The redesign that stopped halfway

1.What the report argues

The foreword of Bain’s Technology Report 2026 sets out the thesis in two sentences. The first is that “productivity and growth depend on how quickly existing organizations can absorb AI and new AI-native organizations can grow.” The second is that “it’s no longer only about access to the best technological capabilities”.

The chapter on the AI native enterprise develops the point. It argues that “the most prominent barriers to AI adoption today are organizational, not technical”, and that when enterprises are asked what holds them back, “they point mostly to technical and talent worries”. In Bain’s reading, most enterprises “remain in the pilot phase of AI absorption”. The work it puts first is process redesign and the modernisation of data and applications, and it argues that the real advantage will come from making the organisation transparent to machines and employees, listing, alongside systems, standards and workflows, its “decision rights”.

That chapter is framing, not evidence. Its survey prints no base, and its stronger claims rest on Bain’s own client experience. What makes the argument worth a board’s time is where the next chapter takes it.

2.Where the gain stalls

The chapter on software organisations opens with a sentence worth testing well beyond software:

“Most companies have changed the easy things and left the hard ones untouched.”Bain & Company, The Half-Finished Redesign, Technology Report 2026, p. 58

The easy things are visible: smaller teams and changing roles. On Auron’s reading, the hard thing is the path a piece of work has to travel before anyone may act on it. Bain puts it plainly: “Code ships faster than ever, but the approvals, handoffs, and escalations around it haven’t moved.” The consequence follows in the next line: “So, the gains stall before they reach a customer.” And the economic point is that “speeding up engineering is wasted if the rest of the organization can’t act” on what it produces.

A later chapter on agentic software development describes the same mechanism from the engineering side. For companies that speed up code generation without redesigning the surrounding process, it says, “the system reverts to the throughput of its slowest human checkpoint.”

For illustration, in a bank, an insurer or a MedTech manufacturer, the equivalent checkpoints are familiar. They include the change advisory board, the credit committee, model approval and claims sign off. Each exists for a reason. The question is whether each still sits in the right place, on the right cadence, once part of the work arrives from an agent rather than a person.

3.Why decision rights are the hard half

Bain offers a tentative reason why this is the step companies skip, “possibly because this is less about tooling and more about redistributing power.” The chapter also reports that central decision making persists, and that this “applies across different levels of AI maturity”.

The remedy Bain proposes is not another reorganisation. It is “distributing decision rights, defining where humans decide, where agents act, and who owns the outcome.” The order matters: “Structure changed first. Accountability has to follow, or the flatter organization simply relocates old bottlenecks.”

Auron’s reading adds one observation for the ROI case. Inside a pilot, the team usually holds the decision. At scale, the decision returns to the standing committee. That is one reason a pilot can look better than the rollout that follows it: the pilot measured the work, while the rollout inherits the old decision path. A business case that prices the technology and the change effort but not the decision path is pricing the easy half.

4.Switzerland and the EU: the supervisors already ask who decides

For regulated firms this is not only an efficiency question. FINMA Guidance 08/2024, Governance and risk management when using artificial intelligence, dated 18 December 2024, is almost two years old and remains the Swiss baseline rather than new law. In it, FINMA reports that it assessed the “definition of responsibilities and accountabilities for the development, implementation, monitoring and use” of AI. It observed that the “development of AI applications is often decentralised, making it challenging to implement consistent standards”, and that “only a few supervised institutions carry out an independent review of the entire model development process”.

The EU AI Act states the same question in legal form for high risk systems within its scope. Article 26(2) requires deployers to “assign human oversight to natural persons who have the necessary competence, training and authority”. Article 14(4) requires the system to be provided so that the people assigned to oversight are enabled, as appropriate and proportionate, among other things to “disregard, override or reverse the output of the high-risk AI system”, as point (d) puts it. Under Regulation (EU) 2026/1744 (CELEX 32026R1744), which amends Article 113 of the AI Act, these obligations apply to stand alone Annex III systems from 2 December 2027.

On Auron’s reading, the two texts make a point that is easy to miss. The word that carries the weight is authority. A firm that names an overseer with the duty to review but not the authority to override has assigned a task, not a decision right.

The practical consequence, on Auron’s reading, is that one map can serve two purposes. The document that shows where humans decide, where agents act and who owns the outcome is the working tool for moving decision points with the workflow, and it also records the responsibilities and the oversight that both texts address.

5.What this means for your organisation

Five questions a board or executive committee can ask of any AI workflow that is about to be funded to scale:

1. For this workflow, can we name where a human decides, where an agent acts and who owns the outcome?
2. Which approval, escalation or sign off does the workflow still pass through, and has its cadence changed since the pilot?
3. Does the person we have named for oversight hold the authority to override, or only the duty to review?
4. Where development sits in the business and standards sit in the centre, who reconciles the two, and how quickly?
5. When did an independent party last review one of our material AI applications from development to use?

6.How Auron approaches it

Auron treats the decision path as part of the business case rather than as an afterthought of governance. Before a workflow is funded to scale, Auron maps where humans decide, where agents act and who owns the outcome, moves the approval points with the workflow rather than leaving them on the old cadence, and records the map so that the same document serves the steering committee and the supervisor. The aim is a shorter path from pilot to P&L and a lower cost of control, because oversight is designed once, with the workflow, rather than added later around it.

7.Limits of this reading

• The source is a consultancy report. Bain states that it has not independently verified the information it relies on. Auron grades the full report C and the two chapters this piece leans on most at D, which is why it uses them for argument and framing only and quotes no figure.
• The software organisation chapter draws on a Bain survey of software companies and, for its finding on central decision making, on a separate Bain brief. Its evidence describes software companies; applying it to banks, insurers or MedTech firms is Auron’s reading, not Bain’s finding.
• Nothing here shows that redistributing decision rights produces a measured return. The case is argued, not measured.
• The regulatory reading is Auron’s framing, not legal advice. Articles 14 and 26 of the AI Act apply to high risk systems within the Act’s scope, on the dates as amended by Regulation (EU) 2026/1744. FINMA Guidance 08/2024 is almost two years old and remains the Swiss baseline rather than new law.

Auron Advisory is formed to build the intelligent enterprise: organisations where AI and disruptive technologies not only increase productivity but open new markets and materially improve the ones already there.

Sources