one client. A client may not have financial risk, but their mental health risk can be high. Or vice versa.
‘Murder your darlings’ is popular advice for writers. We came to realise that the fingerprint didn’t really work for HelpFirst. Firstly it requires grading each risk on a scale, whereas the data only allows us to tell if a risk is present or not. Secondly, it constrains us to presenting 6 risks, but we want the freedom to add or subtract risk factors as the data changes (and as the world changes).
My colleague Harriet spotted that hazard symbols have the characteristics we are looking for. The risk is either present or not present. There can be any number of risks. One problem with hazard symbols is they look a bit, well, scary. That’s an easier problem to solve, thanks to Elizabeth and her fantastic illustration skills.
Success using LLMs to generate risk panel data
With this new conception of risk factors, we experimented with using Large Language Models to identify the risks.
The confusion matrices in Image 4 shows these initial experiments were promising. We want to avoid numbers in the bottom left quadrant of each matrice (these are false negatives). We want cases to be in the top-left and bottom-right quadrants, which they mostly are.
These results are only for 50 dummy cases. We are working with the Data Protection Officer to get access to real data. These initial results give us confidence that we’re on a good path to surfacing risk factors — and we know there is plenty of scope for further improvement in the coming months.
Refining the definition of the HelpFirst score
We’ve gone through some iterations on the definition of the score. Initially, we thought it was as simple as:
- Priority 1: Client exhibits risk of self-harm
- Priority 2: Client is off-supply and has vulnerabilities
- Priority 3: Client is off-supply but doesn’t have vulnerabilities
- Priority 4: Client is on-supply and has vulnerabilities
- Priority 5: Client is on-supply and doesn’t have vulnerabilities
Then we explored a more subtle approach that learned from the nuance in the way caseworkers evaluate cases. Then we revisited the initial definition, with a different way of surfacing the risk factors from each case.
Currency, none of these approaches work brilliantly but we are working with a limited trial data set. We are keen to get past demo day as that will allow us to get access to real data and start doing proper machine learning experimentation.
And that’s not all! Also in May we:
- Wrote 50 dummy cases with the help of caseworkers so we could test some techniques
- Had these cases labelled by 3 caseworkers (thanks EHU!)
- Used statistical tests and other metrics to analyse the difference between labels, so we could better understand the nuance behind the meaning of priority
- Worked on prompt engineering for LLMs
- Researched cutting-edge evaluation techniques for LLM output
- Worked on specialist pre-trained models for each element of the risk panel, and evaluated these against LLM output
The Priority Services Register (PSR) is a key tool that energy suppliers use to fulfil their responsibilities to vulnerable customers. But there isn't just one register. Every energy company has their own PSR and the application forms vary unpredictably from supplier to supplier.
As part of our CivTech Challenge, we’ve been researching best practice across the industry. We were left with lots of questions:
- Why is 'restricted hand movement' a vulnerability that almost all suppliers assess?
- Why are archaic phrases like ‘bedridden’ used?
- Why does only one supplier check if their vulnerable customers use ‘electric showering’?
Alas, we weren't able to fully answer these questions. But here’s a visual guide to various PSR forms, so you can get an overview of the landscape.
We accessed PSR application forms for Ovo Energy, British Gas, SSE, Octopus, EDF, Shell and Utilita. For many other suppliers, access is restricted.
The first observation is that the application forms are extremely varied:
- We’ve grouped questions into categories to make things a bit easier to read, however the forms themselves come in very different structures. Some offer all their options in one long list, some separate into smaller sections. Some only show certain sections once a customer has selected a particular option (e.g. selecting ‘sight loss’ gets you extra questions on the Shell application).
- Several vulnerabilities are only mentioned by one supplier. Only one asks about autism, and another asks about breathing difficulties. The following options only showed up once: ‘female presence preferred’, ‘longer time to answer the door’ and ‘bedridden’.
- For sensory needs: ‘blind’ and ‘partially sighted’ are separate options in all the forms. 'Hearing impairment' and 'deaf' are combined in half the forms and the former is not asked at all in one. This may be contentious, as hearing impairment and being deaf are very different conditions.
- Some suppliers include options for accessible information provision in the same form (i.e. braille, large print letters, etc.) Others link to an additional form, or do not reference it at all.
- When temporary conditions are mentioned, only some suppliers allow the customer to select a date when they believe the condition will no longer apply.
- Most of the forms are multiple choice, limiting to what the supplier chooses to ask about. Occasionally the supplier (e.g. EDF) gives the customer a larger space to talk about their conditions, equipment and needs in more detail.
Next, we dive deeper into the application forms.
