What the survey says

Capgemini's numbers describe a trust problem with a twist. Only 2% of organizations had fully scaled agent deployment, and nearly two in five executives believed the risks of AI agents outweighed the benefits. But trust is not simply collapsing - it is being earned unevenly. Among organizations in the implementation phase, 47% report above-average trust, versus 37% of those still exploring. Trust follows contact, not contemplation. Meanwhile 93% of business leaders believed scaling agents within 12 months would bring competitive advantage, while nearly half of organizations lacked a strategy for doing it. Capgemini projects up to $450 billion in economic value from agentic AI by 2028 - treat that as the vendor's projection, not a forecast - with organizations at scaled implementation projected to realize about $382 million on average over three years, against $76 million for the rest.

What the code says

Snyk's data is behavioral, drawn from what organizations actually commit to their repositories, and it is blunter. The full AI footprint - agent frameworks, MCP servers, retrieval systems, vector databases, datasets - is roughly three times what a model inventory shows, and that ratio held across every region measured. Only 51% of organizations with at least one deployed model declare any dataset in their repositories: half cannot trace, at code level, what data their production models touch. The model mix is also shifting fast: Anthropic's share of enterprise model occurrences rose from 4% to 11% while OpenAI's fell from 44% to 35%. Both findings come with a caveat: Snyk's sample is organizations that use Snyk, and Snyk sells the platform that closes the gaps it measures.

The caveats

The Capgemini survey was fielded in April 2025 - its trust numbers describe mid-2025 sentiment, and attitudes have likely moved since. Snyk's repository scans show what is deployed, not what is used or valuable. Neither dataset tells you whether agents are working; together they tell you how organizations are behaving.

My read

Read together, the two datasets describe an industry deploying faster than it can see. Three implications. First, visibility precedes autonomy: if your AI inventory is a list of models, you are governing a third of your surface - the agents, tools and data pipelines are the majority. Second, the trust numbers are not a reason to slow down; Capgemini's own data says trust grows with implementation. They are a reason to deploy in ways that earn trust - narrow scopes, human oversight, measured expansion - rather than pilots designed to impress a board. Third, the data-lineage gap is the sleeper issue: half of model-deploying organizations cannot link a production model to its data. When the first serious agent incident hits your industry, "we cannot say what it touched" is the answer that turns an incident into a scandal. The companies that scale agents successfully will be the ones that built the map before they needed it.

Sources

• Capgemini Research Institute press release, Trust and human-AI collaboration set to define the next era of agentic AI (16 July 2025, survey methodology included): https://www.capgemini.com/news/press-releases/trust-and-human-ai-collaboration-set-to-define-the-next-era-of-agentic-ai-unlocking-450-billion-opportunity-by-2028/
• Snyk, 2026 State of Agentic AI Adoption: Volume II announcement (3 August 2026, methodology included): https://snyk.io/news/snyk-2026-state-of-agentic-ai-adoption-volume-ii/
• Snyk, 2026 State of Agentic AI Adoption: Volume II (report page): https://snyk.io/lp/state-of-agentic-ai-adoption-vol-II/

The boom, in numbers

• Published GenAI patent families rose from roughly 14,000 in 2023 to over 37,000 in 2025. More than 56,000 new families were published in 2024 and 2025 alone - exceeding the entire cumulative output of the preceding decade (2014 to 2023).
• GenAI now accounts for 8.7% of all AI patent families, up from 6.1% in 2023.
• Large language models overtook generative adversarial networks as the dominant model type, outnumbering them nearly three to one in 2025 (14,100 versus 5,200 published families). Diffusion models rank third. Text-related filings more than tripled in two years, and software/code and 3D modeling are among the fastest-growing smaller categories.

The map is not the product map

WIPO attributes SoftBank's position to its broad AI infrastructure and investment strategy, not to shipping models. Tencent, Ping An and Baidu remain the largest cumulative owners; Alphabet is now the largest US-based GenAI patent owner. The new entrants in the top 25 are telling: State Grid Corporation of China, Nvidia, Inspur Group and Bosch - a utility, a chipmaker, an infrastructure firm and an industrial conglomerate.

