AI Innovation Architecture: who sets purpose when your systems become agents?
Dixon, R.J., 2026e. AI Innovation Architecture: Nested Purpose, Closed-Loop Dynamics, and the Chief Agency Officer. SSRN Working Paper, 52 pages, posted 23 April 2026. Available at papers.ssrn.com/abstract=6546178. In brief. The third paper in my SSRN series. The claim in one sentence: as AI systems move from tools to agents, the defining architectural challenge for an organisation is who sets purpose, who executes, who judges, and how agency is progressively handed over, and that needs a dual operating model, a set of governance rules for nested purpose, and a new seat in the C-suite.
Why I wrote it
The first two papers in this series set up a framework and an economic argument. Purpose, Execution, Judgement described how humans and AI share productive work (Dixon, 2026b). From Knowledge to Purpose argued that economic value is migrating from execution to purpose and judgement as intelligence gets cheap (Dixon, 2026d). Neither looked hard at what happens when the AI in the loop stops being a tool.
It isn't any more. The systems I now watch clients deploy set their own sub-goals, run their own optimisation loops, and increasingly shape the direction of the work they're given. The paper opens with a scenario drawn from my consultancy work: a logistics company running three of them, one optimising delivery speed, one optimising cost, one optimising driver satisfaction, and a head of product who says, "We spend more time arbitrating between our AI systems than we spent doing the work manually." That sentence is the paper. The contradiction between the three systems wasn't a bug in any of them. It was an unresolved strategic contradiction one level up, compounding at machine speed.
What it argues
The paper makes seven contributions, and I'd pick out five for anyone who isn't going to read all 52 pages.
First, the AI-enabled organisation has to run two modes at once. Business as usual is optimised for return on investment. Innovation is optimised for the rate of innovation per unit time, the objective I argued for in the Playbook (Dixon, 2025d). Govern innovation with BAU logic, demanding immediate ROI from experiments, and you kill it. Govern BAU with innovation logic and you destabilise your cashflow. The ratio between the two is a governance choice, and I propose 85 to 15 as a starting allocation, with 15 per cent the minimum viable innovation allocation. That figure comes from consultancy experience, not a controlled study, and the paper says so; it's offered for testing. Before ChatGPT, my estimate is that most organisations ran closer to 99 to 1.
Second, agency is a scale, not a switch. A custom GPT that answers when prompted has negligible agency. A deep research mode sets its own sub-purposes within a bounded task. A desktop agent runs several nested loops across your files and applications and then stops. A persistent agent has a heartbeat and doesn't wait to be asked. Understanding where a given system sits on that scale is, I argue, one of the most important pieces of AI literacy a leader can now develop. And whatever sits on the scale, the primacy of agency belongs to whoever holds the top-level purpose and exercises the final judgement. Everything below, however sophisticated, is crew.
Third, purpose nests, and it nests in more than one dimension. Inside the execution of any PEJ loop sit smaller PEJ loops, each with its own purpose derived from the level above. A team's purpose is often constrained by several higher purposes at once, which creates what the paper calls a purpose topology. When no way of doing the work can satisfy all the constraints, the topology is inconsistent, and the paper's diagnostic rule follows: topological inconsistency at level N is always a signal of unresolved strategic contradiction at level N+1. When teams say they can't reconcile competing directives, the fix is above them, not below.
Fourth, closed loops compound in whichever direction they're pointed. Every PEJ loop is a closed loop: output feeds evaluation, evaluation feeds the next cycle. The paper names three trajectories, and the two that matter most for leaders are the extremes. Pointed well, a loop is convergent innovation, and with second-order learning the system gets better at getting better. Pointed badly, it's the AI slop spiral I described in the PEJ paper, building on Willison's term (Willison, 2024; Dixon, 2026b), and it looks like productivity from the inside. Bostrom's paperclip maximiser becomes, in this frame, a loop with no higher purpose constraining it (Bostrom, 2014). A marketing loop that optimises click-through brilliantly while destroying brand trust is a mild version of the same thing. Loop hygiene, the discipline of monitoring trajectory, governing by purpose rather than method, keeping judgement real, and auditing topology, is what stops it, and it depends on one prerequisite: the loops must be inspectable.
Fifth, purpose-setting authority will be delegated to AI, progressively and in one direction, and that's safe only under three joint conditions: observability, reversibility, and alignment verification. The paper lays out a five-stage delegation spectrum from tool use to strategic partner. Human authority doesn't disappear as you move along it. It moves up.
The Chief Agency Officer
An AI agent that sets its own sub-purposes sits between HR's domain and IT's. Neither is equipped to govern the agency boundary, so there's a gap, and the paper proposes filling it with a Chief Agency Officer who owns the innovation side of the dual operating model, paired with a COO who owns BAU and a CEO who holds the ratio. The CAO's mandate has four parts: innovation, agency governance, loop hygiene, and building AI capability across the five dimensions of Talent, Purpose, AI Literacy, AI Tools, and Data Infrastructure. Every other function then reorients around the dual axis. The CFO measures ROI on one side and AI capability as an intangible asset on the other. The people function recruits for both modes. The CTO runs stable infrastructure and experimental infrastructure. Figure 5 shows the shape of it.
