Brief

IBM Think 2026: From AI Pilots to an Operating Model

In Boston last week, IBM centered its AI argument with bets on sovereignty, agent governance, AI on the mainframe, and a quantum-and-AI flywheel.

Di John Kanan, Arun Ganti, Daniel Webster-Clark, and Ankur Shah

First published on maggio 13, 2026
Tempo di lettura min.


In evidenza IBM moved the AI conversation from model adoption to operating model redesign, and from AI-enabled business to AI-first. Product launches took center stage, including Sovereign Core and IBM Bob, among others. Sovereignty was a major theme, with IBM highlighting its hybrid footprint and global posture as differentiators. Quantum breakthroughs across medical fields and other industries continue to validate the revolutionary potential, and a quantum-and-AI flywheel is beginning to take shape. A year ago at Think, the conversation was about getting from AI prototype to production. This year, the bar moved further. IBM’s stance at Think was that value creation from AI will accrue to organizations that rebuild their operating model around AI, not those that run point workflow optimizations. IBM’s own client zero story, with $4.5 billion in productivity unlocked over three years, was cited throughout Think as a natural proof point. The human side mattered just as much. Eighty-three percent of CEOs in IBM’s 2026 survey said AI success depends more on adoption than on the technology itself, with executive panelists from IBM clients echoing the sentiments live. Bain shares this view that AI adoption—not tooling—limits AI ROI, finding that AI leaders with scale adoption deliver 10% to 25% EBITDA gains. “The enterprises pulling ahead are not deploying more AI—they’re redesigning how their business operates.” Arvind Krishna \| IBM Chairman and Chief Executive Officer The agentic operating layer is taking shape Rob Thomas, IBM’s senior vice president of software and chief commercial officer, organized IBM’s operating model argument around four imperatives: intelligence, action, operations, and trust, supported by marquee launches like IBM Bob (AI coding assistance), Concert (agentic IT operations platform), and Confluent (real-time data streaming). In addition, IBM positioned Apptio’s role in the AI operating model as making AI and broader costs easier to see across the tech stack for ROI understanding. IBM’s architecture aligns well with what Google and Nvidia sketched out earlier this year. Differentiation now sits below the architecture—in identity, runtime control, data context, and model choice. IBM’s acquisition of Confluent ($11 billion) inside watsonx.data is the most strategically important plumbing move IBM has made this cycle. Agents need streaming, governed, contextual data, and most enterprise architectures were built for batch. Confluent streamlines getting the right data to your agent base in real time. IBM focused throughout the week on Bob, an agentic development platform that moves beyond coding assistance to accelerating the full development cycle. It is purposefully model agnostic, dynamically routing tasks across Anthropic's Claude, Mistral, IBM's Granite, or specialized fine-tuned models, optimizing for accuracy, performance, and cost. IBM cited more than 80,000 users and 45% productivity gains. This approach aligns with IBM’s point of view that fit-for-purpose models will ultimately account for the vast majority of AI consumption, which explains why IBM is focusing on model routing more than picking a single frontier model. Governance must move alongside agent proliferation. With IBM citing a 1:120 human-to-non-human identity ratio as the near-term reality, traditional access control simply cannot scale; governance must be built into the data and runtime layers, not retrofitted later. Sovereignty has moved to the front of the agenda Geopolitical tension and regulatory pressure (the EU’s October 2025 sovereignty framework was cited repeatedly) supported the rise of sovereignty from compliance topic to board-level concern. Sovereign Core, now generally available, embeds policy and compliance enforcement at the runtime layer, with continuous compliance monitoring, automated evidence generation, and in-boundary identity and encryption. The "so what" for non-regulated businesses is broader than it looks. As more agents act on enterprise data across jurisdictions, the question of which data crossed which border, under whose authority, and with what audit trail becomes relevant for any global business. Hyperscalers can and will build sovereign cloud regions, but IBM’s combination of on-premises capability, hybrid-by-design architecture, deep regulated-industry footprint, and a non-US ecosystem story (notably the TCS partnership with the Indian government on a national quantum stack) is harder for AWS, Microsoft, or Google to fully replicate. (For more, read the Bain Brief “Cloud Sovereignty: Maintaining Control Without Losing Scale.”) AI on the mainframe is the full-stack story IBM uniquely can tell A theme that gained weight as the week progressed was AI deployed inside the existing IT estate, not alongside it. The z17 ships with the Telum II processor and on-chip AI accelerator, alongside Spyre PCIe accelerators for generative workloads—running 450 billion inferences per day at sub-millisecond latency, enough to move from sampled fraud checks to full-transaction screening. IBM Bob’s Premium Package for Z, in tech preview, layers full-lifecycle agentic SDLC on top: COBOL, JCL, and PL/I refactoring, automated test coverage, and end-to-end modernization. IBM positioned these capabilities as valuable tools for companies sitting on mainframe estates with shrinking COBOL talent pools. Royal Bank of Canada was the cleanest illustration on stage: Across 29 countries and 19 million clients, more than 45% of total workloads now run through hybrid cloud, with an ambition of $700 million to $1 billion in enterprise value by 2027. The broader takeaway: Bringing AI to the data and critical workloads running on mainframes, rather than the other way around, is becoming a more practical option. Quantum took its largest practical step yet IBM now has about 80 quantum computers in cloud-accessible production, and the field has visibly crossed from science to engineering. The most compelling and exciting use case presented at the conference was Cleveland Clinic and Riken’s simulation of a 12,635-atom protein complex (trypsin) on IBM’s quantum hardware. This is roughly 40 times larger than what was possible six months ago, with up to a 210-times improvement in accuracy. Other highlights included momentum on materials engineering in aerospace, risk models in insurance, and nuclear fusion energy. A simulation of a 12,635-atom protein complex on IBM’s quantum hardware was x larger and x more accurate than what was possible six months ago Across these projects and more, IBM shared an emerging best practice pairing quantum and AI as complementary engines: Quantum surfaces information from problems that are exponentially hard for classical computers, and AI learns from it and operationalizes it. For most business leaders, the immediate move is defensive: Post-quantum cryptography migration is the no-regrets action today, and the rationale underlying Bain's 2026 strategic collaboration with IBM. The bigger winners will tackle the harder, more valuable work of identifying the two or three problems that would benefit from quantum once it arrives, and getting data, talent, and partnerships in place now. The bottom line IBM has assembled the most coherent enterprise AI operating model story it has told in years, and the substantive pieces (Confluent inside watsonx.data, Sovereign Core at runtime, Bob across the SDLC) are real. The architectural endpoint is broadly aligned with where other leaders such as Google and Nvidia are heading, but IBM’s hybrid posture, regulated-industry depth, and global sovereignty footprint give it differentiated angles the hyperscalers will likely struggle to match for hybrid enterprises. The quantum-and-AI flywheel is beginning to take shape, and the window for early movers will not stay open forever. For enterprise leaders, the bar has been raised again. Pilots are no longer enough; bolt-on AI is no longer enough. The work is to redesign how the business operates around AI, govern it from the data layer up, treat sovereignty as a strategic concern rather than a compliance issue, take AI-on-mainframe options seriously, and start putting the quantum-and-AI flywheel together before the window narrows.

