The bottleneck is the specification, not the algorithm
The main difficulty in automating business processes with artificial intelligence would not lie in the capabilities of the models, but in organizations’ struggle to translate business needs into executable instructions without ambiguity. This is the thesis presented by Rachel Woods, entrepreneur and former researcher at Facebook AI, in an interview with Isar Meitis, host of the Leveraging AI channel, as reported by AI4Business. According to Meitis, most professionals use generative tools episodically, like an enhanced web search, without integrating them into structured processes. Woods and Meitis converge on one point: defining requirements absorbs the predominant share of the work, and those who know how to do it would have already completed most of the path toward the result.
The four-component structure
The playbook is the model proposed for codifying procedures: according to Woods, it is articulated in four elements valid for any type of activity. The first is the trigger, i.e., the event or temporal condition that starts the process. The second is the inputs, namely the data, documents or raw notes necessary for execution. The third is the steps, the detailed operational sequence with instructions, recipients, evaluation criteria and formatting: in the case of email triage, a playbook can reach eleven steps with decision rules. The fourth is the output, the expected final deliverable, whose precise configuration safeguards quality standards and facilitates integration into broader flows.
This structure is accompanied by a principle of decomposition: language models perform better on simple, circumscribed tasks and are more likely to err when faced with articulated directives. Meitis compares sub-tasks to Lego bricks that can be assembled into larger procedures. Woods invokes the analogy of onboarding a human collaborator: you do not entrust all the work on the first day; you teach one procedure at a time.
Autonomy with a defined perimeter and vendor independence
On the organizational level, Woods argues that autonomy without constraints produces unpredictable results, whereas a clear action perimeter makes it possible to delegate more work. Her company, composed of ten human collaborators, would manage more than one hundred active playbooks in the backend in areas such as marketing, sales, service delivery and leadership. Hence the rule summarized as owning the playbook and renting the technology: operational logic remains an internal asset, while platforms and models change rapidly.
Several aspects remain to be verified: the material does not document measured results, costs or implementation times, and the description of the one hundred playbooks is a statement by the interviewee, not independently verified data. The mention of Google among the companies cited reflects only Woods’s professional past and not a role in the initiative. The method should therefore be read as an operational proposal to be tested, not as consolidated evidence.
