AI & Agentic Systems

AI disappoints us daily, yet we keep paying for it. Expectation-confirmation theory explains the trap.

Organizations keep publishing AI governance policies they cannot enforce. Decoupling explains the gap. Isomorphism explains the copying. Neither explains why we...

Why 88% of organizations use AI while only 7% scale it is not a lag. It is a paradox built into exploration and exploitation.

AI handles the ostensive routine beautifully. But organizational value lives in the performative, where people improvise around what the process diagram never a...

Shadow IT was about unauthorized tools. Shadow AI is about unauthorized autonomous agents making decisions on your behalf. Delegation theory explains why that d...

Spence showed that markets only work when signals are costly enough to separate quality from noise. AI vendor signals are cheap and unverifiable. That is the wh...

Eleven blog posts on agentic AI market sizing, and not one about what happens to decision authority and middle management when algorithms take over routines.

When OpenAI restricted GPT-4's function calling, it made a governance decision disguised as a product update. Boundary resources are where platform power actual...

The productivity paradox keeps showing up because we keep measuring the wrong thing. Use is a proxy. Delegation is the mechanism. With agentic AI, the paradox r...

Orlikowski and Iacono asked IS to theorize the IT artifact. Agentic AI makes that question unavoidable. The artifact is now an actor.

Most delegation theory assumes humans delegate to AI. Stelmaszak et al. (2025) reverse it. When algorithms hand control back, accountability, trust, and organiz...

Most organizations adopted AI. Almost none scaled it. Absorptive capacity is the theory that explains the gap between buying a tool and learning from it.

Gartner says only ~130 of thousands of AI vendors are genuinely agentic. The rest are rebadging. Institutional theory explains exactly why this keeps happening.

McKinsey says 88% of organizations use AI. Only 7% have scaled it. That gap is enormous, and IS research has a theory for exactly why it exists.

Agency theory explains IT outsourcing and governance through principals, agents, and monitoring. AI rewrites the problem in a way the original framework did not...

The platform that controls the agent communication protocol will control the Internet of Agents, just as Google owned search and Apple owned mobile.

Gartner estimates only about 130 of thousands of self-described agentic AI vendors are real. That gap has a name, and it has consequences.

Gartner says 40% of enterprise apps will have AI agents by end of 2026. The same firm says 40%+ of those projects will be canceled by 2027. As an IS researcher,...

Rogers showed that innovation adoption follows an S-curve. Enterprise AI is not stalling. It is exactly where diffusion theory predicts.

Rogers, Granovetter, and Burt explain why the department with the AI champion adopts faster while the department across the building with the same budget does n...

TOE from Tornatzky and Fleischer explains why most AI investments fail while a few transform organizations: the three contexts must align.

Gartner projects the AI agent software market at $376.3B by 2027 and only about 130 real vendors behind it. The gap between the market number and the reality is...

The 80% autonomous resolution prediction is specific enough to be testable. As an IS researcher, the interesting question is not whether the technology can do i...

The AI alignment problem is principal-agent theory with a new degree of freedom: the agent can act faster than human oversight can keep up.

Organizations use AI to optimize existing processes while calling it transformation. March (1991) would call this pure exploitation, and the exploration gap is ...

AI chatbots feel like a face to face conversation but function like a search engine with a confidence problem. Media richness theory explains why that mismatch ...

Thirty percent acceptance is adoption, not effective use. Burton-Jones and Grange defined effective use as faithful domain representation, not output.

Multiagent AI systems are not just tools being used. They act, and their outputs become part of the organizational structure that shapes the next cycle.

Gartner predicts agentic AI will resolve 80% of common customer service issues by 2029. Getting there requires solving problems that have nothing to do with the...

Training AI models consumes staggering amounts of energy. Trist and Bamforth showed in 1951 why optimizing only the technical subsystem creates a collapse.

Meyer and Rowan predicted that organizations adopt structures for legitimacy, not performance. AI ethics boards are the most direct example I have seen.

McKinsey puts the midpoint of workforce automation at 2045. 88% of organizations use AI but only 7% have scaled it. The hype and the timeline are not the same t...

Cohen, March and Olsen showed that solutions search for problems, not the other way around. Most organizations adopted AI exactly that way between 2023 and 2025...

Hallucination is framed as a retrieval bug to be engineered away. Lee and See would call it a trust calibration failure, and that framing changes the solution e...

AI outperforms radiologists at detection but hospitals do not adopt it. Identity theory explains why. The tools that succeed will verify professional identity, ...

Kattnig et al. (2024) showed a model can satisfy every fairness metric and still feel deeply unfair. Amazon's AI hiring tool proved it with real damage.

Bostrom and Heinen said in 1977 that MIS failures are sociotechnical. Half a century later, AI implementation fails the same way.

Economists study macro displacement. IS researchers can study what actually happens inside organizations when AI changes specific jobs, roles, and power structu...

