This Shift considers whether GenAI is delivering on its promised productivity gains and what may be preventing that.
A recent MIT report garnered a lot of attention after it found evidence that questioned the real ROI of early GenAI deployment. The research found that although many organizations have rapidly adopted generative AI, 95% of respondents claimed that they had seen no measurable ROI from these deployments.
There were some critiques of the research methodology which questioned whether it had been too reliant on interviews with senior leaders as well as whether the definition of ROI (scaled deployment and measurable impact on the bottom line within six months) was too narrow; nevertheless, this is another indication perhaps that ROI is taking longer to arrive than the hype around generative AI would suggest.
One of the things that may be holding back those gains is the quality of the work that organisational GenAI deployments are creating.
Thanks to research from Stanford Social Media Lab and BetterUp Labs, there is now a term for low quality GenAI work related content: ‘workslop’. This refers to any AI-generated output that “masquerades as good work but lacks the substance to meaningfully advance a given task” – creating more work rather than less as time is lost fixing or redoing the task.
In the survey 40% of respondents said they had received ‘workslop’ in the prior month, and that each instance can cost up to two hours (or more) in cleanup, review or revision. The research also showed that this phenomenon was undermining organisational relationships, often causing the recipient to think of their colleagues as less reliable or intelligent.
The ability to produce valuable work from GenAI tools is entirely reliant on the skills and thought of the person utilising the tools. However, Econsultancy’s own research shows that training in this area is often lacking or not fit for purpose.
At Econsultancy, bridging this skills gap for marketers has been a keen focus. Even before the emergence of the term ‘workslop’, Econsultancy’s Fast Track to GenAI for Marketing learning programme centred largely on giving teams the knowledge and skills to be able to create real value – rather than content that just has the appearance of it.
There has been a slew of announcements around agentic AI of late, including OpenAI’s “ChatGPT agent” and its AgentKit: an agent builder designed to let developers (and non-developers) visually craft, deploy, and optimise agents.
Google also introduced Gemini 2.5 with browser agentic capability, meaning it can navigate web pages like a human (opening, clicking, dragging), enabling agents that act within browser UIs rather than just via back-end APIs.
Perhaps most interesting though is the announcements around agentic commerce. The idea is these agents would be able to shop on the user’s behalf. Making the biggest splash in the ecommerce space was OpenAI with the launch of “Instant Checkout” and Agentic Commerce Protocol, which effectively enables purchases directly within conversation without leaving ChatGPT.
Progress on these agents is causing ecommerce businesses to take action. Shopify, for instance, has brought in new rules around how AI buying agents can behave on its customers’ websites.
The policy explicitly bans automated scraping, “buy-for-me” agents, and any end-to-end transaction flow that completes checkout without a final human review step.
Alongside this, Shopify have a partnership with OpenAI but there is an interesting tension here between OpenAI’s wish for users to complete transactions within ChatGPT through its agent and Shopify’s wish for control and safe guardrails for its merchants. Who owns the agentic transaction will be a defining question within ecommerce in the coming years.
The other key dynamic involved here, of course, is whether consumers will trust an AI agent with their payment information or with other sensitive actions such as logging in to services?
The potential performance improvements are real, but reliability, trust, and seamless authentication may well remain key barriers. Unless those are resolved, users may hesitate to give agents full autonomy over transactions or personal accounts.