<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Albert Adusei Brobbey on Medium]]></title>
        <description><![CDATA[Stories by Albert Adusei Brobbey on Medium]]></description>
        <link>https://medium.com/@albertaduseibrobbey?source=rss-4cf50d8d7232------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*PpTxHhB5bhkJeDFAx1RDsQ.jpeg</url>
            <title>Stories by Albert Adusei Brobbey on Medium</title>
            <link>https://medium.com/@albertaduseibrobbey?source=rss-4cf50d8d7232------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 15:08:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@albertaduseibrobbey/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Your AI Agent Remembers Things. Who’s Securing Those Memories?]]></title>
            <link>https://medium.com/@albertaduseibrobbey/your-ai-agent-remembers-things-whos-securing-those-memories-97f45fbe6dab?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/97f45fbe6dab</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Tue, 30 Jun 2026 20:28:29 GMT</pubDate>
            <atom:updated>2026-06-30T20:28:29.020Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PINrZSj6n9RculhRIV_d_w.png" /></figure><p>I’ve spent the last several months thinking about a problem that almost nobody in AI security is talking about yet: what happens when AI agents start remembering things across sessions, across devices, and across infrastructure tiers?</p><p>Not remembering in the human sense. Remembering in the systems sense. Vector databases, episodic memory stores, long-term retrieval indices, cloud-to-edge replicas. The agent you deploy in a hospital, a supply chain operation, or a customer service center doesn’t just respond to prompts anymore. It accumulates context. It writes memories. It retrieves them later to inform decisions. And increasingly, those memories get synchronized between a cloud control plane and edge devices that sit closer to the action.</p><p>Here’s the part that concerns me. The AI security community has done excellent work on prompt injection, data poisoning, and model-level risks. But persistent runtime memory doesn’t fit neatly into any of those categories. It’s not training data, because it changes after deployment. It’s not just a prompt, because it persists across sessions and infrastructure tiers. It’s not just a database, because when retrieved, it becomes part of the model’s reasoning context and sometimes its action rationale.</p><p>So what is it? It’s a behavioral influence layer. And right now, most deployments treat it like an ordinary database with a vector index on top.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*r9zk_sL7b_wQBSJGyTbkRg.png" /><figcaption><em>Figure 1. How memory moves between cloud and edge in AI agent deployments, and what can go wrong.</em></figcaption></figure><p><strong>The problem with “just sync it”</strong></p><p>Think about a cloud-to-edge AI deployment. A clinical agent at a hospital edge node keeps patient-specific memory local while synchronizing protocol updates from the cloud. A manufacturing agent caches equipment history near the production line while sending incident summaries to a central operations center. A smart-building agent synchronizes sensor data, tenant preferences, and maintenance tickets across regional gateways.</p><p>In all of these cases, memory records move between tiers. And as they move, something can go wrong that has nothing to do with the model itself.</p><p>A memory record written by an untrusted source gets replicated to an edge node and later retrieved as if it were authoritative clinical guidance. A summary compresses three source memories into one, and in the process, the identity of who originally wrote those memories disappears. A record gets revoked at the cloud after an incident, but the edge node is offline during a network partition, so it keeps retrieving the revoked memory for another six hours. A vector search returns a memory from a different tenant because the embedding was semantically similar to the query, even though the requester had no authorization to see it.</p><p>None of these are model failures. They’re systems failures. And they happen in the gap between “the data synced correctly” and “the data should have been allowed to influence this decision.”</p><p><strong>What I’ve been working on</strong></p><p>My co-author and I have been developing a threat model and mitigation architecture specifically for this problem. The core idea is simple, even if the implementation is not: treat every memory record as a signed envelope rather than a bare vector-store entry.</p><p>That envelope carries provenance (who wrote this), authority scope (who is allowed to retrieve it and for what purpose), causal metadata (what version is this relative to other writes), lineage (if this is a summary, what source records produced it), freshness (when does this expire or require re-validation), revocation state (has this been withdrawn), and a signature binding all of those fields together.</p><p>Synchronization propagates the envelope, not just the payload. Retrieval checks the envelope before admitting the memory into the model’s context. The check is calibrated by action risk: a low-risk personalization query might tolerate bounded stale memory, but a clinical decision or a financial authorization requires complete provenance, current revocation state, and verified lineage.</p><p>The principle is that <strong>semantic similarity is not authorization</strong>. Just because a memory is relevant to the query doesn’t mean it should be allowed to influence the answer.</p><p><strong>What the Signed Memory Envelope Contains</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/650/1*n2lwArxx4K7RVOIIEVWXpA.png" /></figure><p><strong>What the prototype showed</strong></p><p>We built a runnable prototype with one cloud store and six edge stores, using SQLite with vector search, signed envelopes, revocation propagation, and retrieval gates. We compared three designs: baseline synchronization (the way most systems work today), signed hash-log synchronization (a reasonable improvement), and our full design.</p><p>Baseline synchronization returned poisoned memories in nearly half of all retrievals. Signing the hash log didn’t help, because it validated record integrity without checking whether the record should have been retrieved for that requester, at that risk level, at that moment. Our design brought poisoned retrieval down to under 9%, eliminated cross-scope retrieval entirely, and reduced unsafe high-risk retrieval to under 8%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/840/1*2OW_TETMLJKI5E3YKO9z7g.png" /><figcaption><em>Figure 2. Unsafe retrieval rates across three synchronization designs (20 repeated trials, 95% bootstrap CIs).</em></figcaption></figure><p>The cost is a larger metadata envelope per record and a 2% retrieval block rate. For most deployments, that’s a trade worth making.</p><p><strong>Why this matters now</strong></p><p>Every major AI lab is building agents with persistent memory. OpenAI, Anthropic, Google, and dozens of startups are shipping systems that remember, plan, use tools, and act across sessions. Edge AI is growing fast, driven by latency requirements, data residency rules, and the reality that not everything can round-trip to a centralized cloud endpoint.</p><p>The intersection of those two trends, persistent memory and distributed deployment, is where the next class of AI security failures will come from. Not because the models are misaligned, but because the infrastructure that feeds memories into those models wasn’t built with the right security properties.</p><p>I think the AI security community needs to start treating vector stores, episodic memory, and synchronization pipelines the way we already treat identity systems, key management, and audit logs: as security-critical infrastructure that requires provenance, access control, freshness enforcement, and revocation.</p><p>The paper describing this work in full is currently under peer review. I’ll share more when it’s published. In the meantime, if you’re building or deploying AI agents with persistent memory, especially across cloud and edge tiers, I’d be interested in hearing how you’re thinking about memory security. My sense is that most teams haven’t gotten to it yet, and that’s exactly why it matters.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97f45fbe6dab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When AI Makes the Wrong Call: Who Is Accountable?]]></title>
            <link>https://medium.com/@albertaduseibrobbey/when-ai-makes-the-wrong-call-who-is-accountable-42e17650419d?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/42e17650419d</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-governance]]></category>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Thu, 07 May 2026 15:21:58 GMT</pubDate>
