The implementation gap

Implementing AI as part of a broad organizational transformation means aligning people, technology and processes so that they have the best chance of working together, and of delivering the outcomes the organization seeks. The studies on how that succeeds are plentiful and consistent. The gap sits between agreeing with the findings and making them work at the desk. This article first compresses the current findings, then turns to the practical lessons we have learned working with agents.

The literature on why AI transformations fail, and on what makes the few succeed, keeps mounting, and its verdict is remarkably consistent. Technology is rarely the constraint. Deloitte names the dominant pattern layering, agents placed on top of processes built for people,¹ and reports that 84 percent of organizations have never redesigned a job or workflow around AI.² What works, in the same studies, is the reverse motion, work redesigned around the machine and workers who change how they think and judge alongside it. McKinsey finds that successful adopters spend one part on technology, three on process redesign and five on capability building, a ratio most companies invert, and traces stalled rollouts to rational human fears that classic change management never reaches.³ PwC attributes roughly 80 percent of an initiative's value to redesigned work.⁴ IBM's survey of 2,000 CEOs has four in five saying success depends more on people's adoption than on the technology, and shows organizations that redesigned their core functions around AI delivering on their objectives four times as often as those that did not.⁵ Forrester⁶ and the World Economic Forum⁷ add the governing half of the picture, autonomy earned through controls, evidence and named owners. Processes, leadership, incentives and habits of thinking decide the outcome.

Findings like these read as self-evident, and they are easy to agree with from a distance. The gap opens when the actual work starts. Assume, then, that an organization has done what the literature asks. It has invested in capable tools and rewards their use. It has established the data and security guardrails that give agents access to all necessary information within clear boundaries. It has brought its data to a workable level of quality and structure. The mandate has been given to raise productivity tenfold. Everything that follows happens after that point, at the desk where a professional and an agent meet.

Closing the gap, at the level of daily work, means doing everything that raises the chance of succeeding. The habits that follow decide how much the same tools, in the same hands, actually deliver.

Treat an agent organization as a human organization

The most useful habit is a simple one. Expect from agents the level of cognition, and the ways of working, that one would expect when delegating to a strong and trusted team. A message typed into a chat window deserves the same care as a briefing delivered in person to a room of colleagues.

When an output disappoints, resist the conclusion that the technology is not ready. Instead, curiosity is crucial here. The productive question is why it fails, and there are three likely candidates. The ask, the technology, or the task itself.

The three sections that follow take the three candidates in turn, each building on the scenario above.

Analyzing the ask

The first place to look is one’s own message, read the way a colleague would have received it.

• Was the assignment clearly formulated? Most failures trace back to instructions that are missing, vague or misformulated, and to context the agent never received.
  ExampleThe first check is whether the ask named framework A. Assuming the agent used framework B last time, an ask that says “do it like last time” points to framework B without anyone intending it.
• Was the intent visible? Agents, like people, make better decisions when they know what the work is for.
  ExampleThe analysis was meant to defend a position under review, a purpose for which the organization always argues from framework A. Told what the work is for, the agent weighs the choice the way the organization does. Without the purpose, framework B looked equally fit.
• Did the agent know the reasoning path and its key forks? An experienced professional carries a decision tree in mind, and an unnamed fork is a fork the agent takes alone.
  ExampleIf the house view is framework A whenever projections are available, that fork has to be visible to the agent. It can sit in the ask, in a playbook the agent reads, or in the first principles the organization has defined. Left invisible, the fork was taken alone, and the agent chose framework B without knowing there was a choice to defend.
• Was the approach confirmed before the work ran? The cheapest correction happens before the analysis exists. On material cases, the ask can require a short plan first, the intended method and the forks the agent sees, before any work begins.
  ExampleAsked to state its plan, the agent replies that it intends framework B because projections are missing. The mismatch surfaces in one exchange, at the cost of a minute, rather than after a full review of a finished analysis built on the wrong foundation.

Checking the technology

If the cause is not in the ask, the machinery is next.

• The obvious checks come first. The right model for the depth of the task, the required connectors switched on, no necessary tool blocked.
  ExampleWeighing framework A against framework B is a reasoning-heavy judgment. Run on a fast lightweight model, or with the connector to the underlying figures switched off, the choice tips toward the framework that needs less, and that was framework B. What looks like incompetence is configuration.
• Is the required information in the agent’s memory or knowledge base, and is it organized so the agent would actually find it? What a colleague finds by asking, an agent finds by structure, or misses.
  ExampleThe methodology note stating that framework A is preferred was stored in a personal folder the agent cannot see. To the agent, framework B was the only documented option, and choosing it looked like diligence.
• Does the weakness come from how several agents work together? Handoffs lose context between agents the way they do between departments.
  ExampleA research agent handed the drafting agent a summary in which the projections that framework A depends on were compressed away, and from what arrived, framework B looked like the only workable choice. The fix is organizational. A reviewer agent looking over the case, or a playbook rule to always confirm the method with the expert agent team, would have caught the loss.
• Or from the way a single agent works? Every task runs through a loop. The agent reads the assignment, gathers context, chooses an approach, does the work and checks the result. Each of those steps can be configured.
  ExampleThe agent picked framework B when it chose its approach and went straight on to drafting. A loop with one more step, checking the intended method against the organization’s guidance before the work starts, catches a wrong framework on every future run.

Questioning the task itself

Sometimes the ask was clear and the technology sound, and the assignment still could not have succeeded as given. This is the branch of the scenario in which the expectation itself was wrong. Framework A, the approach the professional expected, could not have been applied to this case.

Sometimes that shows only once the work is underway.

And sometimes it was visible before the message was sent.

Working with agents rewards the same qualities working with people does. Clear intent, honest root-cause thinking, and the patience to improve the system so that agent teams grow more valuable day by day.

References

1. Deloitte Insights (2026) ‘AI agents are only the beginning: The path to agentic transformation’, 12 August. Available at: deloitte.com. ↩
2. Deloitte Insights (2026) ‘From AI adoption to AI adaptation’, 9 July. Available at: deloitte.com. ↩
3. McKinsey & Company (2026) ‘How to close the agentic adoption gap’, 7 August. Available at: mckinsey.com. ↩
4. PwC (2025) ‘2026 AI Business Predictions’. Available at: pwc.com. ↩
5. IBM Institute for Business Value (2026) ‘CEOs are reshaping C-suite roles for the AI era’, 4 May. Available at: newsroom.ibm.com. ↩
6. Hopkins, B., Le Clair, C., Pollard, J., Curran, R. and Joseph, L. (2026) ‘The state of agentic AI in 2026: Companies are chasing, few are catching’, Forrester Blogs, 3 June. Available at: forrester.com. ↩
7. World Economic Forum (2026) AI Agents in Action: A Playbook for Trusted Adoption, Authorization and Scaling. Geneva: World Economic Forum. Available at: weforum.org. ↩