Suppose a development team writes code much faster with AI assistance. Releases still follow the same schedule because they wait for the same testing phase and the same approvals. What would have to change for the software to reach its users sooner?
Where the gains stop
In The Goal (1984), Goldratt and Cox describe how a bottleneck limits a system's throughput. Speeding up the work before the bottleneck does not, by itself, increase throughput. In our development example, more changes would be ready for review, while the number released would stay the same.
The first step is to identify what is holding up the work. If testing capacity is insufficient, additional capacity or automated tests may help. If preparing an approval takes too long, that step can also be improved. Neither requires the organisation to change who has the authority to decide.
The situation is different if every change, regardless of its scope or test results, must wait for the next change board meeting. That wait remains even when all the paperwork is ready. If a suitable change is to be released as soon as testing is complete, the organisation must establish who can authorise it without a board meeting. This changes a rule about who may decide.
In Change (1974, p. 10), Watzlawick, Weakland and Fisch distinguish first-order change, which operates within an existing system, from second-order change: “one whose occurrence changes the system itself”. Applied to our example, better preparation for approval is first-order change. Giving a team the authority to release certain changes without the change board is second-order change.
When agents take on tasks
This distinction matters when AI is expected to do more than accelerate individual steps. Suppose an agent can make a code change for a clearly defined task, run the relevant tests and prepare the result for release. If people still have to approve each step separately («Human in the Loop»), the time needed to complete the work remains tied to their availability.
For the agent to work independently within agreed limits, the organisation must first define which tasks those limits cover and when the agent must hand over to a person. Responsibility for the workflow must also be clear. People then set the conditions and oversee execution. This form of oversight is known as “human on the loop”.
Whether this way of working is suitable for the task has to be judged by the quality of the results and the actual time to completion. The agent's technical capability alone cannot answer that question. It does, however, open up a possibility that remains untested as long as the organisation requires the same individual approvals as before.
Who can change the rules
An organisation can change its own rules. Doing so requires a decision by someone with the authority to change them. A team whose remit is limited to following existing procedures cannot abolish a mandatory approval on its own. An outside perspective does not give an external consultant that authority either.
One possible organisational design is a distinct unit with its own decision-making authority. In their work on the ambidextrous organisation (2004), O'Reilly and Tushman describe how units exploring new approaches have their own processes while remaining connected to the established business through shared senior leadership. Applied to AI, such a unit could test which tasks agents can carry out within defined limits.
Autonomy alone is insufficient. Before the trial begins, it must be clear who will compare the results with the existing workflow and decide whether to adopt the new approach in normal operations. An existing development team can also take on this work if it is given the necessary authority. A separate lab with no connection to those responsible for operations may demonstrate a new way of working without anything changing beyond the lab.
For the development team at the start, the trial would succeed if suitable changes could be released sooner, with quality at least maintained. To adopt that way of working in normal operations, the corresponding decision-making authority must carry over as well. Otherwise, the shorter time to release ends with the pilot.
The next article examines what happens to a transformation team asked to prepare changes like these without being given the authority to pursue them: Second-Order Change: Assimilated or Rejected.
References
- Goldratt, E. M. & Cox, J. (1984). The Goal: A Process of Ongoing Improvement. Great Barrington: North River Press.
- O'Reilly, C. A. & Tushman, M. L. (2004). The Ambidextrous Organization. Harvard Business Review, 82(4).
- Watzlawick, P., Weakland, J. H. & Fisch, R. (1974). Change: Principles of Problem Formation and Problem Resolution. New York: Norton. Quotation: p. 10.