In evidenza

IBM moved the AI conversation from model adoption to operating model redesign, and from AI-enabled business to AI-first.
Product launches took center stage, including Sovereign Core and IBM Bob, among others.
Sovereignty was a major theme, with IBM highlighting its hybrid footprint and global posture as differentiators.
Quantum breakthroughs across medical fields and other industries continue to validate the revolutionary potential, and a quantum-and-AI flywheel is beginning to take shape.

A year ago at Think, the conversation was about getting from AI prototype to production. This year, the bar moved further. IBM’s stance at Think was that value creation from AI will accrue to organizations that rebuild their operating model around AI, not those that run point workflow optimizations. IBM’s own client zero story, with $4.5 billion in productivity unlocked over three years, was cited throughout Think as a natural proof point.

The human side mattered just as much. Eighty-three percent of CEOs in IBM’s 2026 survey said AI success depends more on adoption than on the technology itself, with executive panelists from IBM clients echoing the sentiments live. Bain shares this view that AI adoption—not tooling—limits AI ROI, finding that AI leaders with scale adoption deliver 10% to 25% EBITDA gains.

“The enterprises pulling ahead are not deploying more AI—they’re redesigning how their business operates.”

Arvind Krishna | IBM Chairman and Chief Executive Officer

The agentic operating layer is taking shape

Rob Thomas, IBM’s senior vice president of software and chief commercial officer, organized IBM’s operating model argument around four imperatives: intelligence, action, operations, and trust, supported by marquee launches like IBM Bob (AI coding assistance), Concert (agentic IT operations platform), and Confluent (real-time data streaming). In addition, IBM positioned Apptio’s role in the AI operating model as making AI and broader costs easier to see across the tech stack for ROI understanding.

IBM’s architecture aligns well with what Google and Nvidia sketched out earlier this year. Differentiation now sits below the architecture—in identity, runtime control, data context, and model choice.