            <atom:updated>2026-05-07T15:21:58.241Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B7tqt0sddBmXUXnkruzCbg.png" /></figure><p>Last year, a hospital deployed an AI system to optimize patient flow. The engineers built it well. The administrator approved it properly. The operations team monitored it daily. Six months later, a retrospective audit found that the system had been quietly deprioritizing elderly patients in non-emergency queues, not because anyone told it to, but because the efficiency metrics it chased happened to correlate with faster throughput of younger patients.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VUhnkyNoJQaciSA8GtyNrg.png" /></figure><p>One elderly patient died during that period.</p><p>When the hospital’s legal team started asking who was responsible, they hit a wall. The engineers said the system performed to spec. The administrator said she relied on vendor documentation. The operations team said the summary reports showed nothing unusual. Everybody was technically right. Nobody was clearly accountable. And the patient was still dead.</p><p>This is not a story about a bad AI system. It is a story about what happens when organizations deploy systems that make consequential decisions without first deciding which human owns the accountability for those decisions.</p><p><strong>The governance gap nobody is talking about</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LRsQCrjQ5VWyj8OhRi9UsQ.png" /></figure><p>There is a lot of conversation right now about AI safety, explainability, fairness, and bias. These are real problems worth solving. But they are mostly technical problems. The accountability problem is a governance problem, and it sits in a different category.</p><p>When an agentic AI system, the kind that plans, reasons, uses tools, and executes multi-step tasks with minimal human supervision, does something harmful, existing frameworks struggle to answer a basic question: which human, specifically, bears responsibility for this outcome?</p><p>McKinsey surveyed roughly 500 organizations in early 2026 and found that fewer than one-third had reached meaningful maturity in agentic AI governance. Sixty-nine percent of AI governance experts told MIT Sloan Management Review that existing management frameworks are not built for what agentic AI actually does.</p><p>They are right. They are not.</p><p><strong>Why agentic AI breaks traditional accountability</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*muY2MKfLA9gig3JQ7QoTHg.png" /></figure><p>Traditional accountability assumes three things: a human made a decision, that human knew the likely consequences, and the decision caused the outcome. Agentic AI breaks all three simultaneously.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k3to3HttgP5gvueNtGAuWA.png" /></figure><p>These systems operate over days and weeks, not seconds. They adapt to their environments in ways nobody fully predicted. They hand off tasks across networks of specialized agents, so the causal chain between a human’s original instruction and the final outcome can pass through five layers of delegation before anything consequential happens. And they commit real resources, send real communications, and modify real systems while doing it.</p><p>When something goes wrong in that environment, blame tends to collapse onto whoever is standing closest to the incident. Usually someone with the least power to have prevented it. Researchers call this the moral crumple zone: accountability by proximity rather than accountability by position. It produces the wrong incentives, the wrong investigations, and the wrong fixes.<br><strong>What the research proposes</strong></p><p>I recently completed a paper that proposes a different governance architecture. The Distributed Accountability Framework for Agentic AI, DAFA, is built around a simple premise: accountability for agentic AI does not live in one person or one role. It is distributed across a chain of human decisions, and effective governance requires clarity about which human owns what at each point in that chain.</p><p>DAFA has four layers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F04F5RZ6K_vmOaQnPyPE5Q.png" /></figure><p>The Design layer is where AI architects and product owners sit. Their obligation is specification integrity: documenting what the system can do, what objective it optimizes for, and which humans can stop it when something goes wrong. Most organizations treat this as a documentation task. It is a governance obligation. An override architecture that exists on paper but fails at machine speed is not an override architecture.</p><p>The Deployment layer is where business unit leaders sit. Their obligation is contextual accountability: verifying that the system’s design-time specifications match the real-world environment into which it is being deployed. Most organizations skip this. They treat deployment as a configuration event rather than a governance decision. The autonomy calibration decision, how much independent action this agent is authorized to take in this specific context, should be documented and reviewed periodically. Most organizations make this decision once, implicitly, and never revisit it.</p><p>The Operations layer is where I propose a new role: the AI Steward. Not a technical expert. A designated person, with explicit authority and specific information access, who holds accountability for behavioral monitoring of one or more agentic AI systems. This person needs to know what the agent is supposed to do, what anomalous behavior looks like, and what they are authorized to do when they see it. Most organizations do not have this role. Creating it does not require a new department.</p><p>The Outcomes layer is where executive leadership sits. Their obligation is institutional memory: maintaining a structured record of what their agents have done, what went wrong, what was changed, and what was learned. Organizations deploying new agentic systems without that institutional knowledge will make the same governance mistakes repeatedly.</p><p><strong>A new concept: accountability drift</strong></p><p>I also introduce a construct I have not seen named in the existing literature: accountability drift. It describes the structural erosion of an organization’s practical ability to assign and enforce accountability for AI behavior, which happens gradually as agents become more autonomous, time horizons extend, and delegation depth increases. It is not the same as governance failure. An organization can have perfectly intact governance structures and still experience accountability drift, because the agents they are supposed to govern have moved beyond the structures’ design parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RXPUVIoNYFmvJwiJl0mCYA.png" /></figure><p>To operationalize the framework, the paper also proposes the Agentic Accountability Index, a 16-item diagnostic instrument for assessing accountability readiness before deployment and tracking drift over time. It is a proposed instrument that needs empirical validation, but the items are grounded in NIST, ISO, and IT governance research adapted specifically for the agentic AI context.</p><p><strong>Why this matters now</strong></p><p>If you are deploying agentic AI systems today, the question is not whether something unexpected will happen. It will. The question is whether you will be in a position to explain what happened, address the people affected, and prevent the same failure mode from appearing in your next deployment.</p><p>The hospital in the opening story did not fail because its AI system was poorly built. It failed because nobody had answered three governance questions before deployment: who monitors this system’s behavior, at what threshold does a human review become mandatory, and who specifically receives that review request?</p><p>These are not complex questions. They are just questions most organizations have not been asked to answer yet.</p><p><strong>The research</strong></p><p>This paper has been submitted for peer review at Global Business and Organizational Excellence and is available now as a preprint on SSRN. Search my name or the title directly: When Machines Decide: A Distributed Accountability Framework for Agentic AI in Organizations.</p><p>I am also presenting the broader three-paper research agenda on agentic AI accountability at the Americas Conference on Information Systems in Reno, Nevada this August.</p><p>If this is something you are working through in your own organization or research, reach out. This is a problem worth getting right before the incident, not after it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=42e17650419d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Mathematics of
Clinical Disorder]]></title>
            <link>https://medium.com/@albertaduseibrobbey/the-mathematics-of-clinical-disorder-7f5cbfba025c?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/7f5cbfba025c</guid>
            <category><![CDATA[ai-in-healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[process-improvement]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Mon, 06 Apr 2026 14:54:58 GMT</pubDate>
