Design the workflow before scaling the agent

What happened?

DOCUMENTED: McKinsey published “Stacking the odds”, recommending that companies redesign processes, roles and governance around agentic AI rather than place an agent on top of an existing workflow. It is consultancy analysis, not a neutral impact study.

Why does it matter?

LEARNAI ANALYSIS: Many pilots test whether a model can perform a task. Production also requires queue handling, exceptions, data quality, ownership and integration. Without those changes, the pilot becomes another layer rather than a simpler process.

What does it mean for the reader?

PRACTICAL CONSEQUENCE: Map the process before and after the agent. Measure cycle time, errors, human rework and unit cost, not only the model's response rate.

Anthropic measures AI's contribution to AI development

What happened?

DOCUMENTED: Anthropic published a method for measuring the pace of AI development. Its own index says Claude led 26 per cent of work on new models in August, up from less than one per cent in February. “Led” means a task moved from a short instruction to a finished result under human supervision. Anthropic created the metric, and it has not been independently validated.

Why does it matter?

LEARNAI ANALYSIS: The curve says something about the organisation of work, but not necessarily autonomous intelligence. Better tools, more suitable tasks and changed recording practices may also contribute. The company's definition must travel with the number.

What does it mean for the reader?

PRACTICAL CONSEQUENCE: Create an internal measure of agent work that distinguishes generated output, approved output and realised business value.

Novo Nordisk brings Claude into the laboratory

What happened?

DOCUMENTED: Tim Frank Andersen reported on 17 September that Novo Nordisk will test Claude with access to scientific databases for drug discovery. The newsletter also says Claude has been used to write trial reports. The package contained no public timetable or contract value, and a faster reporting workflow does not itself shorten a clinical trial.

Why does it matter?

LEARNAI ANALYSIS: The case places generative AI closer to the company's core scientific work. Benefits may arise in literature search and documentation, while scientific and regulatory approval still requires qualified people.

What does it mean for the reader?

PRACTICAL CONSEQUENCE: Separate information retrieval, analysis and decision-making. Record the data sources and require expert approval at every transition.

Capital requirements become a vendor risk

What happened?

DOCUMENTED: The week's newsletters described several large financing needs. Tim Frank Andersen cited a Financial Times report that OpenAI expects cumulative cash burn of $278 billion by 2030, while SoftBank reportedly borrowed $11.9 billion to support its OpenAI investment. The underlying documents were not available in the package, so these figures remain second-hand reports.

Why does it matter?

LEARNAI ANALYSIS: Model price and features are not the only vendor criteria. Capital requirements, infrastructure commitments and possible price changes affect long-term operations.

What does it mean for the reader?

PRACTICAL CONSEQUENCE: Preserve data portability, document model dependencies and test an alternative supplier for critical workflows. That is ordinary vendor management, not a prediction about any company's survival.