Go back to publications

Five Anchors: Holding Intent Invariant from Discovery to Production in Enterprise AI Delivery

16 September 2026 M. Golombeck Technical Report 10.13140/RG.2.2.31649.36969 English
#Enterprise Architecture#Requirements Engineering#Traceability#Solution Architecture#Project Governance
PDFResearchGate

Most failed enterprise AI projects have no villain and no obviously bad decision in them. They drift, one reasonable step at a time. These are the five checks I run to stop it, and the one property that makes them work as a system rather than a checklist. This paper introduces the term intent drift and treats it as a failure of traceability rather than of engineering.

The framework

Five anchors, each binding a translation to the counterpart whose knowledge it needs, the artefact it produces, and the condition under which the work may continue or must stop.

  1. Separate ambition from requirements, with the business sponsor.
  2. Establish what the data will bear, with data owners and IT operations.
  3. Agree what “good enough” means, with end users and the risk functions that can veto deployment.
  4. Design the smallest sufficient architecture, with enterprise architecture and security.
  5. Validate the economics, with the sponsor and finance.

None of the five is new, and the paper does not pretend otherwise. The contribution is the invariant that holds across them: the measurement closing an initiative has to resolve back to the ambition that opened it. That property is formalised as a chain of justification maps, which makes two things checkable that prose runs together. Scope creep is work in the system that nothing asked for. Quiet de-scoping is part of the ambition that fell out along the way. They have different remedies and are routinely called the same thing in a post-mortem.

What it costs and what it returns

The framework adds process, so it has to justify the overhead. Rather than assert a savings figure, the paper models the anchors as an option to abandon and derives the break-even termination rate: the share of initiatives that must turn out non-viable before the checks pay for themselves. Across plausible parameters that threshold lands between two and thirteen per cent.

Worked through a portfolio of twenty initiatives, one dud a year covers the cost of running the anchors on everything. Every dud after the first is return.

Limitations

There is no empirical validation here. The framework is derived from practice and argued analytically, with no controlled comparison against CRISP-DM or stage-gate governance, and no measurement of the one parameter the economics depend on. The paper states this plainly rather than burying it.

You can also find a shorter, less formal version of this publication on my blog.

Cite this work

@techreport{golombeck2026fiveanchors,
  author      = {Golombeck, Marius},
  title       = {{Five Anchors: Holding Intent Invariant from Discovery to
                 Production in Enterprise AI Delivery}},
  institution = {Independent},
  address     = {Munich, Germany},
  year        = {2026},
  month       = sep,
  language    = {english},
  doi         = {10.13140/RG.2.2.31649.36969},
  url         = {https://doi.org/10.13140/RG.2.2.31649.36969}
}