In plain terms
A calculator augments an accountant; a payroll system automates the monthly payslips. In the first case a person still does the work, and the tool makes them quicker and more accurate. In the second the work happens without anyone doing it, and people deal only with the exceptions. With AI the same choice comes up for every task: should it draft for a person, or act on its own?
Why it matters
The choice sets the business case and the risk. Automation removes labour from a task and demands high reliability, monitoring and a clear answer to who is accountable for errors. Augmentation keeps a person responsible and raises quality or speed, but the saving appears only if the work is reorganised around it. Decide per task, with four questions: what does an error cost, how varied are the cases, how large is the volume, and must someone answer for the decision? Augmentation is not automatically the safe option: a tired reviewer who approves everything is automation without the controls.
Example
An airline receives 1.2 million customer requests a year. Seat changes and receipt copies make up 55% of the volume; they are uniform and cheap to get wrong, so the airline automates them end to end and audits a 2% sample. Compensation claims are varied and costly to get wrong: AI summarises the case, cites the applicable rule and proposes an amount, and a service representative decides. Handling time for those claims falls from 14 minutes to 9.
Most often confused with
Automating a task vs. replacing a job
Automation acts on tasks, and a job is a bundle of tasks. A claims handler reads files, calls customers, negotiates, decides and documents; AI may automate the documenting, assist the deciding and leave the calls untouched. What follows for the job depends on how many of its tasks are affected and on what the organisation does with the freed time: more output, a redesigned role or fewer positions. That is a management decision.
Origin: The idea of augmentation goes back to Douglas Engelbart's 1962 report “Augmenting Human Intellect: A Conceptual Framework”.
Under the hood
Human-factors research treats this as a scale: Parasuraman, Sheridan and Wickens (2000) describe levels of automation across four functions, namely acquiring information, analysing it, deciding and acting, and a system can sit at a different level in each. Practical positions for AI: suggest (a person writes, AI proposes), draft (AI writes, a person edits and approves), act with approval (human-in-the-loop), act and report (human-on-the-loop, with sampling) and act unattended. Decision factors per task: the cost and reversibility of an error, the variety of cases, volume, accuracy measured on a test set, the need for judgement or empathy, and legal accountability. Known risks: automation bias and rubber-stamping; loss of skill in the people expected to handle exceptions; and a harder remaining workload once the easy cases are gone. Economics: automation is counted in cost per transaction, augmentation in throughput, quality and time to competence. Erik Brynjolfsson's essay “The Turing Trap” (2022) argues that incentives favour automation even where augmentation creates more value. Tasks move along the scale as measured reliability improves.