            <atom:updated>2026-04-06T14:54:58.184Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lz9vImPMeTLauYBq5u22Gw.png" /></figure><p><em>A Precise Mathematical Definition of Clinical Process Intelligence</em></p><p>Healthcare has never lacked frameworks for improvement. The problem has never been a shortage of tools. It has been the absence of a unified system that integrates them, one that is mathematically precise, theoretically grounded, empirically testable, and practically deployable.</p><p>Clinical Process Intelligence is that framework. This article presents its complete mathematical definition, from the foundational entropy formula through to the intelligence loop, alongside the intellectual lineage that grounds it in fifty years of established science.</p><h3>Why a mathematical definition changes everything</h3><p>Most concepts in healthcare management are defined loosely. Patient-centered care. High reliability. Integrated delivery. These are aspirations dressed as definitions. Valuable as directions of travel. Difficult to measure. Easy to claim without evidence.</p><p>A mathematical definition does three things a verbal definition cannot. It makes a framework falsifiable, meaning it can be tested and, if necessary, revised. It makes it comparable, meaning two organizations can be ranked objectively. And it makes it manageable, allowing leadership to set targets, track progress, and assign accountability for specific outcomes.</p><p>CPI does not propose principles. It proposes formulas, a measurement model, and a methodology grounded in those formulas from the first engagement to the last quarterly cycle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OIVy7ExvlFsn0zYqMk2AaQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wlriHrt5KJp3GlMjQ-fh4w.png" /></figure><h3>Process entropy H(G): the foundational formula</h3><p>The entire CPI framework rests on one concept borrowed from information theory: entropy. Applied to clinical processes, the formula is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*We_FuAUH1f7P5cgWJ9N4EA.png" /></figure><p>The formula takes each transition’s probability, multiplies it by the log of that probability, sums across all transitions, and negates the result. When every patient follows the same next step, H(G) approaches zero. The pathway is perfectly ordered. When patients scatter unpredictably, H(G) peaks. The pathway is maximally disordered.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aELIEegQcSdJFjMAowCloQ.png" /></figure><h3>Excess entropy ΔH(G): measuring the gap</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TASQNUrJup6FUDY7JRgPlA.png" /></figure><p>When ΔH(G) is greater than zero, the system is more disordered than intended. Avoidable variation is present. The gap is where harm and waste accumulate. When it equals zero, observed care is as ordered as intended. When it falls below zero, observed care is actually more consistent than the reference pathway, which can occur when clinical teams develop tighter working practices than the formal design specifies.</p><h3>Entropy dynamics: why improvement does not last</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Img_L2i2vZZ79oeGC84Wgw.png" /></figure><h3>The convergence condition: the central testable claim</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8oDiz4-Lse6YZjx-JfoLKw.png" /></figure><p>As n increases, as more intelligence cycles are completed, the observed process entropy H(Gn) approaches the reference entropy H(Gref). A healthcare organization that implements and sustains CPI will deliver care that progressively approaches its intended pathway.</p><p>This claim can be tested empirically. If organizations with higher CPI maturity scores show lower ΔH(G) and better clinical outcomes, the convergence condition is supported. If they do not, the framework requires revision. A framework that cannot be proven wrong cannot be trusted.</p><h3>Mutual information I(A;O): separating signal from noise</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aqvo1VQzyHaN-UYkZY5--Q.png" /></figure><p>When I(A;O) is high, the activity is a strong signal. Variation here has real consequences for patients. When I(A;O) is low, the activity is noise. The diagnostic function uses mutual information to rank deviations by outcome-relevance, ensuring attention is directed at variations that actually matter.</p><h3>The CPI Maturity Index: measuring organisational intelligence capacity</h3><p>The entropy formulas define the theoretical target. The CPI Maturity Index measures how much organisational capability exists to pursue it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yV-XyEHtIKvuWJsaipcy7g.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xWq8IBBy9uwKQJxaVgfH0w.png" /></figure><p>The five dimensions are not an arbitrary list. Each one is the organizational capability required to perform one entropy-reduction function. R enables H(G) to be detected as it rises. A enables ΔH(G) to be located and quantified. P enables H(Gt+1) to be estimated. It enables the implementation of entropy-minimizing interventions. L closes the loop, ensuring evaluative outputs update the diagnostic cycle. Remove any dimension and the loop breaks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_GVkhrz7a73rhPr9CVKeTA.png" /></figure><p>The overall score of 3.58 sits just below the Embedded threshold. The L dimension score of 2.8 is the weakest by a significant margin. The learning loop is broken. Without it, the organization generates intelligence but does not compound it. Addressing L is the priority, not improving already-strong dimensions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RJ5Lgh0mbf2iDjOGPUPTzg.png" /></figure><h3>The prescriptive function: selecting interventions by entropy reduction</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qjek7AHZxClQUAFB1Q2LVw.png" /></figure><p>If an intervention is implemented and entropy does not decrease as predicted, the prescriptive model is updated. The system learns from every recommendation it makes.</p><h3>The intelligence loop: how all formulas connect</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z6JNCZ89D3ZSqdZfFInmnA.png" /></figure><p>The loop is closed when E outputs update Lt+1 and Gref. An open loop, one where evaluation does not feed back into sensing, is not CPI. It is a reporting pipeline.</p><h3>The four axioms: structural requirements</h3><p>The framework is governed by four axioms. Violate any one and the system is formally degenerate. It may still produce data or insight, but it no longer qualifies as CPI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2DQlK9QQZdQS-3b68n4Ejg.png" /></figure><h3>Endnotes</h3><ol><li>The synthesis presented in this article as Clinical Process Intelligence is original to our team. The foundational science it draws from is attributed throughout these endnotes. No claim is made to the origination of Shannon entropy, process mining, or continuous improvement methodologies.</li><li>Shannon, C.E. (1948). <em>A Mathematical Theory of Communication.</em> Bell System Technical Journal, 27(3), 379–423. The foundational paper introducing information entropy, H = minus the sum of p(i) log p(i), originally developed to measure the uncertainty in message transmission over noisy channels. The application of this formula to clinical pathway transition distributions is an extension of Shannon’s original work.</li><li>van der Aalst, W.M.P. (2016). <em>Process Mining: Data Science in Action</em> (2nd ed.). Springer. van der Aalst’s work established process mining as a formal discipline bridging data science and process management. His Alpha algorithm and conformance checking methods form the methodological basis for deriving Gobs from EHR event log data. PM4PY, the Python process mining library referenced in CPI implementations, is maintained by the Fraunhofer Institute for Applied Information Technology.</li><li>Munoz-Gama, J., et al. (2022). <em>Process mining for healthcare: Characteristics and challenges.</em> Journal of Biomedical Informatics, 127, 103994. This manifesto paper, co-authored by 55 researchers across institutions including TU Eindhoven, KU Leuven, and Pontificia Universidad Catolica de Chile, formally established the PODS4H research agenda and outlined the key challenges in applying process mining to clinical data.</li><li>HealthProcessAI, Karolinska Institutet (2026). Published in <em>Frontiers in Artificial Intelligence.</em> This work introduced the first operational LLM-enhanced platform for clinical process intelligence, wrapping PM4PY and bupaR process mining tools with a large language model interpretation layer to make outputs interactive and clinician-accessible. It is the closest existing work to the CPI framework proposed here, though it does not include the maturity index, governance structure, or engagement methodology components.</li><li>Institute for Healthcare Improvement. The Plan-Do-Study-Act cycle, developed by W. Edwards Deming and adapted for healthcare by the IHI, provides the foundational continuous improvement logic that the CPI evaluative function formalises mathematically. The distinction CPI draws is between episodic PDSA cycles and the permanent CPI intelligence loop, which keeps entropy reduction ahead of entropy drift continuously. See also: Langley, G.J., et al. (2009). <em>The Improvement Guide: A Practical Approach to Enhancing Organisational Performance</em> (2nd ed.). Jossey-Bass.</li><li>Hair, J.F., Ringle, C.M., and Sarstedt, M. (2011). PLS-SEM: Indeed a silver bullet. <em>Journal of Marketing Theory and Practice,</em> 19(2), 139–152. The second-order latent construct formulation of CPI follows the reflective-reflective hierarchical component model described in this structural equation modelling literature. Factor loadings and measurement error terms in the doctoral formulation correspond directly to standard SEM notation.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7f5cbfba025c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Artificial Intelligence (AI) Is Just Math, So How Does It Understand the World?]]></title>
            <link>https://medium.com/@albertaduseibrobbey/artificial-intelligence-ai-is-just-math-so-how-does-it-understand-the-world-dfa1185dcbae?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/dfa1185dcbae</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-ai]]></category>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Sat, 28 Mar 2026 17:02:41 GMT</pubDate>
