State of the AI Estate 2027
Most enterprises answer the question of how many AI agents they run with a confident guess. This paper brings the evidence together, and shows how to replace the guess with a count.
The census is the oldest act of governance.
Every state that lasted long enough to be remembered counted something. Rome counted its citizens. In 1086 William the Conqueror had England surveyed village by village, and the result became the Domesday Book.
They counted for the plainest of reasons. You cannot tax, defend, feed or protect a population you have only guessed at. A census is where governing starts because it is where wishing stops. Plans made on a guessed number tend to be wrong in the same direction every time, because the guess comes from the part of the population that is easy to see.
Enterprises now hold a new population. It is software that decides, acts and holds credentials. It reads documents, calls tools, moves money in small amounts, writes to systems of record and talks to customers. Most of it arrived without a ceremony. Nobody hired it, nobody onboarded it, and in many companies nobody can say how many of it there are.
That is the subject of this paper. We gathered the best published evidence from Gartner, IDC, EY, Accenture, a16z, the Cloud Security Alliance, IBM and a number of identity security researchers, and we read it together. We expected to find a measurement problem. We found something more useful. Three independent surveys, run by different firms, asked leaders whether they could see their AI agents and identities. In each case between two thirds and 95% said yes. When the evidence was measured, in the same surveys or alongside them, the share with a complete inventory, with automated discovery or with ongoing monitoring fell between 18% and 36%.
Confidence ran between 68% and 95%. Evidence ran between 18% and 36%. That gap is the story of the AI estate in 2027.
We think this gap is the most important and least expensive problem in enterprise AI. It is important because every other control, whether security review, spend discipline, audit readiness or incident response, begins with a list of what exists. It is inexpensive because the first useful version of that list does not require a platform. It requires a definition, a few ways of looking, a named person on every record, and the discipline to repeat the exercise each quarter.
THE LYZR RESEARCH TEAM
November 2026
Seven findings, and what to do about them.
The evidence points one way. Enterprises are running more agents than they have counted, they are confident about a count they have not made, and the cost of the gap is already arriving.
Nobody agrees on the number.
Credible estimates of agents per large company run from about a dozen distinct agents to more than 150,000 embedded agents and identities. IDC expects over one billion active agents worldwide by 2029, forty times the 2025 level. The spread comes from definitions, and your own number depends on yours.
Confidence runs far ahead of evidence.
95% of leaders in one survey believed they had adequate visibility of AI and machine identity exposure, yet only 36% were monitoring it. In another, 95% were confident in their inventory of machine identities, and 27% had automated discovery and lifecycle management. At least half of organizations in the Cloud Security Alliance study were both confident and surprised.
The shadow is the norm.
82% of organizations found shadow agents in the past year. Two thirds of EY’s 975 large-company respondents let employees build agents, and half of them lack a high level of visibility into it. Unauthorized AI tools ran a median of 403 days before detection.
Ownership thins out at the exit.
Only 21% have a formal decommissioning process, and only 32% secure agents with the rigor they apply to human employees. 69% of identity professionals report that between 11% and 50% of their credentials are expired or unused but still active.
The bill is already arriving.
65% had an agent incident in the past year. 99% of companies in EY’s survey reported financial losses from AI risks, averaging a conservative US$4.4 million. Gartner expects the share of enterprise GenAI applications with five or more minor incidents a year to rise from 9% in 2025 to 25% by 2028.
Every rulebook begins with an inventory.
The EU AI Act timetable, NIST’s AI Risk Management Framework and the OWASP Top 10 for Agentic Applications each assume you know what you run, who owns it and what it can reach. Gartner now advises domain-level ownership of every MCP server.
The fix is small and compounds.
A register with one named owner per agent, built from five lenses and repeated quarterly, is the cheapest control on the page. Accenture finds that the 10% of organizations it rates as most security-mature are 69% less likely to face advanced attacks. The association is not proof of cause, and it still points where a leader would want to go.
Five planning assumptions for 2027
Our judgments, in the format Gartner and IDC use, with the confidence we place in each.
| By the end of 2027 | Confidence |
|---|---|
| The agent register is a standard item in audit and insurer questionnaires | 80% |
| The first full census of a large estate finds more agents than any survey figure, and the board is surprised | 70% |
| Retirement of agents becomes a named control with its own owner | 65% |
| Every production agent carries its own identity and its own audit trail | 60% |
| Finance maps model spend to named agents and owners | 55% |
Five moves for the next 90 days
Agree what counts as an agent. Run a scan across cloud accounts, SaaS platforms, code repositories and gateway logs. Ask every team what they have built.