China-based inventors published over 43,000 GenAI patent families in 2024 and 2025, more than the country's entire output from 2014 to 2023. But the fastest relative growth is elsewhere: Japan at a 210% compound annual growth rate, Switzerland 124%, Canada 109%, the United States 92% and Germany 85%. Volume leadership is Chinese; growth leadership is dispersing.

The caveats

Patent counts measure filings, not quality, enforceability or product success. There is a roughly 18-month lag from filing to publication, so the most recent year is inherently incomplete - WIPO's own framing is that the post-ChatGPT surge is still landing in the data. Patents are a map of where organizations place expensive legal bets, not a scoreboard of who is winning.

My read

Three implications. First, freedom to operate is now a real workstream. With 56,000 new families in two years, shipping a GenAI feature in a crowded area without an IP scan is a legal-risk decision, whether or not anyone made it consciously. Second, stop reading model-lab headlines as IP position: the biggest holder is an investment group, and the entrant list says utilities, industrials and chipmakers are staking claims too. GenAI IP is becoming an every-industry game. Third, the globalization matters for talent and policy assumptions - the countries growing fastest are not the ones most strategies were built around. The IP ground is being laid right now, at several times the historical rate. Choosing not to have a patent strategy is still a patent strategy - a default one, written by your competitors.

Sources

• WIPO, Patent Trends Update in GenAI (Technology SPARK Report), Executive summary: https://www.wipo.int/web-publications/spark-patent-trends-update-in-genai/en/executive-summary.html
• WIPO news release, The GenAI Landscape Is Getting Bigger And More Competitive According To New WIPO Report (14 July 2026): https://www.wipo.int/en/web/patent-analytics/w/news/2026/the-genai-landscape-is-getting-bigger-and-more-competitive-according-to-new-wipo-report
• WIPO, Patent Trends Update in GenAI (report landing page): https://www.wipo.int/web-publications/spark-patent-trends-update-in-genai/en/index.html

What the logs show

Growth is real, and about half of it is deepening. Aggregate output tokens produced by ChatGPT Enterprise customers grew roughly sevenfold between June 2025 and March 2026. Within a consistent cohort of firms that adopted between January 2024 and June 2025, tokens grew nearly fourfold - so about half of the growth came from companies that had already adopted, not new logos.

Adoption is concentrated. Among U.S. public companies, adopters are larger, more valuable, and more R&D- and SG&A-intensive than non-adopters. The paper reads this as evidence that early adoption tracks prior investment in intangible and organizational capabilities.

Usage is bottom-up. Active use spans job functions and seniority levels, but early-career workers send many more messages than senior employees - and marketing and communications workers send more than executives.

Usage is broad. The most common tasks are writing, communication and information synthesis, with a long tail across research, planning, data analysis, legal and regulatory work, and finance.

The catch

This is a working paper, subject to change, by authors with financial ties to the vendor whose product generates the data. It covers only centrally administered enterprise seats - not personal accounts, API usage, or competing products. Treat it as one strong lens, not a picture of the whole market.

The measurement gap

A second August 2026 paper, Damon Young's Adoption Telemetry, argues the measurement itself needs to change: instead of surveying people, compute change-management stage progression directly from production usage signals. Young proposes NANTE, a five-stage model with defined telemetry thresholds, published openly so they can be tested and disproven, plus an open-source reference implementation. To his credit, he is explicit that the thresholds are proposed constructs validated only on synthetic populations - a research agenda, not a calibrated model. The contribution is the question, not the answer.

My read

Adoption is a behavior, not a sentiment. Three implications. First, if junior staff use AI the most and executives the least, then in most companies the strategy deck is behind the practice - leadership's job is to learn from the bottom, not only to cascade from the top. Second, define adoption in stages - access, habitual use, workflow change, measurable outcome - and instrument the progression between them. A seat-count dashboard measures procurement, not adoption. Third, vendor telemetry is a gift and a trap: the best behavioral data available, and structurally flattering to the vendor. Pair it with your own instrumentation. The companies that get this right will not be the ones with the best survey scores; they will be the ones that can see, week by week, where use is deepening, where it is stalling, and what it is actually for.

