How DWP is getting AI to work

Balancing speed and safety

Managing tech means accepting that your decisions affect how people work, and even how they live. It requires understanding risk.

When your systems deliver what might be the only financial lifeline for 20 million people, your risk appetite is understandably pretty low. So how do you balance that need for safety against the drive for innovation and the need for speed?

By doing everything "very carefully," says Richard Corbridge, chief digital information officer & director general for DWP Digital, who has big plans for IT at the UK's largest government agency.

"No important decision [at DWP] is made about you by any computer, it is a human that's making the decisions in that space," says Richard, who joined DWP Digital in April last year.

He adds that nothing the department does can be in a "black box" - a long-standing criticism of AI tools.

"It has to be explainable by us, DWP, not the partner that we might be working with."

The partners only provide the tech: the training data all belongs to DWP.

"Maybe it should be called small language models," Richard jokes. "We've been really careful to use generative AI, but do it against ringfenced data and capability... We understand the data we're putting in."

Now read how Oracle is doing the same thing for its customers

It all comes back to transparency and explainability, he says. "If you can't explain to the exec team or my leadership team how it works, then you know what? We shouldn't be doing it at this point, because it's still too unknown as to what's next and what's out there."

Spotting vulnerabilities

Unlike private organisations just dipping their toes into the AI water, DWP is using the tech for more than just basic automation and marketing copy. It's diving right in, because it can't afford to wait.

Even in these days of digital, DWP gets tens of thousands of letters every day. They come in many forms, but according to Richard contain just one message:

"Handwritten, typed, pages and pages, Post-it Notes: all sorts of different ways that people get in touch with us through the post to say, ‘I need help. I'm vulnerable'."

Sorting through more than 22,000 letters every day is a mammoth task, and before AI it was all done manually. Humans would open and read every letter before sorting them by severity, making hard calls about who to help first, a process that could take weeks from receiving the letter.

"We're now doing this on the day that those letters land," says Richard. "We are using a large language model with a view to spotting vulnerabilities, and therefore prioritising all 22,000 letters so the most vulnerable people get looked at and spoken to by a human at pace."

The system, known as White Mail (Richard admits this is "terribly named," but is the historic term DWP uses for unclassified correspondence), has been live for six months, and the results are encouraging: 22,000 documents, both handwritten and typed, are analysed each day. The solution has processed over 2 million documents to date.

Other ways the department is using AI include finding the next best action for work coaches based on policy and citizens' situations; and a tool based on a large language model that can recognise and match clinical terminology.

"You wouldn't begin to believe how many different spellings of ‘pneumonia' there are in a DWP system; so, [we are] using new technology to standardise how pneumonia is recognised in multiple ways in our systems and then translated into a single version of ‘what is pneumonia?', so that when a citizen is talking to anybody in DWP they can see that that person has had that illness and therefore this help needs to be provided for them."

That is at the heart of DWP's embrace of AI: the need to deliver potentially vital help to the most vulnerable people in society

The department was recruiting for a director of AI and innovation when I spoke with Richard, to act as an exemplar and evangelist for new ways of working across government. An interim director, Simon King, has since been found.

But despite the hype, the excitement and the new possibilities, Richard won't overreach.

"Calm and considered are our watchwords - how do we do this and always stay in control?"

That question is on the mind of nearly every IT leader this year, with more than two-thirds calling the tech "promising but immature" in our latest research. Click here to read more.