            <atom:updated>2026-03-28T17:02:41.482Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x_mAYb4xxmwEC0tkadt7LQ.png" /></figure><h3>The AI doesn’t see. It calculates.</h3><p>Here’s something that surprises most people: AI has no senses. It cannot see a photo, hear a song, or read a sentence the way you and I do. What it can do, extraordinarily well, is math. And math only works on numbers. So the entire job of a modern AI system is, first, to convert the world into numbers, and then to find patterns in those numbers.</p><p>Once you understand this, everything about AI becomes less mysterious. The “intelligence” isn’t magic; it’s pattern recognition at a scale no human brain could manage, performed through matrix multiplications, probability distributions, and gradient descent. Strip away the hype, and what you find underneath is a very fast, very large calculator that has seen enough examples to recognize what comes next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LE8ljScG7g_zJiUWA2rWAw.png" /></figure><h3>To an AI, every photo is a grid of numbers.</h3><p>When you look at a photo of a cat, you see fur, eyes, and attitude. When an AI looks at the same photo, it sees a massive grid of pixels , and each pixel is just three numbers: red, green, and blue intensity, each ranging from 0 to 255. A tiny 8×8 thumbnail is already 192 numbers. A 12-megapixel phone photo is 36 million.</p><p>The AI processes every single one of these number triplets, looking for mathematical patterns, edges, textures, curves, that it learned, during training, correspond to concepts like “whisker,” “ear,” or “cat.” Hover over the pixels below to see the numbers behind the color.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qGc3Dqj8XHIcUtF0NPG9-w.png" /></figure><h3>Words become tokens. Tokens become meaning.</h3><p>For large language models, the transformation works differently. Text first gets broken into chunks called tokens, roughly one word or word fragment each, and every token gets assigned a unique ID number. “Hello” might be token #1052. “Artificial” might be split into “Art” and “ificial”, two separate tokens, two separate IDs.</p><p>But the really profound part is what happens next. Each token ID gets mapped to an embedding: a list of hundreds or thousands of numbers that encodes the word’s meaning. Words used in similar contexts end up with mathematically similar embeddings. “King” and “Queen” live close together in this high-dimensional space. “Dog” and “Puppy” are neighbors. The model learns these relationships entirely from patterns in text, nobody defined what “similar” means. Hover the tokens below to watch the embedding values shift.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ABV2u9DT8gj5b8Bnosr-QQ.png" /></figure><h3>The same math, transforming every industry.</h3><p>The numbers-in, predictions-out principle isn’t limited to photos or language. It works in every domain that generates data, which, in the modern world, is every domain. The inputs change. The math is the same. Below is what that actually looks like when it lands in the real world.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3GmbtS4zHiWORw63i87mrQ.png" /></figure><p>A radiologist looking at a chest X-ray is doing something extraordinary , scanning a grayscale image for the faint density variations that signal early-stage pneumonia, a collapsed lung, or a suspicious mass. It takes years of training and a sharp eye. An AI model does something structurally identical: it processes the image as a grid of pixel intensity numbers and looks for the patterns , subtle clusters of higher-density values, irregular edges, asymmetric shadows , that it learned to associate with pathology after training on millions of labeled scans.</p><p>The results have been striking. AI models have demonstrated the ability to detect lung nodules smaller than 3mm in CT scans, nodules that radiologists can miss at a single pass. In diabetic retinopathy screening, AI trained on retinal photographs now matches the diagnostic accuracy of ophthalmologists in detecting the condition early enough to prevent blindness. In pathology, AI models analyzing digitized tissue slides can identify cancer subtypes from cellular patterns with accuracy that rivals specialist panels.</p><p>The reason this works is that diseases leave mathematical fingerprints. A tumor grows with irregular borders, irregular in a specific, measurable way. Pneumonia creates a haze in lung tissue, a density distribution that differs numerically from healthy lung. The AI doesn’t “know” what lungs look like. It knows what the numbers look like when something is wrong.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sG3VkNocPV-QTpY5DnsquQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xttkvbu6fYVeccJXHH5t9Q.png" /></figure><p>Every time you tap your card, a transaction record is born. It contains an amount, a merchant category code, GPS coordinates, a timestamp, a device fingerprint, and dozens of other data points. To a fraud detection AI, all of these immediately become numbers, a vector in a high-dimensional space. The AI’s job is to decide, in under 100 milliseconds, whether this vector looks like you or whether it looks like someone who stole your card.</p><p>This works because your spending behavior is mathematically consistent. You buy coffee at similar times, in similar places, for similar amounts. You don’t buy electronics in Singapore at 3 a.m., three minutes after a coffee in Austin. That sequence of transactions doesn’t just seem suspicious; it’s statistically anomalous in the space of your historical transaction vectors. The distance between the new transaction and your normal cluster of behavior, measured mathematically, is enormous.</p><p>Beyond individual fraud detection, AI is reshaping how credit risk is assessed. Traditional credit scoring uses a handful of variables. Modern AI models ingest thousands , including behavioral patterns like the time of day someone checks their balance, or how consistently they maintain a small buffer above zero. These patterns, invisible to a human analyst, are mathematically predictive of repayment behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tEN95Fa2a0dlf6KdiM0gnA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*236liyTJyK-GDaGlPkA6vA.png" /></figure><p>For fifty years, predicting how a protein folds from its amino acid sequence was considered one of biology’s grand unsolved challenges , the so-called protein folding problem. In 2020, DeepMind’s AlphaFold essentially solved it. The achievement wasn’t a biological breakthrough in the traditional sense. It was a mathematical one.</p><p>A protein is a chain of amino acids. Each amino acid can be represented as a number. A protein sequence is therefore a sequence of numbers. AlphaFold treats this sequence as input to a transformer-style neural network, the same architecture that underlies large language models, and predicts the 3D coordinates of every atom in the resulting folded structure. The model learned to do this by training on the roughly 170,000 protein structures that scientists had painstakingly determined over decades using X-ray crystallography.</p><p>The downstream implications are enormous. Drug discovery begins with understanding protein structures , a drug molecule works by binding to a specific protein in a specific way. AlphaFold has now predicted the structures of over 200 million proteins, essentially the entire known proteome of life. What used to take a research team years now takes seconds. Researchers are using these predicted structures to design new drug candidates for diseases that previously had no viable molecular targets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dj9o8Ckl7_8EG8o45Uc7nQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PNucAac_BgS4ILORi0NNpA.png" /></figure><p>A self-driving car exists in a state of constant numerical overwhelm. Each lidar pulse returns a point in 3D space, x, y, z coordinates, plus a reflection intensity value. A single rotation of the lidar array generates 100,000 to 2 million such points. Cameras generate pixel grids at 30–60 frames per second. Radar adds another layer of range and velocity measurements. All of this, more than 1 terabyte of sensor data per hour, pours into the vehicle’s AI systems, which must classify every object, track its motion, predict where it will be in 2 seconds, and plan a safe path, all in under 50 milliseconds.</p><p>The AI doesn’t “see” the pedestrian stepping off the curb. It detects a cluster of lidar points arranged in a shape consistent with a human torso, assigns it a velocity vector based on the change in that cluster’s position between frames, and computes a predicted trajectory. If that trajectory intersects with the vehicle’s planned path before the vehicle can stop, the AI outputs a braking command. Entirely mathematical. Entirely numbers.</p><p>One of the most difficult challenges is what the industry calls the “long tail” , the enormous variety of rare, unexpected scenarios that a human driver handles by understanding the world, and a self-driving car must handle by having seen enough numerical examples to generalize correctly. A mattress on the highway. A school crossing guard waving cars through a red light. These edge cases are where the math is hardest, and where the most active research is focused.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NunRdQ16wLry_BPSaxb4RQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_O_WGYQxVvJis0GfZK3uhg.png" /></figure><p>A healthy industrial motor has a signature , a precise pattern of vibration frequencies, a characteristic current draw, a normal operating temperature range. These aren’t metaphors. They are literal waveforms of numbers, streaming off sensors attached to the machine 24 hours a day. When a bearing begins to fail, it introduces a specific sub-harmonic vibration signature , a frequency that wasn’t there before, encoded as a numerical deviation in the sensor stream. The machine doesn’t look different. It doesn’t sound different to the human ear on the factory floor. But the numbers have changed.</p><p>AI models trained on historical failure data learn to recognize these numerical precursors, the specific pattern of deviations that precede a catastrophic failure by days or weeks. A factory running predictive maintenance no longer replaces parts on a fixed schedule, wasting components that still have life. And it no longer waits for machines to break, halting production lines at enormous cost. It replaces components exactly when the numbers say to, not a day too early, not a day too late.