Generative AI will be great for generative AI consultants

Receive free Artificial intelligence updates

We’ll send you a myFT Daily Digest email rounding up the latest Artificial intelligence news every morning.

Here’s an unusual thing, a sellside note on generative AI that does more than swallow and regurgitate the hype. It’s from JPMorgan analysts Tien-tsin Huang et al, who cover IT services at the bank.

GenAI represents “the biggest tech wave since cloud or mobile”, so will be a “multi-faceted revenue driver for IT services and BPO [business process outsourcing] providers,” they write. For the next few years, however, most spending will be to organise messy databases, disabuse management of delusions and direct capex towards tasks suitable for turbocharging with what’s in effect a whizzy form of autocomplete.

JPMorgan sees no shortage of applications for GenAI. But to demonstrate the limitations over the average wage slave, it highlights how quickly the probability-weighted algorithm powering ChatGPT gets distracted or bored whenever the job has fixed parameters:

GenAI is prone to errors. It is not artificial general intelligence, or AI indistinguishable from human intelligence. It cannot code perfectly, and it cannot set up a real enterprise’s IT architecture at all, just like it can’t value a company perfectly and it can’t set up a real equity portfolio at all. For established enterprises with material business momentum, critical information systems, and customer relationships to maintain, reliability of response is paramount. GenAI is a phenomenal productivity enhancer, but it’s not nearly good enough to replace most of its users.

That’s how it’s worked so far, with generative AI taking on the role of wingman for software developers, digital artists and content mill workers. At a corporate level, the projects currently being worked on are mostly to build walled-garden AI tools that don’t risk data leakage and will probably be restricted forever to the company intranet.

The first public-facing deployments (aside from toys like Midjourney and ChatGPT) are likely to be aimed at easing bottlenecks in customer service; stuff like call centres, where companies have fine-tuned large-language models with their own proprietary data, JPMorgan says.

By 2027 or thereabouts, the more conservative enterprises will be willing to take the lead from the early movers so AI “could expand dramatically and exponentially”, JPMorgan says. That’s when stuff like hyper-personalised media and robot PAs become a mainstream reality.

But for most companies, the preparatory work is going to be the difficult and expensive bit. The average corporate IT network is just too much of a clusterflux to use as feedstock for generative AI, particularly when no one can predict or even explain what it’ll be doing with the data.

Here in full is JPMorgan’s conclusion:

Despite the significant automation, engagement, and other opportunities, real widespread adoption of genAI remains years away. Early enthusiasts have passed the torch to the innovators in the typical adoption cycle, but the mainstream market has serious barriers to overcome to really capitalize on the opportunity that genAI presents.

Business leaders’ first questions for their technology advisors upon interacting with genAI tools are (1) how can we use it in our organization and (2) what do we need to do to get it rolled out. Each question is critical. Jack Dorsey, Co-Founder and Block Head at Block, encouraged companies to approach investments in genAI from a use case perspective rather than a technology perspective in his comments at our 2023 TMC Conference in Boston, suggesting that the hype around the technology could lead companies to spend aimlessly and thus realize inferior returns. IT Services companies can help operating companies target their genAI spending towards the areas with the highest returns. ACN [Accenture] cited on its F3Q23 earnings call that the company completed 100 genAI projects amounting to ~$100 million in sales in the prior four months; while this early traction demonstrates strong positioning for ACN and reflects the company’s robust client relationships, that $100M is still a tiny figure on ACN’s $60B+ revenue base, and the bulk of those engagements likely represent preliminary, exploratory projects as clients begin to figure out how they want to use genAI. These exploratory engagements can include conversations around what kind of model to use, how to train it, quantifying the lift required to prepare the training data, etc.

Legacy businesses must accelerate their digital transformations with emphasis on data readiness to capture the full potential benefits of genAI. Enterprises should benefit the most from genAI models when they crack them open and fine tune them with their proprietary data; to do so, companies must get their data in relatively good shape, which for most enterprises continues to be a challenge. Unified data assembled neatly in a modern database hosted in the cloud is optimal for fine-tuning LLMs, but most companies’ data is anything but unified on anything but modern database infrastructure. The “tech debt” that we discuss extensively in describing the drivers of digital transformation spending at legacy enterprises complicates genAI implementation just as it complicates other technology initiatives. Enterprises with faulty data housed in disparate, legacy databases will benefit from consolidating and modernizing their data estates in their efforts to implement genAI. The work required to ready enterprises’ data to fine tune models will probably dwarf the work of actually fine tuning the models. The longstanding supply-demand gap for engineering talent (which remains even after loosening of supply from cost rationalization at unprofitable and large-cap tech) ensures that enterprises will require tech services firms’ assistance with this data cleansing, migration, and unification work.

Mobile took around eight years to approach saturation, and cloud is still not approaching saturation after over a decade of heavy investments. From ACN to GDYN [Grid Dynamics], the largest and smallest IT services providers in our coverage (and beyond) say cloud work remains responsible for the bulk of their revenue growth. Tech debt is so substantial at most legacy institutions that any new technology faces a very steep climb. Therefore, despite our excitement about the potential for genAI to propel profitability across our coverage, we expect that it will take a few years before genAI contributes meaningful cost savings and revenue growth to our coverage group and several more years before these contributions ramp to their full potential.