Gartner's May 2026 finding that AI layoffs free up budget without delivering returns is the most important warning CIOs are not taking seriously.

AI models predict patterns in observed data but never access the generative mechanisms that produce them. Critical realism tells us why that ceiling is not a te...

Gartner says two-person AI-augmented teams will deliver what twenty-person teams do today. Baird and Maruping predicted this years ago. The real question is not...

Network effects theory explains why AI markets concentrate around a few providers and why the data flywheel makes it nearly impossible for new entrants to catch...

Social cognitive theory predicts that AI adoption clusters not by access but by observation: people adopt when they see peers succeed with AI, and the gap betwe...

Star and Griesemer defined boundary objects in 1989. AI output fits the definition perfectly. And that is why every governance policy that treats it as a single...

Alter's work system framework already has a slot for AI. It is not the technology element. It is the participant element.

Almost every large organization has run an AI pilot. Very few of those pilots make it to production. That gap is not a data science problem.

Almost every AI policy is a find-and-replaced IT acceptable-use policy. DiMaggio and Powell explain why that pattern is not lazy. It is institutional isomorphis...

Gartner predicts over 40% of agentic AI projects canceled by 2027. The pattern looks a lot like ERP in the 1990s, and the failure mode is the same.

The EU has comprehensive legislation. The US has executive orders. China has sector-specific rules. The UK has no AI-specific law. Organizations operating globa...

IS research methods assume technology is stable and humans adapt. AI inverts that assumption, and our methods have not caught up.

AI safety warnings raise threat appraisal sky high. Without coping appraisal, they produce fear control, not safety.

When 80% of unauthorized AI use is internal policy violations, AI security platforms are governance products, not security products.

Access to AI tools is not the barrier. The belief that you can use them effectively is what separates who gains and who falls behind.

Every major cloud vendor now sells GPU clusters. The hardware is commodity. The capability to use it is not. Here is what Carr and Barney teach us about AI moat...

Goodhue and Thompson said performance follows fit, not features. AI is a family of tools, and the wrong one for the task will fail the same way every time.

The New York Times lawsuit against OpenAI is the Knowledge-Based View of the firm becoming visible in court and in every training data licensing negotiation.

Transaction cost economics predicts that when AI cuts coordination costs, organizations should restructure. But they are not. What gives?

AI trust repair is not about rebuilding a relationship. It is about recalibration. The IS trust literature has been explaining this for two decades.

Pfeffer and Salancik said dependence rises when resources are critical and concentrated. AI model providers are the most concentrated critical resource in IS hi...

When an algorithm screens job applicants or recommends bail conditions, it is not neutral infrastructure. It is policy, running at scale, with no judgment calls...

AI agents that plan, act, and iterate without human input are moving into enterprise deployment. The accountability questions this raises do not have clean answ...

Most people think digital privacy is about data breaches. Leidner and Tona's CARE framework says it is really about whether your dignity can survive digitizatio...

Gartner's April 2026 CEO survey puts 80% of chief executives on record expecting AI to force operational overhauls. McKinsey's data shows only 7% of organizatio...

Rule-based bots were honest about what they couldn't do. LLM-based bots can sound confident while being wrong. That's a more dangerous failure mode.

Gartner's April 2026 forecast shows AI spending nearly doubling to $2.5T in 2026. CIOs are now expected to demonstrate returns on that investment. As an IS rese...

CDSS tools fire alerts, suggest diagnoses, flag drug interactions. But when the algorithm is wrong and the doctor follows it anyway, who is responsible for the ...

Gartner projects public cloud reaching $1.48 trillion by 2029, doubling from 2025. Getting there requires physical infrastructure, energy capacity, and organiza...

I rolled my eyes when a study partner said critical realism was important. Three reads later I apologized, at least internally.

Studying institutional theory for comps changed how I read the news. Every AI-first press release is mimetic isomorphism. I cannot unsee it.

Hardware-level encryption during processing changes the risk side of the privacy calculus. IS privacy models were built for a world that no longer exists.

Gartner forecasts $788 billion in data center systems spending in 2026. That number is infrastructure for AI, and infrastructure has always been more political ...

DORA 2024 found 90% of organizations have adopted at least one platform engineering practice. The gap between adoption and value is a sociotechnical problem, no...

Compliance asks 'are we allowed to do this?' Ethics asks 'should we do this?' In most organizations, the first question is answered and the second is never aske...

Digital provenance (BOMs, attestation databases, watermarking) gives users verifiable information about system origin and integrity. It is the first trust calib...

The DORA 2024 report found AI adoption improves deployment throughput but damages delivery stability. As an IS researcher I think that pairing is the most impor...

Gartner says 60 percent of enterprise GenAI models will be domain-specific by 2028. But a domain-specific model cannot compensate for an organization that lacks...

When AI runs on a factory sensor or hospital monitor instead of the cloud, the technical subsystem becomes physically distributed and harder to monitor. STS the...