</p><p>The same principle extends to entire industrial processes. In semiconductor fabrication, AI monitors hundreds of process parameters simultaneously, chamber pressure, gas flow rates, plasma power, wafer temperature, and detects the subtle correlations that predict whether a batch will yield a high percentage of good chips or a defective one, in time to adjust the process mid-run.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bJ_vMaDajfUqSZyQaGXjMw.png" /></figure><p>Traditional weather forecasting runs on numerical physics simulations, equations that model how air masses move, how heat transfers, and how moisture condenses. These simulations are themselves math. But they are expensive, slow, and constrained by the computational cost of solving partial differential equations at every point in a global grid.</p><p>In 2023, Google DeepMind released GraphCast, an AI model that produces 10-day global weather forecasts at 0.25° resolution, that’s a grid point every 28 kilometers across the entire planet, in under 60 seconds. For comparison, the European Centre for Medium-Range Weather Forecasts (ECMWF), one of the best traditional forecast systems in the world, requires hours of computation on a supercomputer to produce similar forecasts. GraphCast outperformed ECMWF on 90% of forecast targets tested.</p><p>The AI didn’t learn physics equations. It learned, from 40 years of historical atmospheric data, the statistical patterns of how the atmosphere evolves, patterns that implicitly encode the physics, without needing to solve the equations explicitly. This is a profound shift: instead of simulating the world from first principles, AI learns the world’s behavior from data and reproduces it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z6DLjntrmdOY3hdjB9J60Q.png" /></figure><h3>Embeddings: where math captures meaning.</h3><p>Of all the ideas in modern AI, embeddings might be the most philosophically striking. The claim is audacious: that is, the relationship between concepts, the connotations of words, and the similarity of ideas can be <em>represented as geometry</em>.</p><p>In an embedding space, “Paris is to France as Berlin is to Germany” isn’t a language puzzle; it’s a vector equation. You can literally take the embedding for “Paris,” subtract “France,” add “Germany,” and arrive near the embedding for “Berlin.” The structure of human thought, encoded as coordinates in a high-dimensional space, turns out to follow rules as clean as arithmetic.</p><p>This is why AI can translate languages without a dictionary, understand analogies, and generate coherent long-form text. It isn’t following grammar rules; it’s navigating a learned geometry of meaning, making the statistically most likely next move through that space. And the same principle applies across every domain: in medical imaging, “diseased tissue” and “healthy tissue” become geometrically separable clusters. In fraud detection, “legitimate transaction” and “fraudulent transaction” occupy different regions of feature space. The math is universal. Only the input changes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U5SiO4BJ2GRIxYzNLsWZcA.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dfa1185dcbae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rich Data, Poor Insight: Rethinking Process Excellence in Healthcare]]></title>
            <link>https://medium.com/@albertaduseibrobbey/rich-data-poor-insight-rethinking-process-excellence-in-healthcare-4129d270366e?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/4129d270366e</guid>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[ai-in-healthcare]]></category>
            <category><![CDATA[lean-six-sigma]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Fri, 20 Mar 2026 16:21:47 GMT</pubDate>
            <atom:updated>2026-03-20T16:21:47.090Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*w26V5HMaxbVM-DQ0.png" /></figure><p>There’s a particular kind of frustration that lives in healthcare operations.</p><p>You’ve done the work. You ran the project. You mapped the process, identified the waste, built the fishbone diagram, and ran the A3. Six months later, the numbers improved, and then, quietly, they didn’t. The old patterns came back. The bottlenecks returned. And somewhere in a folder on a shared drive, the project report sits untouched.</p><p>This isn’t a failure of methodology. It’s a mismatch between the tools we use and the environment we’re operating in.</p><p>Healthcare is not a factory floor.</p><h3>The Promise That Lean Six Sigma Delivered, And Where It Stopped</h3><p>Let’s be clear: Lean Six Sigma changed healthcare for the better. It gave operations leaders a structured language for improvement. It brought rigor to processes that had been running on habit and institutional memory. It reduced waste, standardized workflows, and in many cases, saved lives.</p><p>But it was designed for a different era.</p><p>Lean thinking emerged from manufacturing environments where inputs are controlled, outputs are predictable, and improvement cycles can afford to be deliberate. DMAIC works beautifully when you have time to define, measure, analyze, improve, and control. When the process holds still long enough to study it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*-3YVmIoyKmjbSHri.png" /></figure><p>Clinical environments don’t hold still.</p><p>Patient volumes shift by the hour. Staff mix changes by the shift. A surge in the ED ripples into the lab, the pharmacy, the floors. A single delayed discharge creates a chain reaction that no process map fully captures , because by the time the map is drawn, the conditions have already changed.</p><p>This is the core tension: <strong>we are applying periodic tools to continuous problems.</strong></p><h3>Data Rich, Insight Poor</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*0pny3ln4LPFgZAiD.png" /></figure><p>Here’s what makes this especially hard to accept: hospitals have never had more data.</p><p>EHRs capture thousands of data points per patient encounter. Bed management systems track capacity in real time. Nurse call systems, lab interfaces, scheduling platforms, and supply chain tools are generating a continuous signal. The data exists.</p><p>But most of it sits in silos. And most of what gets analyzed gets analyzed after the fact, in monthly dashboards, quarterly reviews, and annual performance reports. By the time the insight surfaces, the moment to act on it has long passed.</p><p>We’ve built sophisticated data collection systems and connected them to retrospective reporting. That’s not intelligence. That’s documentation.</p><h3>The Question We Should Be Asking</h3><p>The problem isn’t that healthcare leaders don’t care about process excellence. They do, deeply. The problem is that the dominant model of improvement still treats process as something you fix and then move on from.</p><p>But what if process improvement wasn’t a project?</p><p>What if it were a capability embedded, continuous, and intelligent enough to adapt in real time as conditions change?</p><p>What if your operations didn’t just report on what happened last month, but detected what was about to happen, and responded before it did?</p><p>This is not a futuristic idea. The technology exists. The data exists. What’s been missing is a framework that brings them together in a way that’s clinically meaningful, operationally practical, and built for the complexity of real healthcare environments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*VmGZxqEuDDvDR7TG.png" /></figure><h3>Something Is Shifting</h3><p>Across the industry, a quiet evolution is underway. Health systems that are moving beyond traditional process improvement aren’t abandoning Lean or Six Sigma; they’re recognizing that those methodologies need a foundation. A layer of continuous intelligence that feeds into and elevates everything else.</p><p>They’re asking different questions. Not just <em>“what went wrong?”</em> but <em>“what is happening right now, and what will happen next if we don’t act?”</em></p><p>The shift isn’t from process improvement to AI. It’s from episodic improvement to embedded intelligence.</p><p>There’s a name for what’s emerging on the other side of that shift.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4129d270366e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Fixing Processes: Build Processes That Fix Themselves]]></title>
            <link>https://medium.com/@albertaduseibrobbey/stop-fixing-processes-build-processes-that-fix-themselves-b0d1445f206b?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0d1445f206b</guid>
            <category><![CDATA[continuous-intelligence]]></category>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Mon, 24 Nov 2025 17:36:22 GMT</pubDate>