Put one named person on every record. Triage the agents that nobody owns and the agents that nobody knew about.
Run the census again. Compare the two counts. Report the difference to whoever owns risk, and set the cadence for the next one.
Hold one rule above the rest. An agent with no owner does not make production calls. Everything else follows from being able to enforce that.
Read these first.
The number nobody has
A dozen, or a hundred and fifty thousand.
There is no accepted number for how many agents a large enterprise runs. That fact deserves more attention than any single estimate, because it tells you the instrument for measuring it does not yet exist in most companies.
Start with what the sources actually say. Gartner reported in 2025 that the average Fortune 500 company ran fewer than fifteen agents, and projected in April 2026 that it would run more than 150,000 by 2028. OutSystems, surveying 1,900 IT leaders in the winter of 2025 and 2026, found about twelve distinct agents per organization, rising to about twenty by 2027. IBM, at its Think conference in 2026, put a large enterprise at roughly 1,600 agents by the end of 2026. At the planetary scale, IDC projects more than one billion actively deployed agents by 2029, forty times the 2025 level, taking over 217 billion actions a day.
These figures do not contradict each other as much as they appear to. They count different things.
Ask how many devices a household owns. Count the laptops and the answer is three. Count the phones, TVs, watches, bulbs and thermostats that touch the network and it is closer to forty. Both answers are correct. The first counts what somebody chose to buy. The second counts everything that is connected. The agent estimates differ for the same reason. A survey of distinct agents that a team deliberately built returns a small number. A projection that includes every embedded assistant and every machine identity that can act returns a very large one.
Adoption claims diverge for the same reason
The same spread appears when surveys ask whether organizations are using agents at all. In a May 2025 Gartner poll of 147 CIOs and IT leaders, 24% had deployed a few agents and 4% had deployed more than a dozen, so 28% in total, while half said they were still researching and experimenting. Okta’s research, conducted among 261 senior executives, found that 91% of organizations already use agents. SpyCloud’s survey of 750 security leaders found that 91% use AI tools or agents with access to internal systems, applications or data. Permiso’s survey of more than 500 security and identity professionals found that 92% have agents accessing production or sensitive data.
The distance between 28% and 92% reflects both the passage of time and the width of the question. Deploying a few agents on purpose is one thing. Having agents with access to production data, which includes assistants embedded in the software you already pay for, is another. Seen from a distance, almost every large organization has agents. Seen close up, most cannot say which ones.
Volume will outrun review
The count matters more each quarter because the volume is rising faster than any manual review can follow. IDC expects Global 2000 agent use to grow tenfold by 2027, with token and API call volumes growing a thousandfold.
The model layer is multiplying as well. In a16z’s survey of 100 enterprise CIOs across 15 industries in May 2025, 37% of respondents used five or more models, up from 29% a year earlier, and leaders expected their large language model spend to grow about 75% over the following year.
The rapid acceleration and increasing agency of AI agents necessitates a shift beyond traditional human oversight.Avivah Litan, VP Distinguished Analyst, Gartner, June 2025
The sources agree on direction and disagree on size. For planning, treat the lowest estimate as what a team would admit to, the middle estimates as what a scan would find, and the highest as what your SaaS vendors are quietly adding.
- Write a one-paragraph definition of an agent for your organization, with an example that is in and an example that is out. A definition that fits on one page can be used by a hundred teams.
- Decide whether embedded vendor assistants count. If they do, your number will be larger, and your task will be harder and more honest.
- Record the date with every count. A number without a date is a rumor by the next quarter.
- Platform vendors shipping built-in agent inventories that auditors accept
- Analyst definitions converging, which would make surveys comparable
- The ratio of agents to employees appearing in earnings calls and annual reports
The confidence paradox
Sure of the speed, and just pulled over.
If you wanted to design a survey to detect overconfidence in AI visibility, you would ask two questions. Do you believe you can see it, and what are you doing to see it. Several research teams, without coordinating, did exactly that.
The pattern is consistent enough to name. In SpyCloud’s survey of 750 security leaders, 95% believed they had adequate visibility into AI and non-human identity exposure, and 36% were monitoring it. In Permiso’s survey, 95% were confident in their inventory of non-human identities with access to critical systems, but only 27% had fully automated both discovery and lifecycle management, and about half discovered new identities only through periodic audits or after an incident. In the Cloud Security Alliance’s study with Token Security, 68% were confident in their visibility of agents and 82% had found shadow agents in the previous year.
