Skip to content

The Handoff to the Owner

A handoff is finished when the owner's people can run and change the system without the team that built it. Until then, the delivery team has only moved on, and what it built stays behind with nobody who understands it.

This is the second handoff in delivery, from the delivery team to the owner who runs the result. When to Embed an Engineer names what kills pilots at this point: nobody owns the system once its builders leave. Staffing the Delivery Team covers the first handoff, including the owner and the delegate this page relies on.

Why AI systems need an owner after launch

Launch starts the owner's job. Software usually costs more to keep running and changing than it did to build. Maintenance typically consumes 40 to 80 percent of lifetime cost, averaging about 60 percent (Glass, 2003).

AI systems add upkeep of their own. Researchers at Google found it common for machine learning systems to incur "massive ongoing maintenance costs" (Sculley et al., 2015). One reason is that the world they depend on keeps moving: "the external world is rarely stable" (Sculley et al., 2015).

US banking regulators already expect banks to monitor traditional models as products, clients, data, and markets change (OCC, Federal Reserve, and FDIC, 2026). Their 2026 guidance leaves generative and agentic AI out of scope (OCC, Federal Reserve, and FDIC, 2026). For those systems, the owner has to decide how the watching gets done, under the organization's own governance.

Plan for this before launch, not after. Umbrex lists a monitoring plan among a project's few non-negotiables (Umbrex, 2025).

What has to transfer

The handoff is complete only when each of these has a named person and a line in someone's budget:

  • The running costs. Hosting, model usage, and licenses move from the program's build budget to the owner's operating budget.
  • Incidents. Name the person who takes the call when the system gives a wrong answer on a Monday morning.
  • The tests. The team built checks from the delegate's domain questions. Name who reruns them after any change to the model, the prompts, or the data.
  • Changes. Name who may change prompts, rules, and data sources, and who approves those changes.
  • The record. Logs, decisions, and the reasons the team built the system the way it did.

These map onto The Three Accountabilities. The owner takes the Owner seat and answers for the outcome. An architect answers for the system. A verifier checks that it still works, and the delegate, whose questions shaped the tests, is the natural choice. If you cannot name all three people, the handoff has not happened.

Test the handoff before the team leaves

Before the delivery team leaves, ask the owner's people to make one real change without them. They might update a rule, add a new case, or handle an incident from start to finish. If they cannot, the knowledge is leaving with the builders.

Run the test while the builders are still there to help, because a failed test after they have moved on becomes a stalled system.

When the team should not leave

Some work should stay with the team that built it. When a use case is central to how a business unit competes and changes often, keep a small team with that unit permanently. Amazon's Werner Vogels states the principle plainly: "you build it, you own it" (Vogels, 2026).

Microsoft's guidance for a maturing center of excellence also moves delivery closer to the teams that carry the work. It says to transfer AI delivery to the platform teams and turn the center into an advisory group (Microsoft, 2025).

A handoff suits a workflow that stays stable once built. A permanent team suits one that changes every month.

Measure it a year after launch

Staying in production is hard even for experienced organizations. Gartner published survey results on this in 2025. At high-maturity organizations, 45% of leaders said their AI initiatives stay in production three years or more (Gartner, 2025). At low-maturity organizations, 20% said the same (Gartner, 2025). Both figures are leaders reporting on their own programs.

So judge a use case a year after launch, not on launch day. Measuring Adoption covers how to measure real use rather than logins. Then ask three questions: is it still in use, is it still correct, and who fixed it the last time it broke? If the answer to the last question is the delivery team, the handoff did not happen.

Sources

  • Robert L. Glass — Facts and Fallacies of Software Engineering (Fact 41), Addison-Wesley (ISBN 0-321-11742-5), 2003. Maintenance typically consumes 40 to 80 percent (average, 60 percent) of software costs. Therefore, it is probably the most important life cycle phase of software. View source · verified 2026-07-01 · primary
  • D. Sculley et al. (Google) — Hidden Technical Debt in Machine Learning Systems, Advances in Neural Information Processing Systems 28 (NIPS 2015), 2015. Using the software engineering framework of technical debt, we find it is common to incur massive ongoing maintenance costs in real-world ML systems. View source · verified 2026-09-13 · primary
  • D. Sculley et al. (Google) — Hidden Technical Debt in Machine Learning Systems, section 7 (NIPS 2015), 2015. Experience has shown that the external world is rarely stable. This background rate of change creates ongoing maintenance cost. View source · verified 2026-09-13 · primary
  • OCC, Federal Reserve, and FDIC — Supervisory Guidance on Model Risk Management, Bulletin 2026-13, section V, 2026. Ongoing model monitoring involves an evaluation of the extent to which a model is performing as expected given potential changes in products, exposures, activities, clients, data relevance, or market conditions. View source · verified 2026-09-13 · primary
  • OCC, Federal Reserve, and FDIC — Supervisory Guidance on Model Risk Management, Bulletin 2026-13, footnote 3, 2026. Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance. View source · verified 2026-09-13 · primary
  • Umbrex — Chief AI Officer Playbook — The Chief AI Officer Mandate, 2025. Define a minimum viable lifecycle with a few non-negotiables: named business owner, baseline, approved data access, risk review, monitoring plan, and adoption approach. View source · verified 2026-09-13 · primary
  • Werner Vogels, All Things Distributed — A return to two-pizza culture, 2026. They were expected to own a problem and deliver with autonomy, and ownership meant the same thing it has always meant at Amazon: you build it, you own it. View source · verified 2026-09-13 · primary
  • Microsoft — Cloud Adoption Framework — Establish an AI Center of Excellence, 2025. Transfer AI delivery to the platform teams... Replace the CoE gatekeeper model that blocks work with an advisory group that sets guardrails. View source · verified 2026-07-02 · primary
  • Gartner — Gartner Survey Finds 45% of Organizations With High AI Maturity Keep AI Projects Operational for at Least Three Years, 2025. Forty-five percent of leaders in organizations with high AI maturity said their AI initiatives remain in production for three years or more to ensure sustained impact and value, according to a survey by Gartner, Inc. This compares to only 20% in low-maturity organizations. View source · verified 2026-09-13 · ⚠ secondary mirror