Last week, we shared our ‘human first’ framework for AI – a hierarchy for deciding where AI adds value and where it risks eroding quality or hindering development. In summary, we: automate the no-brainers, use discretion on tasks where AI can accelerate thinking, and protect work requiring deep human judgment.
But knowing where we might want to use AI doesn’t tell you how it can add value. And this leads us to the problem of ‘workslop’.
The workslop problem
If you’ve used the latest AI tools, you know how powerful they’ve become. You also know that despite the hype around ‘agents’ and ‘autonomous workflows’, there are surprisingly few complex tasks where AI produces client-ready work – especially in the context of low investment use of a chat interface.
This has sparked a backlash. The term ‘workslop’ captures the problem: used without discernment, AI just accelerates poor outputs. Work gets reviewed more times than it otherwise would, creating a vicious cycle where we’d have been better off without AI in the first place.
Try asking Copilot to generate a client-ready report from raw data, or a questionnaire from research questions. You’ll see the problem immediately.
Why this happens
After GPT-4, most performance gains have come from post-training improvements – essentially, spending more compute power on reasoning. Instead of making one giant leap to an output, the model chunks your task into smaller, manageable steps, and works through those.
“But I still can’t just upload my data and get a client-ready report. Why?”
Because AI isn’t an expert in breaking down every niche process in every industry in a way that leads to quality. It has a generic reasoning algorithm that won’t always match industry best practice for decomposing complex tasks. Even if you’re brilliantly smart, a non-expert can’t just go from inputs to high-quality outputs in one big leap. Unless you break a task down into the right simple steps, you’ll miss opportunities to add quality.
The fundamental point
You can’t delegate the production of high-quality complex outputs if you don’t understand the granular processes that ensure that quality. The first step is documenting those processes. This has been a strategic focus for us at IFF in the past few years, developing a comprehensive library of IFF Way guides, which help our team quickly remind themselves of the best way.
The second step is delegating those steps effectively – be that through good management practices working with a research team, or working with AI through good prompt engineering.
However, these are necessary but not sufficient steps – without the experience and expertise to review quality, you may still end up with slop. Clearly therefore, learning and development remains essential. Which is why IFF continue to invest heavily in our award-winning Trainee Research Programme – to instil solid fundamentals. We’re also committed to a real focus on ongoing development and opportunities to learn from experienced senior researchers on the job.
AI is not a magic wand that replaces expertise; it’s a tool that can replicate processes. To make quality replicable, you need granular understanding of what processes lead to quality, and granular control over their execution. Without robust process understanding, the technology is worthless to us. With them, it can be accelerated. The final step though – evaluating the quality of a complex output – still relies on a level of human experience and expertise which are yet to be reduced to an algorithm.
Our ‘north star’ remains being human first. That means recognising that human expertise is the prerequisite to getting the most out of AI. By doubling down on the ‘IFF Way’, a rigorous focus on process, and investment in learning and development, we ensure AI remains a tool for operational excellence, not a replacement for it.
Next week: why automation will never fully replace experts, and why the evaluation of quality still requires human judgment.