Sources

• Chatterji, Holtz, Rakholia, Tambe and Weeratunga, How Organizations Use AI: Evidence from ChatGPT (working paper, last updated 24 August 2026): https://arxiv.org/abs/2608.12236
• Young, Adoption Telemetry: Measuring Enterprise AI Adoption from Production Signals (arXiv, 22 August 2026): https://arxiv.org/abs/2608.23617

This is all verifiable at the source. The Commission's AI Act Service Desk publishes the consolidated Article 113 text with the amendments marked, plus an official implementation timeline. Both are worth bookmarking, because they are now the only version of the calendar that counts.

What actually applies today

The delay is surgical, not general. Reading the official timeline, a large share of the AI Act is already live:

• Since 2 February 2025: the general provisions - definitions and the Article 4 AI literacy obligation - plus the first prohibited practices.
• Since 2 August 2025: the rules for general-purpose AI models, the governance structure, and the penalties framework.
• Since 2 August 2026: the transparency obligations of Article 50 (including labeling of synthetic content), the innovation-support measures, and - critically - the start of enforcement at national and EU level for GPAI models, prohibitions, transparency rules and AI literacy.

And one date lands before the postponed high-risk wave: 2 December 2026, a month from now, when new prohibitions take effect for AI systems generating non-consensual sexual deepfakes and child sexual abuse material, alongside the transitional deadline for providers of synthetic-content systems already on the market before August 2026 to comply with Article 50(2).

Why Brussels blinked

The Commission's own FAQ is candid about the motivation: stakeholder consultations through 2025 "revealed implementation challenges" in rolling out the Act, and the Digital Omnibus is framed as a simplification measure to make compliance workable - part of a broader competitiveness push. Reading between the lines, the standards, guidance and national machinery needed to enforce Annex III were not going to be ready in time, and industry was not either.

My read

Deadline relief is not risk relief. Three things did not move with the dates. First, enforcement is already running for everything except high-risk systems, and the December 2026 deepfake prohibitions are weeks away. Second, the harms Annex III targets - a discriminatory hiring screen, a biased credit decision - are not legal until December 2027; they are enforceable reputational and legal problems right now under existing discrimination, consumer-protection and data-protection law, and no board should treat the Omnibus as permission to ship them. Third, the amendment itself is the lesson: this law will keep changing. The companies that memorized the 2024 timeline built compliance programs around dates that no longer exist. The durable capability is not a deadline on a slide; it is an inventory of your AI systems, a named owner for each, and a standing habit of checking the consolidated text rather than the conference talk. Sixteen extra months is enough time to build that muscle slowly and cheaply. It is also exactly enough time to waste and end up crash-buying consultants in the autumn of 2027. The Omnibus bet is that most companies will do the second. It is worth deciding, explicitly, to be the counterexample.

Sources

• European Commission, AI Act Service Desk, Article 113: Entry into force and application (consolidated text as at 27 July 2026, Digital Omnibus amendments marked): https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-113
• European Commission, AI Act Service Desk, Timeline for the Implementation of the EU AI Act: https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act
• European Commission, AI Act Service Desk, FAQ (Digital Omnibus section): https://ai-act-service-desk.ec.europa.eu/en/faq
• Regulation (EU) 2024/1689 (EUR-Lex record): https://eur-lex.europa.eu/eli/reg/2024/1689/oj

The macro math is not the problem

Start with the aggregate, which is genuinely reassuring. The WEF's Four Futures for Jobs report, drawing on its Future of Jobs analysis, projects that global macrotrends will create around 170 million new jobs by 2030 while displacing about 92 million - a net gain of roughly 78 million. AI adoption itself is nearly universal at the level of experimentation: the share of businesses using AI in at least one function has risen from 55% in 2022 to 88% in the latest estimates cited in the same report.

But net numbers hide the distribution, and the distribution is where the 12% lives. Jobs created and jobs displaced are not the same jobs, held by the same people, in the same places. The executives surveyed seem to know this: 24% expect AI to create new jobs, but more than twice as many expect displacement, and a quarter predict AI will increase industry concentration - a polite way of saying the gains pool at the top.