IBM’s acquisition of Confluent ($11 billion) inside watsonx.data is the most strategically important plumbing move IBM has made this cycle. Agents need streaming, governed, contextual data, and most enterprise architectures were built for batch. Confluent streamlines getting the right data to your agent base in real time. IBM focused throughout the week on Bob, an agentic development platform that moves beyond coding assistance to accelerating the full development cycle. It is purposefully model agnostic, dynamically routing tasks across Anthropic's Claude, Mistral, IBM's Granite, or specialized fine-tuned models, optimizing for accuracy, performance, and cost. IBM cited more than 80,000 users and 45% productivity gains. This approach aligns with IBM’s point of view that fit-for-purpose models will ultimately account for the vast majority of AI consumption, which explains why IBM is focusing on model routing more than picking a single frontier model.

Governance must move alongside agent proliferation. With IBM citing a 1:120 human-to-non-human identity ratio as the near-term reality, traditional access control simply cannot scale; governance must be built into the data and runtime layers, not retrofitted later.

Sovereignty has moved to the front of the agenda

Geopolitical tension and regulatory pressure (the EU’s October 2025 sovereignty framework was cited repeatedly) supported the rise of sovereignty from compliance topic to board-level concern. Sovereign Core, now generally available, embeds policy and compliance enforcement at the runtime layer, with continuous compliance monitoring, automated evidence generation, and in-boundary identity and encryption.

The "so what" for non-regulated businesses is broader than it looks. As more agents act on enterprise data across jurisdictions, the question of which data crossed which border, under whose authority, and with what audit trail becomes relevant for any global business. Hyperscalers can and will build sovereign cloud regions, but IBM’s combination of on-premises capability, hybrid-by-design architecture, deep regulated-industry footprint, and a non-US ecosystem story (notably the TCS partnership with the Indian government on a national quantum stack) is harder for AWS, Microsoft, or Google to fully replicate. (For more, read the Bain Brief “Cloud Sovereignty: Maintaining Control Without Losing Scale.”)

AI on the mainframe is the full-stack story IBM uniquely can tell

A theme that gained weight as the week progressed was AI deployed inside the existing IT estate, not alongside it. The z17 ships with the Telum II processor and on-chip AI accelerator, alongside Spyre PCIe accelerators for generative workloads—running 450 billion inferences per day at sub-millisecond latency, enough to move from sampled fraud checks to full-transaction screening. IBM Bob’s Premium Package for Z, in tech preview, layers full-lifecycle agentic SDLC on top: COBOL, JCL, and PL/I refactoring, automated test coverage, and end-to-end modernization. IBM positioned these capabilities as valuable tools for companies sitting on mainframe estates with shrinking COBOL talent pools.

Royal Bank of Canada was the cleanest illustration on stage: Across 29 countries and 19 million clients, more than 45% of total workloads now run through hybrid cloud, with an ambition of $700 million to $1 billion in enterprise value by 2027. The broader takeaway: Bringing AI to the data and critical workloads running on mainframes, rather than the other way around, is becoming a more practical option.

Quantum took its largest practical step yet

IBM now has about 80 quantum computers in cloud-accessible production, and the field has visibly crossed from science to engineering. The most compelling and exciting use case presented at the conference was Cleveland Clinic and Riken’s simulation of a 12,635-atom protein complex (trypsin) on IBM’s quantum hardware. This is roughly 40 times larger than what was possible six months ago, with up to a 210-times improvement in accuracy. Other highlights included momentum on materials engineering in aerospace, risk models in insurance, and nuclear fusion energy.

A simulation of a 12,635-atom protein complex on IBM’s quantum hardware was

larger and

more accurate than what was possible six months ago

Across these projects and more, IBM shared an emerging best practice pairing quantum and AI as complementary engines: Quantum surfaces information from problems that are exponentially hard for classical computers, and AI learns from it and operationalizes it.

For most business leaders, the immediate move is defensive: Post-quantum cryptography migration is the no-regrets action today, and the rationale underlying Bain's 2026 strategic collaboration with IBM. The bigger winners will tackle the harder, more valuable work of identifying the two or three problems that would benefit from quantum once it arrives, and getting data, talent, and partnerships in place now.

The bottom line

IBM has assembled the most coherent enterprise AI operating model story it has told in years, and the substantive pieces (Confluent inside watsonx.data, Sovereign Core at runtime, Bob across the SDLC) are real. The architectural endpoint is broadly aligned with where other leaders such as Google and Nvidia are heading, but IBM’s hybrid posture, regulated-industry depth, and global sovereignty footprint give it differentiated angles the hyperscalers will likely struggle to match for hybrid enterprises. The quantum-and-AI flywheel is beginning to take shape, and the window for early movers will not stay open forever.

For enterprise leaders, the bar has been raised again. Pilots are no longer enough; bolt-on AI is no longer enough. The work is to redesign how the business operates around AI, govern it from the data layer up, treat sovereignty as a strategic concern rather than a compliance issue, take AI-on-mainframe options seriously, and start putting the quantum-and-AI flywheel together before the window narrows.