The seventh and last contribution is an image I've found useful with leaders: the Child Emperor, a leader holding formal authority over capability that exceeds any single human's capacity to comprehend or directly control. It isn't pejorative. It's structural, and it draws on the old distinction between formal and real authority (Aghion and Tirole, 1997). The Child Emperor's authority stays meaningful only through governing by purpose and boundary rather than reviewing every cycle, and through continuous learning, because the execution multiplier means a small improvement in the quality of human purpose-setting now unlocks a disproportionate increase in output. The bottleneck was never execution. It was purpose.
What it means in practice
Map every significant human-AI interaction, identify the purpose topology, and test it for consistency. Audit your purpose statements for embedded method, using the minimum viable constraint test: could this purpose be achieved by a radically different method than the one I'm imagining? If not, it contains method and it's capping the innovation of everyone below it. Measure the BAU to innovation ratio and report it to the board. Treat the time between spotting a new AI capability and operationalising it as a performance metric, the organisation's innovation clock speed. Delegate purpose-setting to AI first where errors are cheap and reversible.
The dual operating model sits on well-worn ground: March's exploration and exploitation (March, 1991) and the ambidexterity literature that followed. What's new is the claim that agency allocation and innovation are converging into a single activity, and the governance role that follows from it. Ten formal propositions are set out for empirical testing, and the limitations section is candid that the paper presents no empirical evidence yet. If you run agentic systems in anger and want to test any of this, I'd like to hear from you.
How to cite it
Dixon, R.J., 2026e. AI Innovation Architecture: Nested Purpose, Closed-Loop Dynamics, and the Chief Agency Officer. SSRN Working Paper. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6546178
References
Aghion, P. and Tirole, J., 1997. Formal and real authority in organizations. Journal of Political Economy, 105(1), pp.1–29. Available at: https://doi.org/10.1086/262063.
Bostrom, N., 2014. Superintelligence: Paths, Dangers, Strategies. Oxford: Oxford University Press. Available at: https://global.oup.com/academic/product/superintelligence-9780199678112 [Accessed 4 September 2026].
Dixon, R.J., 2025d. The AI Transformation Playbook: A Step-by-Step Guide to Becoming an AI-Enabled Organisation. Manchester: Dixon AI Press. Available at: https://www.amazon.co.uk/dp/1919400508.
Dixon, R.J., 2025. The AI Transformation Playbook: my first book, and why I wrote it. [online] robdixon.ai, 19 December. Available at: https://www.robdixon.ai/post/the-ai-transformation-playbook [Accessed 4 September 2026].
Dixon, R.J., 2026b. Purpose, Execution, Judgement: A Framework for Human-AI Collaboration and Organisational AI Capability in the Post-ChatGPT Era. SSRN Working Paper. Available at: https://dx.doi.org/10.2139/ssrn.6282200 [Accessed 4 September 2026].
Dixon, R.J., 2026d. From Knowledge to Purpose: The Shifting Bottleneck of Value Creation and Five Domains of Human Value as We Approach AGI. SSRN Working Paper. Available at: https://dx.doi.org/10.2139/ssrn.6396778 [Accessed 4 September 2026].
Dixon, R.J., 2026e. AI Innovation Architecture: Nested Purpose, Closed-Loop Dynamics, and the Chief Agency Officer. SSRN Working Paper. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6546178 [Accessed 4 September 2026].
Dixon, R.J., 2026. Purpose, Execution, Judgement: the paper behind the framework. [online] robdixon.ai, 10 March. Available at: https://www.robdixon.ai/post/purpose-execution-judgement-pej-framework-paper [Accessed 4 September 2026].
Dixon, R.J., 2026. From Knowledge to Purpose: what happens to human value when intelligence is cheap. [online] robdixon.ai, 23 March. Available at: https://www.robdixon.ai/post/from-knowledge-to-purpose-paper [Accessed 4 September 2026].
March, J.G., 1991. Exploration and exploitation in organizational learning. Organization Science, 2(1), pp.71–87. Available at: https://doi.org/10.1287/orsc.2.1.71.
Willison, S., 2024. Slop is the new name for unwanted AI-generated content. [online] Simon Willison's Weblog, 8 May. Available at: https://simonwillison.net/2024/May/8/slop/ [Accessed 4 September 2026].
Further reading. Added September 2026: later work that builds on this paper.
Dixon, R.J., 2026i. Stop Putting IT in Charge of AI (and 8 Other Mistakes Organisations Keep Making). [online] LinkedIn, 8 August. Available at: https://www.linkedin.com/pulse/stop-putting-charge-ai-8-other-mistakes-organisations-rob-dixon-ge0he [Accessed 4 September 2026].
Dixon, R.J., 2026. The CEO's Guide to AI: what I told a room full of CEOs in Chicago. [online] robdixon.ai, 3 September. Available at: https://www.robdixon.ai/post/ceos-guide-to-ai-chicago-keynote [Accessed 4 September 2026].
Rob Dixon is Founder and CEO of Dixon AI Ltd and lectures at Manchester Metropolitan University. He writes and speaks on AI capability, strategy, and human-AI collaboration. Bibliography · Google Scholar · ORCID · LinkedIn · dixonai.com
Comments