            <atom:updated>2025-11-24T17:36:22.868Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1015/1*B12vcEQMPffAkZVcwZMP5g.png" /></figure><p>For decades, organizations treated improvement as an event. A problem emerged, leadership approved a charter, a team formed, data collection began, and analysis followed. Months later, the solution reached the control phase. The model was effective in a world where processes were stable and change occurred periodically. It no longer aligns with the pace of modern operations.</p><p>Today, most of the variation that damages performance happens in the space <em>between</em> projects. By the time a DMAIC cycle closes, the environment has already shifted. A new bottleneck forms. A metric drifts. A workaround appears on the frontline. Waste quietly regenerates while the next improvement sprint waits its turn in the portfolio. Organizations aren’t failing because Lean Six Sigma stopped working. They’re failing because the cadence of improvement no longer matches the cadence of change. That gap is where Continuous Intelligence alters everything.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1015/1*KF8oZElgLLjHkuebzfEj7g.png" /></figure><p><strong>The Problem With “Sprints”</strong></p><p>Improvement projects assume that progress behaves in neat cycles. In reality, the most damaging waste occurs when nothing is being analyzed. Demand spikes on a weekend. A supplier delay cascades into overtime. A policy update increases task time by 12 seconds and nobody notices until the next quarterly review. Humans detect only what they have time to measure. AI detects what happens whether anyone is watching or not.</p><p>Traditional DMAIC will always have value, but its one-time-per-problem rhythm is mismatched to environments where problems evolve continuously. The industry has spent 20 years accelerating projects. AI solves a different problem, eliminating the <em>waiting</em> between projects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*25JkgVJPdAbFYpDBTjwfmw.png" /></figure><p><strong>Kanban for Intelligence</strong></p><p>Continuous Intelligence behaves like a live Kanban board, but the cards represent process conditions, not tasks. Real-time data flows through the system:</p><ul><li>When variation enters the process, the board “pulls” investigation.</li><li>When a leading indicator predicts a bottleneck, the board “pulls” intervention.</li><li>When drift threatens quality, the board “pulls” escalation.</li></ul><p>Instead of pushing improvement initiatives based on opinion, AI pulls improvement cycles based on evidence. The role of practitioners changes: from monitoring the data to monitoring the <em>models</em> that monitor the data. The feedback loop becomes the product.</p><p>Improvement stops being a project pipeline and becomes a living workflow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1016/1*Jgc0Rook7lX4RF77ZpScqA.png" /></figure><p><strong>Empowerment Shift on the Frontline</strong></p><p>In project-based improvement, frontline workers serve mostly as data providers. They answer surveys, time studies, motion analyses, interviews. Their contribution is critical, but passive. Someone else diagnoses the problem and implements the change.</p><p>Perpetual improvement flips that relationship.</p><p>Frontline workers become the governors of the boundaries, not the gatherers of data. They define:</p><ul><li>What “normal” looks like.</li><li>What variation is acceptable.</li><li>When the AI is allowed to intervene.</li><li>When human judgment overrides automation.</li></ul><p>The frontline stops waiting to be rescued by a project team and instead becomes the authority that determines how aggressively the system should adapt. That’s empowerment that actually changes culture, not slogans.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1012/1*ldKbhpcEn3toDHiZTeWHow.png" /></figure><p><strong>The Perpetual Audit</strong></p><p>AI never sleeps, never forgets, and never waits for quarterly KPI reviews. It performs what Lean practitioners have always wanted but could never execute manually: 24/7 Control phase.</p><p>It asks:</p><ul><li>Is this pattern normal?</li><li>Is this deviation temporary or structural?</li><li>Does this deviation require a correction now?</li><li>Who needs to be alerted, and how urgently?</li></ul><p>The Control phase no longer fades after the project closes. It becomes a standing operating condition, constantly comparing reality to standard work, triggering either automated correction or human intervention.</p><p>The result isn’t more control.<br> It’s <em>sustained</em> control.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1009/1*v9xeIAUDBVylS3Zqo--MoA.png" /></figure><p><strong>What This Transition Means for Leadership</strong></p><p>Moving from projects to perpetual flow isn’t about technology. It’s about identity.</p><ul><li>PMOs become flow orchestration hubs, not backlog schedulers.</li><li>Black Belts become architects of guardrails, not guardians of toolkits.</li><li>Operators become performance stewards, not task executors.</li><li>Executives stop celebrating big wins and start celebrating <em>consistent ones.</em></li></ul><p>Success shifts from “How many projects did we complete?” to “How long did the system stay in the optimal zone?”</p><p>Organizations that make this shift early will outperform those still running cyclical improvement programs simply because they improve in the spaces where others are idle.</p><p>Project-based improvement built the last era of operational excellence.<br> Perpetual flow will define the next.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0d1445f206b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Shift from Continuous Improvement to Continuous Intelligence]]></title>
            <link>https://medium.com/@albertaduseibrobbey/for-decades-lean-six-sigma-has-been-the-gold-standard-for-process-excellence-098579c884c2?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/098579c884c2</guid>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Sat, 15 Nov 2025 13:46:03 GMT</pubDate>
            <atom:updated>2025-11-15T13:47:39.865Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1012/1*EQ5Fev83-NLuN27HQ7pHCw.png" /></figure><p>For decades, Lean Six Sigma has been the gold standard for process excellence. But what happens when machines can identify waste, predict variation, and improve processes on their own?</p><h3>The Legacy of Lean Six Sigma</h3><p>Lean Six Sigma has powered some of the most impressive operational transformations in modern business history. From manufacturing floors to healthcare systems, its structured approach to eliminating waste and reducing variation has saved billions and fundamentally changed how organizations think about quality and efficiency.</p><p>The methodology gave us powerful tools: DMAIC frameworks, statistical process control, value stream mapping, root cause analysis. It taught generations of professionals to see processes through a critical lens, to measure what matters, and to drive change through data and discipline. These principles remain as relevant today as they were when they revolutionized manufacturing in the 1980s and 90s.</p><p>But here’s the reality we must confront: traditional Lean Six Sigma was designed for a different era. An era where data was scarce and expensive to collect. Where analysis happened in monthly reviews, not milliseconds. Where human experts needed weeks to identify patterns that algorithms can now surface in seconds.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1016/1*tG1OAYQEhv3b306PWwN7kA.png" /></figure><h3>The Limits of Manual Excellence</h3><p>Traditional continuous improvement operates on a fundamentally reactive cycle. A problem emerges, we form a project team, we collect data, we analyze root causes, we implement solutions, we validate results. This cycle typically takes months, sometimes years.</p><p>Meanwhile, modern businesses generate data continuously. Customer behaviors shift daily. Supply chains fluctuate hourly. Market conditions evolve in real-time. The gap between when problems occur and when we respond to them has become a critical vulnerability.</p><p>Consider what happens in a traditional Six Sigma project: a Black Belt spends weeks gathering historical data, analyzing it in Minitab, building fishbone diagrams, and presenting findings to stakeholders. By the time solutions are implemented, the process may have already changed. The insights, while methodologically sound, are historical snapshots of dynamic systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1016/1*hgNV8OZtVsviVFcZJnIZ0g.png" /></figure><h3>Enter the Age of Continuous Intelligence</h3><p>We’re witnessing the emergence of a new paradigm: Continuous Intelligence. This isn’t about replacing Lean Six Sigma — it’s about augmenting and accelerating it through AI, automation, and real-time analytics.</p><p>Continuous Intelligence means processes that self-monitor, predict their own failures, and adapt without waiting for human intervention. It means moving from periodic audits to perpetual optimization. From reactive problem-solving to proactive prevention.</p><p>Here’s what this looks like in practice:</p><p>Automated Waste Detection: AI systems continuously analyze operations, identifying the eight wastes of Lean faster and more comprehensively than any kaizen team. They spot patterns across vast datasets that would be invisible to human observers — correlations between seemingly unrelated variables, micro-inefficiencies that compound into major losses, emerging trends before they become problems.</p><p>Predictive Quality Control: Instead of detecting defects after they occur, machine learning models predict quality issues before they happen. They learn to recognize the subtle signs of impending failures, adjusting process parameters in real-time to maintain optimal performance. Statistical process control charts are updated continuously, with AI automatically identifying special cause variation.</p><p>Dynamic Process Optimization: Algorithms don’t just maintain processes at their current capability, they continuously experiment, learn, and improve. They run thousands of virtual experiments, optimizing multiple variables simultaneously in ways that would be impossible through traditional DOE methods.</p><p>Intelligent Resource Allocation: AI predicts demand fluctuations, optimizes scheduling, and dynamically allocates resources across the value stream. What once required complex planning cycles now happens automatically, responding to changing conditions.