Displacement is specific, and so is the shortage

The OECD's Skills in the AI Age report (July 2026) adds the texture the aggregate hides. Its analysis finds the occupations most at risk of automation are not the ones most exposed to AI; displacement risk concentrates in routine, repetitive roles, while many AI-exposed occupations - engineers, for example - rest on non-routine skills that are harder to automate. At the same time, the OECD reports that skills shortages rank among the top barriers to AI adoption itself. And the skills actually demanded are telling: across the vacancy data the OECD cites, 72% of openings in highly AI-exposed occupations ask for at least one management skill, and 67% ask for business-process skills. The scarce resource is not prompt whisperers. It is people who can run redesigned workflows.

So the labor market is simultaneously short of AI-complementary skills and at risk of shedding routine work. Both can be true, and both are management problems.

The organizational multiplier

Microsoft's 2026 Work Trend Index - a vendor report, worth reading with that in mind - surveyed 20,000 workers who use AI across 10 countries (fielded February to April 2026 by Edelman Data x Intelligence) and analyzed trillions of anonymized Microsoft 365 signals. Its central finding: organizational factors - culture, manager support, talent practices - account for twice the reported AI impact of individual effort alone. Workers are already using AI in advanced, resourceful ways; most organizations, in Microsoft's framing, are not keeping up.

That finding rhymes with the WEF's scenario analysis. The Four Futures report does not predict one 2030; it lays out four, split by how fast AI advances and how ready workforces are. The difference between the good futures and the bad ones is not model capability. It is whether companies treat talent strategy as part of AI strategy - upskilling at scale, redesigning jobs rather than deleting them, aligning incentives so productivity gains show up in pay and not only in margin.

My read

The 12% wage figure is not a forecast; it is a confession. Executives who expect AI to lift margins but not wages are describing a plan in which the gains bypass the workforce - and then, often in the next breath, worrying publicly about why they cannot find AI-ready talent. The OECD data says the talent shortage and the displacement risk are two ends of the same transition, and the Microsoft data says the return on AI roughly doubles when the organization invests in the people side. The leaders who come out well in the WEF's scenarios are not the ones with the best models. They are the ones who decided the 12% was a number they intended to change. If you run an AI program, that is the question to put to your own plan: when the savings arrive, where exactly do they go - and who, on your current org chart, is being trained to do the work that remains?

Sources

• World Economic Forum, Four Futures for Jobs in the New Economy: AI and Talent in 2030 (2025; executive survey of 10,000+ leaders globally): https://reports.weforum.org/docs/WEF_Four_Futures_for_Jobs_in_the_New_Economy_AI_and_Talent_in_2030_2025.pdf
• OECD, Skills in the AI Age (July 2026): https://www.oecd.org/content/dam/oecd/en/publications/reports/2026/07/skills-in-the-ai-age_e8d8c1e6/972bd15e-en.pdf
• Microsoft, 2026 Work Trend Index Annual Report (vendor-published; 20,000 AI-using workers across 10 markets, Feb-Apr 2026): https://assets-c4akfrf5b4d3f4b7.z01.azurefd.net/assets/2026/09/2026_Work_Trend_Index_Annual_Report_090326_6a99e3f096b6e.pdf

Both numbers can be true at once. Reading them together is more useful than quoting either one alone.

The perception ledger and the P&L ledger

Wharton measures perception, and it says so. Leaders were asked whether they see positive returns, and roughly three-quarters do; four in five expect positive returns within two to three years. Notably, the optimism rises with altitude: 81% of VP-and-above respondents believe ROI is positive, versus 69% of mid-managers, who are presumably closer to the actual workflows. The same study finds real momentum underneath the perception: 82% of leaders use generative AI at least weekly (up 10 points year over year), 46% daily, and 72% now formally track ROI with structured, business-linked metrics like profitability and throughput.