EY’s survey of 975 C-suite leaders at companies with revenue above US$1 billion adds a revealing detail. On average, only 12% of respondents correctly identified the appropriate controls for five AI risks. Chief risk officers scored 11%, slightly below average. The people who most believed they were in control were rarely tested on it.
Why confidence is cheap and counting is not
None of this requires anyone to be careless. Confidence comes from experience and from the tools a team already has. A security team that has run cloud posture checks for years has every reason to believe it sees the cloud. An identity team that manages service accounts has every reason to believe it sees service accounts. What neither team has is a place where an agent, a model, a tool, a credential and an owner appear on one line. Each team sees its own slice well and assumes the union is covered.
A captain sails with a chart last updated in 2022. The coastline on the chart is accurate where it was surveyed. New sandbanks have formed since. The captain is not careless. The captain is confident, and the chart gives no sign of what it leaves out. A census is the new survey of the coast.
The cost of a wrong belief is paid later
A wrong belief about visibility has a particular economic property. It costs nothing until the day it matters, and then it costs a great deal. The Cloud Security Alliance found that 65% of organizations had an agent incident in the past year. IBM’s breach research, which we cover in Chapter 5, found that high levels of shadow AI added about US$670,000 to the average breach. Both are payments on a debt that the confident belief allowed to accumulate.
Visibility is the one area of security where being sure is a risk factor. A leader who says ‘I do not know yet’ will ask for a count. A leader who says ‘we have it covered’ will not.
- Ask each team that claims visibility one question. Can you show me the list, and when did it last change.
- Separate belief from evidence in your own reporting. Report the number of agents you can name, the number you have found since the last count, and the number with a named owner.
- Treat a rising count as a sign of progress. The first census usually finds more than anyone expected.
- Cyber insurers adding agent inventory questions to renewal forms
- Audit firms asking for the date of the last agent count
- Boards asking for the number of agents alongside the number of employees
The shadow estate
They gather where building is easy.
A shadow agent is one that runs without the knowledge of IT or security. Nobody has to be reckless to make one. It takes a person with a problem, a login and an afternoon.
The Cloud Security Alliance, in a study with Token Security published in April 2026, asked where shadow agents were most commonly found. Internal automation environments led with 51%, and LLM platforms followed with 47%. Respondents could choose more than one. Both are places where a single person can ship something useful with very little ceremony, which is also why they are valuable.
The surrounding evidence points the same way. Gartner’s survey of 302 cybersecurity leaders, run from March to May 2025, found that 69% suspect employees are using prohibited public generative AI tools. LayerX’s browser telemetry in 2025 found that 71% of generative AI connections from enterprise browsers used personal accounts, and that 58% of the connections made with corporate accounts bypassed single sign-on. In 1Password’s 2026 research, 82% of CIOs said employees create agents faster than IT can govern them.
The most useful single dataset on this comes from EY. In its Responsible AI Pulse survey of 975 C-suite leaders at companies above US$1 billion in revenue, two thirds said their organizations allow employees to build or deploy agents in some form. Only 60% of those have formal, organization-wide policies for it, and half said they lack a high level of visibility into how employees use agents.
The clock that runs while nobody looks
Detection time is the part of the shadow problem that people find hardest to believe. Reco’s 2025 research found a median of 403 days before unauthorized AI tools were detected in the organizations it studied. That is a median, which means half took longer.
Four quarter-end closes fit inside that window, along with a product launch, a performance review cycle and, in many organizations, a reorganization. By the time an agent is found it has a history. It has been trusted by colleagues and wired into other work. The longer it runs, the more the question shifts from ‘what is this’ to ‘what will break if we switch it off’, and the harder it is to answer.
If a review does not look for these tools, the median 403 day run passes through every review in between. Arithmetic by Lyzr Research on the Reco median.
The estate is also being rebuilt from underneath
Okta reported from its own platform data, covering November 2024 to October 2025, a 650% year-over-year increase in centrally managed service accounts, and access requests per company that were up 1,140% over two years. Those are the plumbing of automation. Agents run on them. Whatever you believe about agent counts, the number of non-human things with access is rising steeply.
The builder who moved on
A finance operations analyst builds a small agent that reads supplier emails, matches them to purchase orders and drafts payment holds. It works so well that three colleagues start to depend on it. A year later the analyst moves to another team. The agent keeps running on the analyst’s personal API key and on a shared mailbox login. Nobody is told. When the key is rotated for an unrelated reason, payment holds stop, a quarter-end close is delayed, and it takes two days to find out why.
Nothing about this story needs a malicious person. It needs a missing record. A register entry with a named owner and a review date would have flagged the move on the day the analyst changed teams.