88% of organizations say they use AI. Only 7% have actually scaled it. That gap is not a rounding error.

The EU AI Act hits all three of Scott's institutional pillars at once. Here is why that makes it the most consequential IS policy since GDPR.

The EU AI Act is now in force. If you build or deploy AI that touches EU users, you are in a regulated space whether or not you planned for it.

The EU AI Act's high-risk requirements hit in August 2026. Organizations with EU operations using AI in hiring, credit, or healthcare are not ready.

TAM was built for email and word processors. Now it explains AI agents, blockchain, and VR. The IS field knows how to replace theories, so why does it keep choo...

Cohen, March and Olsen showed that in real organizations, solutions arrive before problems. The AI strategy wave is the clearest recent example.

AI spending jumped from $1.5 trillion in 2025 to $2.5 trillion in 2026. Before celebrating, it is worth asking what the money is actually going toward.

McKinsey estimates $2.6T-$4.4T in annual GenAI value. McKinsey also reports only 7% of organizations have fully scaled AI. That gap has a name in IS research.

Gartner's April 2026 forecast puts GenAI model spending growth at 80.8% for 2026. That number is real. What it is buying is more complicated.

AI investments are growing 35%+ year over year. Boards want ROI. The measurement frameworks for GenAI ROI barely exist yet.

Enterprise AI governance involves at least four distinct problems. Most organizations are still treating them as one.

Global IT spending reaches $6.31 trillion in 2026, a 13.5% jump from 2025. AI is the engine. Data centers, software, and security are all growing fast.

Pfeffer and Salancik predicted this. When one supplier controls access to a critical resource, power concentrates. NVIDIA at 80% GPU market share changes everyt...

Gartner predicts guardian agents will claim 10-15% of the agentic AI market by 2030. That prediction reveals a governance problem nobody has solved.

Most technology resistance is not about the technology. When a tool threatens who someone believes they are, rejection is identity-consistent, not irrational.

Gartner found 54% of I&O leaders are adopting AI to cut costs. But AI spending itself is heading to $2.5 trillion in 2026. The math needs to include both sides.

Most organizations adopting AI are not making rational decisions. They are responding to coercive, mimetic, and normative pressure that makes copying feel like ...

DORA 2024 found a direct correlation between internal developer platform quality and the ability to extract value from AI tools. That finding reframes platform ...

IS sits between technology and management, and that positioning opens more paths than most people realize, including some that didn't exist five years ago.

DeLone and McLean's IS success model applies to AI systems when you expand 'user' to include both humans and the downstream systems that consume AI output.

Digital infrastructure consumes significant electricity. IS researchers studying digital transformation should be asking who measures, reports, and governs the ...

LLMs are different from earlier automation. The right question is not which jobs will disappear but which tasks inside each job will change, and what that leave...

McKinsey's GenAI economic potential estimate is real and grounded in methodology. The way it gets used in boardrooms strips out the conditions that make it mean...

Most ML models fail after deployment, not during training. MLOps is the discipline that finally takes production seriously.

The same path dependence and absorptive capacity dynamics that explain organizational learning traps also explain why models degrade when trained on their own o...

Gartner says multiagent systems inquiries surged 1,445% but the real gap is not better agents. It is coordination architecture.

Open-weight AI is digital commons at global scale. The question is not open versus closed but governance design, and Ostrom's principles tell us where to start.

RLHF mirrors Giddens' duality of structure, creating a recursive loop where human preferences shape AI outputs and reshape human preferences.

Weick's sensemaking theory says people act first and make sense later, preferring plausibility over accuracy. That explains everything from the Twitter rebrand ...

Employees are pasting internal documents into consumer AI tools. The risk profile is different from shadow IT, but the pattern is identical.

Ferneley and Sobreperez's workaround types explain why employees bypass official AI tools. Shadow AI is task-technology fit failure, not compliance failure.

McKinsey's State of AI 2025 puts GenAI's economic potential at $2.6T to $4.4T annually. DORA 2024 shows why most software teams are not capturing their share of...

Why 'use' is the wrong construct for agentic AI, and what delegation theory actually means for how we study human-AI interaction.

Synthetic data offers a way to train AI on sensitive domains without exposing real people's records. The privacy tradeoff is real but often misunderstood.

Trust, trustworthiness, reliance, and delegation operate at different levels with different antecedents. Treating them as interchangeable is not a measurement c...

AI-native platforms let anyone generate code from a prompt. Trist and Bamforth showed in 1951 why optimizing only the technical subsystem creates a maintenance ...

Weick said people prefer plausible stories over accurate ones. XAI built tools for accuracy. It has been solving the wrong problem since 1995.

When AI generates unverified data at scale, the assumption that data is trustworthy by default becomes the most dangerous assumption in your architecture.

Gartner predicts 50% of organizations will adopt zero-trust data governance by 2028. The driver is a problem AI has created: we cannot tell where data came from...