BCG measures something harder: substantial value in the form of revenue and cash flow. By that standard, only 5% of firms - BCG calls them "future-built" - qualify. Another 35% are scaling and seeing some returns, and the remaining 60% (14% stagnating, 46% emerging) report minimal gains. The future-built 5% are pulling away: BCG reports 1.7x revenue growth, 1.6x EBIT margin, and 3.6x three-year total shareholder return compared with the lagging groups. One caveat from BCG's own methodology note: results reflect the business area each respondent knows best, not always the whole company.

A third data point sits between the two. In Google Cloud's 2026 ROI of AI survey of 2,403 senior executives, 86% agreed that AI has let them scale revenue or output without proportional increases in operating costs. That is again an agree/disagree perception question - and the survey is published by a company that sells AI - but it shows how broadly the "AI pays" belief now holds among executives.

Why the gap is organizational, not technical

If three-quarters of leaders see returns but only 5% see substantial value, the binding constraint is not model quality or access to tools. Everyone has the tools. BCG's comparison of what future-built companies do differently reads like an org chart critique, not a technology shopping list:

• Leadership owns it. Nearly 100% of future-built organizations report deeply engaged C-suites, versus 8% of laggards - a 12x difference. They are 3x more likely to have appointed a chief AI officer and run AI as a board- and CEO-sponsored multiyear program rather than a collection of pilots.
• They redesign workflows instead of sprinkling AI on old ones. Future-built firms have 5x more AI workflows deployed or scaled, with value-based prioritization and rigorous tracking of results.
• They aim at the core of the business. About 70% of AI's value sits in core functions - sales, marketing, operations - not support functions, up from 62% in BCG's 2024 study.
• They invest in people, not just platforms. The five strategies in BCG's playbook include broad-based upskilling and an AI-first operating model built on human-machine collaboration, alongside the data and platform foundation.

There is one more number in the BCG study worth sitting with: agentic AI, barely mentioned in 2024, already accounts for about 17% of total AI value in 2025. The value pool is shifting toward systems that act, and those systems demand exactly the workflow redesign and governance that the 60% have not done.

My read

The Wharton optimism is not wrong; it is measuring an earlier point on the curve. A team that saves an hour a day with a drafting assistant is genuinely "seeing positive returns." But that kind of win is cheap, widespread, and easy for competitors to copy, which is why it barely registers in BCG's revenue and cash-flow ledger. The uncomfortable implication of the VP-plus-versus-middle-manager gap (81% versus 69%) is that confidence about AI value is strongest furthest from the work. If you lead an AI program, the question worth asking in your next review is not "are we seeing returns?" - everyone says yes to that - but "which P&L line moved, and by how much?" BCG's data suggests the honest answer, for most companies, is still "none of them, yet." The fix is unglamorous: executive ownership, workflow redesign, tracking against a small number of core-business outcomes. None of it requires waiting for a better model.

Sources

• BCG, The Widening AI Value Gap, September 2025 (Build for the Future 2025 Global Study, n = 1,250 firms worldwide): https://media-publications.bcg.com/The-Widening-AI-Value-Gap-Sept-2025.pdf
• Wharton GBK, 2025 AI Adoption Report (~800 US senior decision makers at enterprises with 1,000+ employees and >$50M revenue, fielded June 26 - July 11, 2025): https://ai.wharton.upenn.edu/wp-content/uploads/2025/10/2025-Wharton-GBK-AI-Adoption-Report_Full-Report.pdf
• Google Cloud, The ROI of AI 2026 (N = 2,403 senior executives; vendor-published survey): https://services.google.com/fh/files/misc/google_cloud_roi_of_ai_2026.pdf

The headline number: serving a 7-billion-parameter model is 10 to 30 times cheaper than a 70-to-175-billion-parameter model, measured in latency, energy and compute. And specializing a small model takes a few GPU-hours of fine-tuning, not weeks.

This matters because agents change what a model is for. A chatbot holds open-ended conversation, and generality earns its keep. An agent, in the paper's framing, performs a small number of specialized tasks repetitively: plan, call a tool, check the result, repeat. Paying frontier-model prices for encyclopedic generality on every one of those calls is like hiring a professor to stamp forms. Surveys cited in the paper say more than half of large IT enterprises already run AI agents, with 21% having adopted just within the last year. The bill for all those repetitive calls is becoming a board-level topic.

