Global Advisors designed and built an internal consulting delivery platform in which humans and AI work through the same explicit reasoning structure. The platform models the path from client context and engagement scope through problems, issues, hypotheses, analyses and presentation outputs. It binds analytical artefacts to downstream deliverables and makes state, provenance, review and access part of the product.

The work addressed a central risk of generative AI: fluent output can conceal weak method, poor scoping and missing evidence. The response was not a larger prompt. It was software that constrains work within a visible consulting argument.

Maturity: operational internal workbench with active deployment; several advanced template, orchestration and review capabilities remain in development.

Traditional consulting delivery distributes structure across senior judgement, working files, meetings, analysis code and presentation decks. That model can produce excellent work, but it makes method difficult to reuse and creates a supervision bottleneck. Free-form AI can accelerate drafting while making the underlying weakness worse: plausible material appears before the problem, hypothesis and evidence are properly linked.

Global Advisors needed a system that could:

• make the consulting argument visible and editable;
• constrain AI and less-experienced contributors to relevant context;
• reuse analysis rather than regenerate it for every slide;
• preserve evidence, version and freshness;
• protect reviewed or frozen work;
• allow scoped external contribution without exposing an entire engagement; and
• keep presentation output connected to the analysis that supports it.

At the centre is a fixed-depth hierarchy representing the consulting problem. Each node has a type, parent context, state, references and downstream relationships. Selecting a node opens a workspace appropriate to that level of the argument.

The hierarchy is simultaneously:

• a navigation model;
• a reasoning scaffold;
• a context boundary for AI;
• an access-control surface;
• a state and review model;
• a memory structure; and
• a dependency graph for analyses and slides.

This was the most important product decision. Tools and models operate inside the structure; they do not replace it.

The implementation moved well beyond a concept demonstrator:

• a relational system of record for engagements, hierarchy nodes, analyses, outputs, references and presentation objects;
• typed interfaces for tree loading, node editing, movement, ordering, dependencies, consumers and versions;
• a browser workbench with hierarchy navigation, context-sensitive editors, search, filters and batch operations;
• governed draft, refine and commit flows tied to a selected node and its permitted context;
• revision-conflict protection and atomic readiness updates so a partial batch cannot silently appear complete;
• provenance records linking generated or edited content to a tool run, source context and node version;
• stale and frozen states that affect behaviour rather than serving as decorative labels;
• reusable analytical outputs with explicit downstream consumers;
• presentation objects bound to governed outputs instead of detached copied images;
• a deck-preview workspace that keeps slide content and non-slide reasoning visible together;
• a companion inside the native presentation environment for binding objects, capturing selection, invoking governed AI actions and checking templates;
• role- and engagement-scoped access, including controlled onboarding and approval evidence;
• API-first boundaries so browser, presentation, bot and workflow clients use the same governed core; and
• deterministic evaluation across source fidelity, factual consistency, patch validity and presentation compliance.

The platform rests on a practical observation: an inexperienced consultant and an unconstrained agent often fail in similar ways. Both can create output that is confident, weakly reasoned, out of scope and difficult to audit.

The same harness helps both:

• visible problem structure reduces context drift;
• templates distribute method and standards;
• node scope limits what a contribution is meant to answer;
• states make review and readiness explicit;
• provenance makes claims traceable;
• reusable outputs discourage repeated, inconsistent calculation; and
• dependencies expose what becomes stale when an input changes.

AI is therefore embedded in the delivery system rather than placed beside it as a general-purpose chatbot.

Native presentation work cannot be ignored

Consultants still deliver through presentation software. Building a separate AI interface would have forced users to choose between governance and their normal authoring environment. The companion surface kept native editing while round-tripping governed identities, context and validation.

Access must follow the work object

External or specialised contributors may need one task, not the engagement. The hierarchy provided a natural unit for scoped context and contribution, reducing unnecessary disclosure.

Leaders considering AI-enabled professional work should ask:

• What is the explicit structure of good work in this domain?
• Can both people and AI operate inside that structure?
• Which outputs should become reusable assets?
• How are evidence, state and freshness preserved into the final deliverable?
• Where is human judgement required?
• How will the system change supervision, roles and capability development?

The strategic opportunity extends beyond task automation to redesigning how quality work is produced and improved.