Our Take
The framing of automation as a binary (machine or human) misses the actual disruption: workflows that require humans and machines to operate together in sequence, and nobody has designed those interfaces yet.
Why it matters
If you're building AI products or deploying them in production, you're already past the "can it work" phase. The question now is how to architect handoffs between AI and human judgment without creating bottlenecks or liability gaps.
Do this week
Map your current AI workflows: identify every point where output passes from machine to human, then ask whether that human is checking work or making a new decision. Design accordingly.
The automation question has a false premise
The Financial Times piece challenges the standard framing of artificial intelligence in the workplace: "Can a machine do this job?" That binary formulation assumes displacement. But in practice, most deployed AI systems don't replace jobs; they change the shape of them.
The article suggests the real question should be different. Not whether a machine can perform a task, but how human and machine work sequences together, what judgment layers remain, and who owns the failure modes when they intersect.
The design problem is upstream
Most enterprise AI deployments today operate inside a workflow, not as a replacement for one. A radiologist still reviews the scan; the model flags regions of interest first. A contract reviewer still approves terms; the model extracts provisions. A customer service agent still handles escalations; the model routes and summarizes incoming requests.
This is not a transition state. It's the actual shape of work. And it requires rethinking how we measure success. A model that is 95% accurate at flagging anomalies doesn't reduce headcount by 95%; it changes the skill profile of the person reviewing anomalies and increases their throughput. But it also introduces new friction: false positives that waste time, false negatives that create liability, and handoff delays where information gets lost in translation between machine output and human input.
The cost isn't always labor. It's often operational complexity. You now have two systems to monitor, two failure modes to explain, and regulatory exposure on the interface between them.
Design for the junction, not the replacement
If you're evaluating AI for a workflow, stop measuring against the human baseline alone. Instead, audit the handoff points.
Ask: Where does the machine output become human input? What does the human need to see to make a confident decision? What can they not see that would change the outcome? Where does the human decision create feedback that could improve the model, and is your system capturing it?
Then ask the harder question: If this fails, who is liable? If the model confidently recommends X and the human does Y because the presentation was unclear, who owns that divergence?
Most teams skip this. They optimize for model accuracy and assume human oversight will handle the rest. It won't. Oversight at scale requires different tooling, different training, and often different hiring profiles than the original job required. Design for that now, or you will pay for it later in error rates, staff churn, and compliance exposure.