The evidence is not just one vendor

NVIDIA's paper is an argument, so I went looking for independent measurements. The strongest I found is a 2026 study in the IJCAI proceedings from DFKI, Germany's AI research center, which benchmarked small models on accuracy and energy use across five tasks.

On GSM8K, a math reasoning benchmark, the best large model they tested, MistralLarge-3 at 675 billion parameters, consumed about 32 times the energy of the best small model, Qwen-3-4B, for roughly a 2-point absolute accuracy gain. The diminishing returns show up inside the small range too: on another benchmark, moving from Qwen-3-8B to Qwen-3-14B for a single percentage point of accuracy nearly doubled the inference energy.

The same study delivers the honest counterpoint, and it is the part most "small is better" takes skip. Smaller is not automatically cheaper. Comparing each small model against every larger one, the smaller model burned more energy in 14 to 43 percent of pairwise comparisons, depending on the benchmark. The reason is output length: a small model that rambles spends its savings on extra tokens. Size and cost-per-token alone do not determine efficiency. Behavior does.

Meanwhile, the enterprise constraint is real. Analyst firm Ecosystm found that 54% of organizations in Asia Pacific identify infrastructure limitations as the biggest barrier to AI deployment, and only 21% say they have the compute and infrastructure capabilities to support AI growth. Small models attack exactly that constraint. They run on consumer-grade hardware, at the edge, even offline, which also keeps data in-house.

Where small already works

In NVIDIA's tests on three open-source agent frameworks, specialized small models could potentially replace 40 to 70 percent of the LLM calls. The pattern is routing: a small model handles the routine steps, and the frontier model sees only the hard residue. Add fine-tuning that costs GPU-hours and finishes overnight, and each step of an agent can have its own specialist instead of renting one generalist.

What I would do with this

This section is my opinion, built from the sources above.

1. Map your agent's calls before choosing models. Most steps in an agentic workflow are narrow and repetitive, and those are the expensive ones to get wrong.
2. Default every step to a small model. Escalate to the frontier model only on measured failure, not on vibes.
3. Measure output length, not just model size. The IJCAI data says verbosity is an energy line item, so prompt and tune for terse answers.
4. Fine-tune instead of stacking ever-longer prompts. At a few GPU-hours per behavior, specialization is an afternoon job.
5. Keep the frontier model for the calls that genuinely need it, and be honest about how few those are.

Limits of this analysis

NVIDIA's paper is a position paper, meaning advocacy, from a company whose hardware powers both ends of this market. The IJCAI numbers come from specific benchmarks, models and one hardware setup, and the authors note results can vary with deployment. Ecosystm's figures cover Asia Pacific only. What I trust is the direction across all three: for repetitive agent work, model size is a cost decision more than a capability one. Exact savings will vary by workload.

Sources

• Belcak et al., NVIDIA Research, Small Language Models are the Future of Agentic AI (arXiv, 2025, v3 September 2026): https://arxiv.org/pdf/2506.02153
• Reichwald et al., DFKI, Mapping the Efficiency Landscape of Small Language Models (IJCAI 2026 proceedings): https://ijcai.org/proceedings/2026/0627.pdf
• Ecosystm, Right-Sizing Enterprise AI: The Rise of Small Language Models (July 2026): https://ecosystm.io/wp-content/uploads/2026/07/Ecosystm-Report_report-right-sizing-enterprise-ai-rise-of-small-language-models-Jul26.pdf

Even discounted for vendor optimism, the gap is the story of 2026. AI agents are reaching production faster than the identity programs meant to govern them, and on Okta's numbers only one in ten organizations has a well-developed strategy for the identities those agents use.

We spent two decades building identity programs for people: badges, single sign-on, multi-factor authentication, offboarding checklists. AI agents arrived faster than that machinery could adapt. What is left is a workforce most companies are not set up to count, attribute, or offboard.

The workforce you cannot see

Non-human identities are not new. Service accounts, API keys, OAuth tokens and bots have been multiplying for years. Research cited in a 2026 Cloud Security Alliance paper puts the ratio at 45 non-human identities for every human one across the average enterprise, and 144 to 1 in cloud-native environments, up from 92 to 1 in the first half of 2024. One published audit of a Fortune 500 financial institution found 4.2 million non-human identities against roughly 50,000 human user accounts.

