Software delivery is a domain I know well, I am both the subject matter expert and the agentic engineer, which makes reengineering it relatively straightforward. But the same logic applies anywhere complex human workflows exist.

Take the generation of a credit report by a credit analyst working for a credit bureau. In my experience working in the business information services space, most recently at the BIIA 2026 Annual Conference in Manila, where I spoke on a panel on applied intelligence, this is rarely a single person’s job. There are researchers who gather raw data. Analysts who structure and interpret it. Reviewers who quality-check the output. An operations head who validates before the report goes to the client. This is already a squad. The roles already encode a separation between execution and judgment, even if nobody has mapped it explicitly.

At that conference, one theme surfaced in almost every session: the industry knows AI is coming, wants the efficiency, and is deeply uncertain about accountability. If an AI system generates a credit inference and it turns out to be wrong, who is responsible? The consensus answer was clear: the human expert. Which means human accountability is non-negotiable. What is negotiable is where in the pipeline the human intervenes, at every step, or only at the moments that actually require judgment.

With the right subject matter expert involved, someone who understands the domain, the rules, the exceptions that workflow can be mapped, the execution steps identified, and agents embedded to handle them autonomously within guardrails. The analyst does not disappear. She moves from spending sixty percent of her time chasing data sources and formatting outputs, to spending that time on the inferences, the edge cases, the judgment calls that a credit bureau client is actually paying for.

There is one more thing that every AI adoption initiative underestimates, and I want to name it directly: the mindset work.

Process reengineering can be designed on a whiteboard. The technology can be built. The guardrails can be specified. But none of it will work if the people whose workflow is being redesigned do not understand what is happening to them, why it is happening, and what their role looks like on the other side.

The fear I described at the opening of Part 1: the quiet sense that you might be training your replacement does not resolve through a strategy deck or a town hall. It resolves through deliberate, sustained engagement with the people closest to the work. Not a one-time change management session, but an ongoing effort to help operational teams understand that agents executing routine tasks is not a threat to their expertise. It is a release valve for it.

I have seen this play out on both sides. In teams where the mindset work was done, where people understood the process redesign before it happened, where their own knowledge was treated as the foundation the agents were built on the adoption was faster, guardrails were better calibrated, and the efficiency gains were real. In teams where it was skipped, the same technology produced friction, workarounds, and eventually a quiet reversion to the old process with an agent sitting unused in the background.

The subject matter expert is not just a source of process knowledge. She is the person whose buy-in determines whether the reengineered process actually runs. Treating her as a stakeholder to be managed rather than a co-designer of the new workflow is the most common and most costly mistake in AI adoption.

The final point is this: the first redesign is not the last one.

As the technology matures and the team’s confidence grows, the boundary between what agents handle and what humans handle will shift. Tasks that required human judgment today because the agent was not reliable enough, the guardrails were not tight enough, the trust was not yet established, will move into the execution layer over time. The human experts will move further up the value chain. The agents will handle more, with less supervision.

This is not a transformation that happens once. It is a continuous recalibration, driven by evidence, enabled by trust, and constrained by accountability at every stage.

The goal was never to move from human-centric to agent-centric. That framing creates the fear I described at the start and that fear is the biggest implementation risk in the room. The goal is a process where agents do what they do best: consistently, at scale, within defined bounds and humans do what they do best: the things that require experience, judgment, and accountability that cannot be delegated.