Shadow agents are a byproduct of useful work done fast. The remedy that fits is a lightweight way to declare an agent, with the benefit that declared agents are supported and undeclared ones are not.
- Make declaring an agent easier than hiding one. A short form, a named owner and a review date is enough to start.
- Offer something in exchange for registering. Support, a sanctioned model, a budget line or a security check that saves the builder time.
- Scan the three places the evidence points to first. Internal automation, LLM platforms and personal-account connections.
- Detection times shrinking in vendor telemetry as scans become routine
- Personal-account use of AI tools declining as sanctioned alternatives spread
- Employee-built agents appearing in internal audit plans
The owners
Agents get a badge and skip HR.
Consider what a company does for a new employee. A role is written down. Access is granted for the role. Someone reviews that access at intervals. When the person leaves, there is an exit process that removes the badge. Each step has an owner.
Now compare how many organizations run a similar process for agents. The Cloud Security Alliance and Token Security asked.
Organizations are reasonable at the first two steps and weak at the rest. Only 21% have a formal decommissioning process and 19% are confident they fully retire agents. In Okta’s research, 32% say they secure agents with the same rigor they apply to human employees, and 10% say they have a mature strategy for managing them. SpyCloud found that 56% of AI privileges have formal governance and ownership, while 41% rely on informal processes or partial ownership.
An agent that nobody retires keeps its credentials and its access. It is a former employee whose badge still opens the door.
The crowd is larger than the list
The ownership problem is made bigger by the sheer number of machine identities that agents join. Entro Labs measured 144 non-human identities for every human in cloud-native environments in the first half of 2025, up from 92 a year earlier. Rubrik’s research puts the broader enterprise figure at about 45 to 1. These are service accounts, API keys, tokens and certificates, and agents are the newest members of that population.
At those ratios a 10,000 person company has somewhere between 450,000 and 1.4 million machine identities. Nobody reviews that many things by hand. Permiso’s survey adds a corollary. 69% of respondents said between 11% and 50% of their credentials are expired or unused but still active in production. Applied to 10,000 non-human identities, that range is 1,100 to 5,000 credentials that could be used and that nobody is using. Permiso also found that 95% say AI systems could create or modify identities without traditional human oversight, and 91% expect AI-generated identities to grow in 2026.
The Cloud Security Alliance found that 68% of respondents cannot reliably distinguish agent actions from human ones. Gartner’s Aaron Lord put the remedy in one line in April 2026. Authentication and authorization for agents should be tailored to them, not inherited from human user roles.
For every agent that can act on production systems, ask whether you could name the person to call at 2 a.m. If the honest answer is no, the agent has no owner, whatever the register says.
Arithmetic by Lyzr Research on published ratios. Your own ratio will differ, which is the reason to count.
A mental model for the whole estate
Every agent in your estate sits in one of four rooms, decided by two questions. Is it counted, and does it have an owner?
- Add one field to every agent record, a named person. Not a team, not a mailbox, not a project code.
- Tie agent review to events people already run. Reorganizations, role changes, exits and quarter-end closes are all natural moments to ask who now owns an agent.
- Write the exit step first. If you cannot say how an agent is retired, do not allow it into production.
- Identity vendors adding first-class agent identities
- Retirement of agents appearing in HR and IAM runbooks
- Per-agent audit trails becoming a default in agent platforms
The bill
The debt comes due in three places.
A wrong belief about visibility costs nothing until it matters. Then it shows up as an incident, a breach premium or a project that gets canceled. The data from the past year shows all three.
Incidents are common and mostly quiet
In the Cloud Security Alliance study with Token Security, 65% of organizations had an agent incident in the past year. Of those, 61% reported data exposure and 43% reported operational disruption. The Cloud Security Alliance separately reported that 53% had agents exceed their intended permissions.
Identity is where many of these incidents begin. SpyCloud’s 2026 survey of 750 cybersecurity leaders found that compromised non-human identities were the leading entry point into the enterprise, at 31%, ahead of phishing and social engineering at 17%. 68% of organizations had an identity-based event and those that did averaged eight of them. The same survey found that 91% of organizations use AI tools or agents with access to internal systems, and that only 36% monitor AI and non-human identity exposure.
Losses are nearly universal
EY’s Responsible AI Pulse is the broadest look at what AI risk has cost. In a survey of 975 C-suite leaders at companies with revenue above US$1 billion across 21 countries, 99% reported financial losses from AI-related risks. Nearly two thirds, 64%, reported losses above US$1 million. EY conservatively estimates the average loss among companies that experienced risks at US$4.4 million. The most commonly reported risks were non-compliance with AI regulations at 57%, a negative effect on sustainability goals at 55%, and biased outputs at 53%.