AI agents turn a hygiene problem into an authority problem. A service account runs the same job every night. An agent decides which job to run. It reads email, queries the warehouse, calls vendor APIs and chains those steps together at machine speed. Every one of those actions rides on a credential someone issued and nobody reviewed.

The governance vacuum

The CSA paper, The Non-Human Identity Governance Vacuum, assembles findings from several industry surveys and reports, each with its own sample. The direction across them is consistent:

• 78% of organizations have no documented policy for creating or removing AI identities.
• More than 16% do not track the creation of AI-related identities at all.
• 51% report no clear ownership for AI identities.
• Only 20% have formal API key rotation procedures, and 24% take more than 24 hours to rotate or revoke a credential they already know is exposed.

The cost shows up in breach data. IBM's 2025 Cost of a Data Breach report, cited in the same paper, found that 97% of organizations that suffered an AI-related breach lacked proper AI access controls. Shadow AI, where employees point unapproved tools at company data, played a role in 20% of breaches and added roughly $670,000 to the average bill.

This is not a tooling gap you can buy your way out of. It is an ownership gap. Identities nobody owns do not get rotated, scoped, reviewed or revoked.

The standards are catching up

Two 2026 efforts are worth reading in full.

In February, NIST's National Cybersecurity Center of Excellence published a draft concept paper on applying existing identity standards to software and AI agents. The premise is blunt: the standards for identification, authentication and authorization already exist. The work now is demonstrating how they apply to agentic architectures, not inventing new ones.

In March, the Coalition for Secure AI published version 1.0 of its Agentic Identity and Access Management guidance. Its imperatives read like a job description for the next three years of IAM work: treat agents as first-class identities with their own lifecycle, eliminate standing privilege in favor of short-lived, task-scoped credentials, keep an agent's baseline permissions distinct from the rights of the human it acts for, and be able to prove control on demand, meaning you can reconstruct which agents existed and what they did.

Okta frames the same shift from the vendor side: identity is becoming the control plane for autonomous systems, because it is the one layer that sees every action an agent takes.

What I would do Monday morning

This section is my opinion, built from the sources above.

1. Count your agents before you govern them. You cannot secure identities you have never inventoried, and 16% of organizations are not even tracking new ones as they appear.
2. Give every agent a named owner. Not a team. A person who answers for what it does and confirms it still needs to exist.
3. Issue credentials like concert tickets, not apartment keys. Short-lived, scoped to the task, expired when the task ends.
4. Separate what the agent may do from what the human behind it may do. Delegation is where intent gets lost.
5. Practice firing one. Pick an agent, revoke its credentials, and time the drill. In the surveys CSA cites, roughly a quarter of organizations need more than a day to rotate or revoke an exposed credential. That is too long.

Limits of this analysis

The numbers here come from vendor-published research and industry surveys with different methodologies and populations, and some, like Okta's, do not describe their sample at all. Treat any single figure as directional. What I trust is the pattern: across four independent sources, identity governance is lagging agent deployment by a wide margin. This is analysis, not legal or security advice.

Sources

• Okta, The AI Agent Era: Rethinking Identity, Trust, and Control (2026): https://www.okta.com/sites/default/files/2026-02/The-AI-agent-era.pdf
• Cloud Security Alliance, The Non-Human Identity Governance Vacuum (May 2026): https://labs.cloudsecurityalliance.org/wp-content/uploads/2026/05/CSA_whitepaper_nonhuman_identity_agentic_ai_governance_v1-csa-styled.pdf
• Coalition for Secure AI, Agentic Identity and Access Management v1.0 (March 2026): https://www.coalitionforsecureai.org/wp-content/uploads/2026/04/agentic-identity-and-access-control.pdf
• NIST NCCoE, Accelerating the Adoption of Software and AI Agent Identity and Authorization, draft concept paper (February 2026): https://www.nccoe.nist.gov/sites/default/files/2026-02/accelerating-the-adoption-of-software-and-ai-agent-identity-and-authorization-concept-paper.pdf

What this is not: SR 26-2 is supervisory guidance, not law, and it is written for banking organizations, mostly those over $30 billion in assets. "Out of scope" does not mean unregulated: existing obligations, from consumer protection to safety and soundness, still apply to whatever a bank deploys. But the directional signal reaches well beyond banking. For now, agent governance will be practitioner-led in every industry.