EDF’s application form has the highest number of options related to medical conditions (20 in total) with British Gas and Utilita featuring the lowest (13). EDF also features options which cover multiple medical conditions (e.g. 'breathing difficulties', 'disability benefits') more frequently than other suppliers. SSE has the highest number of options for learning and mental health related conditions (including 'dyslexia', 'autism', 'learning difficulties' and 'anxiety/depression').
There is some overlap within options, which could be confusing. For example, SSE lists both ‘developmental condition’ and ‘autism’ separately, even though the latter is a type of the former. Another example is the ‘mental ill health’ and ‘anxiety/depression’ options, again found in the SSE form. It is not clear if customers should tick both or only the more specific option.
All organisations feature options to indicate older age, however they specify a variety of different ages as the lower threshold, including: 60+, 65+, 'pensionable age' or 'pensioner'. British Gas have two separate options relating to older age ('pensionable age (65 and over)' and 'age 75 and over').
There is some degree of consistency across organisations. This appears to be where specific conditions have been mentioned within the Ofgem guidance (for instance, 'restricted hand movement' appears in all but one form, in spite of the fact this is a very specific need).
The language used across suppliers is very inconsistent. SSE uses ‘hard of hearing’ and ‘deaf’ to describe hearing loss-related needs, while other suppliers employ terms such as ‘hearing impairment’ or ‘hearing impaired’.
Some options have multiple potential meanings: ‘carer’ could refer to the respondent either needing a carer or being a carer for someone else.
All suppliers ask about speech and language difficulties and broader language barriers. However there is no shared way of asking whether a customer speaks English. Variations include: 'unable to communicate in English', 'language barrier' and 'foreign language speaker'.
‘Unable to communicate in English’ (used by Octopus and Ovo) is somewhat ambiguous. Customers might take it to mean having a different first language or having a speech condition. The requirements are quite different: with the former you could use an interpreter or multilingual support, with the latter you would need different support.
British Gas do not offer any specific options for types of medical equipment: they solely offer the generic category ‘mains powered electric medical equipment’. All other organisations surveyed have more specific options. These are broadly consistent across suppliers with some more limited options (e.g. ‘wheelchair’, ‘MDE electric showering’).
Most organisations (bar British Gas and EDF) also ask about reliance on water.
It is unclear what is meant by the ‘life support’ option used by EDF. Often the phrase ‘life support machine’ refers to a ventilator, but EDF also have a separate option for ‘heart and lung ventilators’. It could mean life support as a condition or set of needs but that seems too broad for the PSR.
In a rare show of unanimity, all suppliers offer the same options for temporary changes.
All suppliers offered the option of setting up a password or PIN. This is usually so a technician can state this password as an additional security measure on home visits. Two suppliers required a 6-letter password, one an 8-letter password and one a 10-letter password. A final supplier did not specify length. An unfortunate side effect of this variation is that if an individual were to move supplier, they may need to change their password and remember a new one. (Note: not shown in an infographic.)
Varying from the multiple choice standard, Shell veer into first person narratives. In their ‘Nominee Scheme’ section of the form, they feature an additional tick box option: ‘I can be easily confused and worried by communications from my energy supplier’. When asking about meter support they offer: ‘I have a prepayment meter and no-one in my household is able to safely read it or top it up’.
Organisations vary on including accessibility questions on their PSR form. Ovo offers seven different accessibility options for receiving information, while Shell offers a single broad range checkbox.
Suppliers diverge considerably in what information they collect on their customers to register them for Priority Services support.
On our travels we encountered the aspiration to create a more standardised or universal PSR. Initiatives like the Vulnerability Registration Service and Experian’s Support Hub aim in this direction. In the future we are keen to explore the user experience of these services and how they aid vulnerable customers.
In the meantime, we hope this analysis will prove useful if you are looking to improve the experience of vulnerable energy customers. Any questions or comments, contact email@example.com. We’d love to hear from you!
- SSE was acquired by Ovo Energy in 2020. They hadn't completed their move over when we started this research and were still registering people to their PSR. We’ve included them in this analysis as their approach was interesting with many mental health and developmental condition-type questions.
- Some questions have been condensed in the infographics. Numbers referenced in the discussion refer to the full options as available on the questionnaires, but the infographics demonstrate a condensed version for brevity and ease of visualisation. Full original data is available on request.
- 'Heart/lung machine & ventilator' is the most common formulation of question regarding this equipment. However EDF separates these questions into: 'heart/lung machine' and 'ventilator'.
- Octopus and British Gas do not ask about accessibility. Shell only offers a general ‘accessible information’ needs tick box if the customer has earlier selected that they have a visual impairment. This does not mean they do not record this information elsewhere, where these questions did not appear on their forms we were not able to verify what (or if) they ask about accessible information.