</p><h3>The Human Element Evolves</h3><p>This shift doesn’t diminish the role of process excellence professionals; it elevates it. Instead of spending time on data collection and manual analysis, practitioners focus on strategic questions: Which processes should we optimize? What outcomes truly matter? How do we balance competing objectives? What ethical guardrails should guide our automation?</p><p>The skills that matter are changing. We still need to understand variation and waste, but now we also need to interpret algorithmic outputs, design intelligent systems, and manage the human-AI collaboration. The future belongs to professionals who combine process expertise with data science literacy and strategic thinking.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1013/1*W5GFjUCnR5uZ_cmUri28CA.png" /></figure><h3>Building the Bridge</h3><p>The transition from Continuous Improvement to Continuous Intelligence isn’t instantaneous. Organizations need to build the infrastructure: connected sensors, data pipelines, analytics platforms, AI capabilities. They need to develop new competencies and change management approaches. They need to maintain the discipline and rigor of Lean Six Sigma while embracing new technologies.</p><p>The most successful organizations are taking a hybrid approach. They’re using AI to identify opportunities and generate insights, then applying Lean Six Sigma discipline to validate findings and manage implementation. They’re automating routine optimization while reserving human judgment for complex, contextual decisions.</p><h3>The Next Frontier</h3><p>We’re at an inflection point in process excellence. The organizations that thrive in the next decade won’t be those that cling to traditional methods or blindly adopt new technologies. There’ll be those who thoughtfully integrate the wisdom of Lean Six Sigma with the power of Continuous Intelligence.</p><p>The goal remains unchanged: to eliminate waste, reduce variation, and deliver value. But the velocity, scale, and sophistication with which we pursue that goal are being transformed.</p><p>The question isn’t whether to make this shift, it’s how quickly and how well. Because while you’re planning your next improvement project, your competitors might be building systems that improve themselves.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=098579c884c2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Does My Phone Know What I’m Thinking? The Truth Behind Creepy Ads]]></title>
            <link>https://medium.com/@albertaduseibrobbey/why-does-my-phone-know-what-im-thinking-the-truth-behind-creepy-ads-6390c863e48d?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/6390c863e48d</guid>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Tue, 02 Sep 2025 16:52:12 GMT</pubDate>
            <atom:updated>2025-09-02T16:52:12.982Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*ElHBNq-pxQAyKE4YS4wWkg.png" /></figure><p>Most people have had this experience. You mention something casually to a friend, a new vacuum cleaner, a vacation idea, a pair of hiking boots, and later that day, your social media feed seems to respond with ads for that very thing. Sometimes it feels even stranger: you only <em>think</em> about a product, and the next morning you see it promoted online. The coincidence feels too sharp to ignore, leading many to wonder whether our phones are secretly listening to us.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*as59GIwIDKsUHHQl" /></figure><p>The unsettling truth is that in most cases, microphones are not the culprit. The reality is far more complex and, in many ways, more concerning. Companies rarely need audio recordings to create the impression of mind-reading. Instead, they rely on an ecosystem of sophisticated tracking, predictive algorithms, location data, and the quirks of human psychology.</p><p><strong>The Power of Tracking and Prediction</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*_fj-7y7h-HiURIsmm9CyfQ.png" /></figure><p>Smartphones and apps generate enormous amounts of behavioral data. Every search you perform, every post you like, every purchase you make, and even the amount of time you pause on a video creates signals that are collected and analyzed. This data is tied together through methods such as device IDs, cookies, and invisible tracking pixels embedded in websites and emails. Together, these signals form a digital fingerprint, a constantly updated map of who you are, what you do, and what you are likely to want.</p><p>Advertisers use this fingerprint to target ads with uncanny precision. When your online behavior even slightly aligns with a marketable pattern, ads appear almost instantly. The system doesn’t need to listen to your voice to know what you’re interested in. The data trail you leave behind already tells the story.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3e0dUEcslVpZdogY" /></figure><p><strong>Prediction Disguised as Telepathy</strong></p><p>Algorithms have become so advanced that they don’t just reflect your current interests; they predict your future ones. By analyzing vast patterns across billions of users, machine learning models can infer what you may want before you consciously decide. If people who share your demographics and online habits recently purchased hiking gear, the algorithm assumes you may be next in line.</p><p>This predictive element explains why ads sometimes arrive before you take direct action. It feels as though your phone overheard your conversation, but in reality, the system recognized the signals leading toward that purchase. In some cases, it knew what you were going to want before you even admitted it to yourself.</p><p><strong>How Social Circles and Location Play a Role</strong></p><p>Targeting doesn’t stop at the individual level. Ad networks map relationships and use proximity to make connections. If a close friend has been searching for baby strollers and you spend time together in person, algorithms recognize the connection and often show ads to both of you. This can happen through location services, Bluetooth, and Wi-Fi signals, which confirm that you share physical space.</p><p>The effect is subtle but powerful. You may see an ad not because of your own behavior, but because the people around you expressed interest. When this overlaps with a real conversation you’ve had, it can feel almost supernatural.</p><p><strong>The Psychology That Fuels the Illusion</strong></p><p>Even when algorithms alone can’t explain the eerie feeling, our minds help fill the gap. Cognitive biases make the experience more dramatic than it is. The frequency illusion, also known as the Baader–Meinhof effect, makes us notice something more once it’s on our mind. Ads for a product may have been appearing for days, but you only notice them once you start thinking about that product in real life.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*O_44JePU6u1ivBh8OWkJJA.png" /></figure><p>Confirmation bias adds another layer. We tend to ignore the countless irrelevant ads we scroll past each day, while remembering the one ad that matched our private thought or conversation. A handful of coincidences stand out, reinforcing the impression that our phones are eavesdropping.</p><p><strong>The Question of the Microphone</strong></p><p>Companies such as Meta and Google have consistently denied using microphones for ad targeting, and so far, independent researchers have not found evidence of constant audio surveillance. There are good reasons for this. Streaming and transcribing audio from billions of users would be technically inefficient, easily detected by security experts, and legally catastrophic.</p><p>That said, our devices do listen locally for wake words such as “Hey Siri” or “Okay Google.” After activation, small snippets of audio are processed in the cloud. Companies claim this information is not used for advertising, but the fact that it exists leaves enough gray space to fuel public suspicion.</p><p><strong>Why It Feels So Invasive</strong></p><p>The reality is that advertisers don’t need microphones because the system already works. By combining relentless data harvesting, predictive modeling, and network effects, ads can appear with such precision that they feel psychic. The unsettling part is not hidden surveillance but the way ordinary data is transformed into powerful predictions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*KbBGtXPf8sGn5kRqP_kRYw.png" /></figure><p>This raises deeper concerns. If platforms can predict what we want before we know it ourselves, where does personalization end and manipulation begin? Ads no longer just reflect demand; they help create it, steering our choices in ways that blur the line between autonomy and influence.</p><p><strong>What You Can Do</strong></p><p>While you can’t erase your digital shadow, there are ways to reduce your exposure. Reviewing app permissions and revoking access to the microphone and location data from unnecessary apps is a start. Both iPhone and Android allow you to limit ad tracking or reset advertising IDs. Using privacy-focused browsers such as Brave or Firefox, along with extensions like uBlock Origin or Privacy Badger, can help block some forms of tracking. Most importantly, being selective with the apps you install reduces the number of companies collecting your data in the first place.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*-QHg411rbrv_HjPmDjw97g.png" /></figure><p>Your phone is not listening in the literal sense, but the system behind it is sophisticated enough to make it feel that way. What we experience as “mind-reading ads” is the result of predictive algorithms trained on massive data trails, combined with human biases that amplify coincidences into something uncanny. The danger lies not in secret microphones, but in a digital ecosystem that already knows us so well it can anticipate our needs, and perhaps even shape them, before we speak them aloud.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6390c863e48d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Inside the Mind of a Doctoral Researcher Navigating the Realities of the Business World]]></title>