On April 17, 2026, the Federal Reserve, the OCC and the FDIC issued SR 26-2, a revised guidance on model risk management. It replaces SR 11-7, the 2011 letter that taught a generation of risk teams what a model inventory, validation and effective challenge look like. Fifteen years of supervisory experience, condensed into one document.

And then, in a footnote, the agencies step around the biggest story in the room: generative AI and agentic AI models "are novel and rapidly evolving. As such, they are not within the scope of this guidance."

The systems your organization is deploying fastest are the ones the flagship framework does not cover.

Read the footnote twice

The footnote does not stop at the exclusion. Its next sentence says a banking organization's own "risk management and governance practices should guide the determination of appropriate governance and controls for any tools, processes, or systems not covered in this document."

Regulators did not write the rules for AI agents. They handed practitioners the pen.

This is the real story of SR 26-2. The rest of the letter is genuinely useful: a sharper risk-based approach, materiality thinking, proportionality for smaller institutions. But the footnote is where leadership lives. The agencies are saying, in plain supervisory language: you understand these systems better than we can write rules for them right now. Show us what good looks like.

Why an agent is not just another model

Model risk management was built for artifacts. A model is a thing you can version, validate before first use, and monitor on a cycle. It takes inputs and returns estimates. Its worst day is a wrong number.

An agent is different in kind. It has goals, tools, credentials and memory. It calls APIs, moves data, drafts the email, touches the production system. Its worst day is not a wrong answer. It is a wrong action, executed a thousand times at machine speed before anyone notices.

That is why simply stretching SR 11-7 thinking over agents will not work. The unit of governance is no longer the model. It is the actor.

What I would do on Monday morning

The spine of model risk management still works. It just needs to be pointed at behavior, not artifacts. Five moves borrow that spine and extend it:

1. Inventory agents, not just models. Every agent gets a name, an owner, a list of tools and permissions, the data it can touch, and its blast radius. Most organizations cannot produce this list today.
2. Tier by autonomy and blast radius. SR 26-2's materiality thinking applies directly: an agent that summarizes documents is not an agent that moves money. Governance effort follows consequence, not novelty.
3. Validate behavior, not just outputs. Test trajectories, not single responses. Red-team the agent against the actions it is allowed to take before you widen those permissions.
4. Keep effective challenge. Someone independent needs standing to question what the agent is allowed to do, not only how well it does it. The developers of the agent cannot be its only reviewers.
5. Monitor in real time, with a kill switch and an audit trail. Periodic review cycles assume a static artifact. Agents need continuous observation, clear escalation, and a log you can reconstruct an incident from.

None of this waits for a regulator. The World Economic Forum's work on agent evaluation and governance, and the security research coming out of firms like Sophos, show the practitioner community is already building these foundations. The question is whether your organization is building its own or waiting to inherit someone else's.

The gap is the opportunity

Adoption is not waiting. Wharton and GBK find roughly three in four enterprise leaders reporting positive ROI from generative AI, while BCG finds only about five percent of companies achieving AI value at scale. My read of that gap: the difference is not better models. It is operating discipline, and I believe governance sits at the core of it.

There will be an SR letter for AI agents one day, in banking and beyond it. The leaders who treat today's gap as an assignment will write the practices it cites. The ones who wait will inherit them.

I know which side of that footnote I want to be on.

Sources

• SR 26-2, Revised Guidance on Model Risk Management (Federal Reserve, OCC, FDIC, April 17, 2026)
• AI Agents in Action: Foundations for Evaluation and Governance (World Economic Forum, 2025)
• The Widening AI Value Gap (BCG, 2025)
• 2025 AI Adoption Report (Wharton Human-AI Research and GBK Collective)