Artificial intelligence pilots at HSBC, Morgan Stanley and Societe Generale have uncovered a range of challenges that affect how the projects might transition into production.
Speaking at SIBOS 2017 in Toronto last month, HSBC’s global head of operations Kirsty Roth warned that early successes with derivatives of enterprise AI technologies had not necessarily uncovered potential roadblocks to production deployment.
“It’s been somewhat easy - in a funny way - to get going using sample data, [but] then you hit the real problems,” Roth said.
“I think our early track record on PoCs or pilots hides a little bit the underlying issues.
“Then we get into what data are we comfortable with using on a public cloud, or can we even agree as an organisation on what we think the data rules are by country - believe me that’s a pretty good debate between regulatory, legal and the techs.
“It’s about what does live [in production] really mean and how is it going to be sustainable?”
HSBC had taken a dual-prong approach to enterprise AI in the experimentation phase.
On one hand, the institution had looked at its problems and how AI might be able to assist in solving them - what Roth called a “pretty old fashioned use case approach” to the technology.
But HSBC also did some work using the technology as a starting point.
“There are more vendors who want to sell you something in this space than you can count so sometimes [it’s a case of a] solution looking for a problem,” Roth said.
However, HSBC was happy to entertain some technology experimentation to learn capabilities before there was an obvious use case for them.
“[There were] things that we saw that had worked in other places and we could see quite quickly how we could apply them,” she said.
Societe Generale Securities Services’ head of business solutions, Matt Davey, said his organisation encountered similar types of challenges with its early uses of robotic process automation (RPA), a specific type of AI.
“We’ve done quite a bit of work with RPA recently and I have to say we’ve been a bit disillusioned with that experience,” Davey said.
“When I talked to people internally, there was quite a lot of negative comment about RPA and the fact it was really like a macro technology.
“I started in systems in the late ‘80s and macros have their value, but I think that some of the challenges are if you need to ‘parameterise’ all … aspects of a process for an RPA then there’s a cost in doing that. If you need to add people to monitor what the RPA is doing, there’s another cost there.
“So when you look at the total cost of ownership there’s quite a question mark over whether it’s worthwhile.”
Davey said Societe Generale also saw potential flow-on issues from widespread use of RPA as-is.
“When you put that macro or RPA in, it has the effect of setting your legacy systems in aspic,” he said.
“It’s very hard to make changes to those systems from that point because you disrupt the RPA that you’ve put in on the top, so it’s something of a one-step move: it stops you from really making changes to your legacy platforms from there.
“I think one of the themes of AI is how do the banks in particular integrate these technologies onto the spaghetti of middle and back office platforms that most of the banks have.
“That’s a real challenge that I think is highlighted by some of the experiences that we’ve had with RPA.”
However, he noted that those challenges had not necessarily soured appetite for RPA or AI in general. Rather, they had informed the path on moving some of these technologies out of experimentation and into production use.
“The follow-on to RPA is we’re now looking at RPA with a combined AI component so that you have the AI component managing the state of whatever the process is,” Davey said.
“We haven’t stopped it, we’re just changing the tool.”
Davey also said he did not “want to sound negative on AI”.
Like others, his challenges simply stemmed from the fact that “the PoC is the easy bit: it’s how you get that into production and shift the balance” that causes enterprises some real problems.
“We’ve got a lot of proof-of-concepts,” he said.
“We’ve got some great examples of AI being used on legal documents where we can look and parse the clauses on a legal document.
“A lot of clauses will be standard but you’re looking for the 20 percent or so that are non-standard.”
Morgan Stanley’s global head of client technologies Louella San Juan noted the firm had experienced “success with unstructured data and running that via machine learning models”, particularly those from the open source - rather than vendor - worlds.
Get tips on how Australian enterprises are tackling artificial intelligence - download iTnews' report to read individual case studies as well as analysis of the broader state of play.