            <link>https://medium.com/@albertaduseibrobbey/inside-the-mind-of-a-doctoral-researcher-navigating-the-realities-of-the-business-world-4f07a929f25f?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f07a929f25f</guid>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Mon, 05 May 2025 15:12:00 GMT</pubDate>
            <atom:updated>2025-05-05T15:12:00.546Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LxyF70R0UxZzmnYR" /></figure><p>If you had told me a few years ago that I’d be deep in academic journals while also running real-world business projects, I would’ve raised an eyebrow. Not that I didn’t value research, but because, like many professionals, I didn’t perceive its practical application in my daily life. I was used to solving problems, hitting targets and delivering results. What more could I want?</p><p>Then I started my doctoral journey and everything changed..</p><p>For a time, I kept these two worlds apart. Work was nimble, responsive and people-driven. The research was slow, methodical, and riddled with frameworks I couldn’t even pronounce correctly at first glance. Now, though, over time, they seemed to be talking to each other.. What seemed like two different languages became one conversation: practice tells me what is happening. Theory begins to explain why.</p><p><strong>When Frameworks Start Talking Back</strong></p><p>Instinct is something we all tend to rely on, especially in the world of real-world strategy. You see a pattern, follow your gut, and move. That works — until it doesn’t.</p><p>I remember when we adopted dashboards and still faced systematic delays. We had meetings full of conflicting assumptions and a complete lack of alignment. That’s when I applied the Fishbone Diagram, a concept I encountered through Lean Six Sigma. This process of breaking the problem down into categories such as methods, people, environment, and systems turned a general feeling of frustration into a structured map. As it turned out, what we took to be a simple resource issue actually involved outdated processes, poor system integration, and a single process checkpoint that was completely out of alignment with the rest of its business area.</p><p>That was a shift. Theory didn’t just add structure, it unlocked better decisions.</p><p><strong>How Research Changes the Way You Think</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dk5VNlk3rcIkahqr" /></figure><p>Doctoral research teaches you to question assumptions. It challenges you to back up observations with patterns, logic, and evidence. That attitude didn’t slow down my strategic work , it deepened it.</p><p>I enter high-stakes conversations now with fewer assumptions and more impertinent questions. I don’t just ask, “What’s the issue?” I ask, “What are the underlying variables we haven’t named yet?” The possibility of zooming out before zooming in has helped make my thinking clearer and more relevant in complex systems.</p><p>The further I dive into research, the clearer it becomes to me how much modern institutions are missing when they ignore theory. Not because theory has everything figured out, but precisely because theory re-frames the question.</p><p>Leadership models, decision frameworks, and operations design are more than academic artifacts. Instead, they keep you from being reactive and add a layer of intentionality to the strategy. When you’re navigating uncertainty, theory provides you with mental infrastructure. It allows you to distinguish signal from noise. It teaches you to dig into why a solution worked, not just that it did.</p><p><strong>So why does real-world strategy need theory?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*swyyd66KshbNhnuY" /></figure><p>Because experience shows you <em>what</em> is happening. But if you want to lead lasting change, you need to understand <em>why</em>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f07a929f25f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Won’t Replace You, But a Leader Who Understands It Will]]></title>
            <link>https://medium.com/@albertaduseibrobbey/ai-wont-replace-you-but-a-leader-who-understands-it-will-ec5bd1bfdadf?source=rss-4cf50d8d7232------2</link>
            <guid isPermaLink="false">https://medium.com/p/ec5bd1bfdadf</guid>
            <dc:creator><![CDATA[Albert Adusei Brobbey]]></dc:creator>
            <pubDate>Tue, 22 Apr 2025 19:12:10 GMT</pubDate>
            <atom:updated>2025-04-22T19:12:10.743Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FeFGzhf35gNPyVz6" /></figure><p>The fear that artificial intelligence will eliminate jobs isn’t new. It echoes every major technological shift, from the steam engine to the internet. But that fear is now taking a more nuanced form, not <em>just</em> about being replaced by machines, but being outpaced by people who understand how to use them better.</p><p>The real risk isn’t AI itself — it’s irrelevance.</p><p><strong>The Tools Are Evolving. Are You?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1grVBgnzTXAMIj85" /></figure><p>AI isn’t a threat to humans. It’s a threat to traditional thinking.</p><p>Leaders who are paying attention are not asking, “Will AI take our jobs?” They’re asking, “How can we integrate AI into our decision-making to move faster, see clearer, and operate leaner?” The difference in mindset is already separating future-ready organizations from the ones scrambling to keep up.</p><p>Take finance, for instance. In forward-thinking companies, CFOs aren’t just reviewing quarterly results and working with predictive analytics to make real-time capital allocation decisions. In marketing, AI is enabling hyper-personalization at scale, creating micro-campaigns based on behavioral data without human bottlenecks. And in operations, supply chains are being dynamically optimized through AI-driven forecasting.</p><p>The people leading these shifts aren’t data scientists. They’re leaders who understand data, question assumptions, and recognize that competitive advantage now depends on how fast you can synthesize insight and act on it.</p><p><strong>AI Is Not Magic. It’s Leverage.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/0*l8mFiXcRjNs3dNOr" /></figure><p>Let’s make one thing clear: AI is not a silver bullet. It can’t replace judgment, ethics, or human empathy. But it does compress timelines. It processes complexity at scale. It frees up capacity for higher-order thinking.</p><p>The leaders who get this aren’t trying to replace human talent — they’re augmenting it. They’re creating teams that pair strategic acumen with technical fluency. They’re not asking their teams to code models from scratch , they’re asking better questions of the data, building workflows that loop human insight with algorithmic efficiency.</p><p>In short, they’re leading differently.</p><p><strong>The Gap Is Not Technical. It’s Strategic.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CyCyrg7EZUhwbDwv" /></figure><p>Ironically, the biggest obstacle to AI adoption isn’t the technology itself — it’s leadership hesitation. In many companies, AI is still trapped in the basement of IT or siloed within data science teams. Meanwhile, the executive suite remains passive, skeptical, or simply unaware of what’s possible.</p><p>That gap is costing organizations millions in missed opportunities.</p><p>According to a 2023 study by PwC, <strong>AI-driven organizations are 2.5 times more likely to report significant performance gains</strong>. But most executives still struggle to translate those insights into strategy. They ask for dashboards, not decisions. They fund pilots but lack a change model for adoption. They focus on AI tools instead of AI outcomes.</p><p>Leadership that fails to evolve with AI doesn’t just slow innovation — it makes companies vulnerable to disruption from more agile, more informed competitors.</p><p><strong>What Makes a Leader AI-Literate?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ewLAuxElgdIi6aje" /></figure><p>You don’t need to be a programmer. But you do need to understand how AI creates leverage:</p><ul><li><strong>You need to know where AI can improve decision quality, not just automate tasks.</strong></li><li><strong>You need to distinguish between a data problem and a people problem.</strong></li><li><strong>You need to build cross-functional teams that include business, data, and technical voices at the same table.</strong></li><li><strong>You need to champion experimentation and model smart risk-taking.</strong></li></ul><p>And most of all, you need to be willing to unlearn parts of your leadership playbook.</p><p>The leaders thriving in this new era aren’t the loudest or the most tech-savvy. They’re the ones who recognize that technology, used strategically, amplifies judgment, improves clarity, and accelerates impact.</p><p><strong>Final Thought</strong></p><p>AI won’t replace your job. But if you’re a leader who doesn’t evolve, someone else, someone who understands how to think with machines, will replace <em>you</em>.</p><p>The future won’t belong to those who resist automation. It will belong to those who lead with it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ec5bd1bfdadf" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>