Shadow AI adds a premium to every breach
IBM’s 2025 Cost of a Data Breach report, covering 600 breached organizations, found that high levels of shadow AI added about US$670,000 to the average breach cost of US$4.63 million. Of the organizations that reported breaches of AI models or applications, 13% of the sample, 97% lacked proper AI access controls.
The trend line points up
Gartner’s April 2026 predictions put numbers on the direction. By 2028, 25% of enterprise GenAI applications will experience at least five minor security incidents per year, up from 9% in 2025. By 2029, 15% will experience at least one major incident per year, up from 3% in 2025. Gartner’s Aaron Lord attributes part of the rise to the way the Model Context Protocol was designed, with interoperability and ease of use first.
The readiness gap
Accenture’s State of Cybersecurity Resilience 2025 surveyed 2,286 executives, 80% of them CISOs and 20% CIOs, at organizations with revenue above US$1 billion in 17 countries. It found that 90% of organizations are not adequately prepared to secure an AI-driven future. 63% fall in what Accenture calls the Exposed Zone, 27% in the Progressing Zone, and 10% in the Reinvention Ready Zone. 77% lack essential data and AI security practices, and only 22% have clear policies and training for generative AI use.
And projects that never ship
Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating cost, unclear business value and inadequate risk controls. Two of those three reasons are estate problems. Cost and risk controls both depend on knowing what is running. IDC’s Asia/Pacific FutureScape reaches a related view, predicting that by 2027 half of AI applications and services will fail to progress beyond proof of concept.
Money is following the problem. Gartner expects AI governance spending to reach US$492 million in 2026 and surpass US$1 billion by 2030, and expects more than 40% of enterprises to have had a shadow AI incident by 2030. They are predictions, and we treat them as such. Even if they are only roughly right, the budget for this work is arriving.
The invoice nobody could explain
A finance team reviewing cloud and model spend at the end of a quarter notices a model provider invoice that has doubled. The invoice is paid by a corporate card owned by a team that has since been reorganized. Three internal tools call the provider. One of them has no owner on record. The team spends a week tracing usage back to a prototype that was never shut down and that loops on a retry whenever an upstream service is slow.
The cost is small compared with a breach. The lesson is not small. Spend attribution and security attribution rely on the same record.
The cost of the gap does not arrive as one event. It arrives as incidents, premiums, audit findings and projects that stall. All of them are easier to explain when there is a count.
- Put a number on the cost of not knowing. Take your last breach exercise or incident review and ask which agents were involved, who owned them, and how long it took to find them.
- Ask finance for model and API spend by cost center and by named agent. The gap between those two views is your shadow estate in dollars.
- Report the three counts leadership should see each quarter. Agents named, agents owned, agents retired.
- Cyber insurance exclusions or premiums tied to AI governance
- Breach reports that cite unknown or unowned agents
- Finance and FinOps tools adding per-agent cost views
The rulebook
Every rulebook begins with a list.
It is easy to read the AI governance landscape as a pile of unrelated demands. Read closely, nearly every one of them quietly assumes the same first step.
The EU AI Act clock
The dates below are as reported in July 2026 by the Data Protection Report, and they have moved before. Duties for general-purpose AI models have applied since August 2, 2025. Article 50 transparency duties apply from August 2, 2026, with a grace period on marking until December 2, 2026. High-risk systems listed in Annex III have been delayed to December 2, 2027, and those in Annex I, covering products already regulated elsewhere, to August 2, 2028. One secondary note we reviewed gave different high-risk dates, and we used the Data Protection Report. Confirm the current text with counsel.
The lesson in this calendar is not any single date. It is that you cannot tell which dates apply to you without a list of the systems you run and what each one does. Obligations attach to systems, not to organizations.
NIST asks for it by name
The NIST AI Risk Management Framework asks, under GOVERN 1.6, for mechanisms to inventory AI systems, in accordance with organizational risk priorities. It is one sentence in a long document, and many other controls in the framework depend on it.
OWASP now has a list for agents
The OWASP Top 10 for Agentic Applications, published on December 12, 2025, names ten risks. Reading them with an operator’s eye, nearly every one is easier to defend against if you have a register. The table pairs each risk with what the register gives you. The pairing is our analysis.
| Risk | What it means | What the register gives you |
|---|---|---|
| ASI01 Agent Goal Hijack | Manipulated inputs redirect an agent’s objective | Which agents read untrusted content |
| ASI02 Tool Misuse and Exploitation | Legitimate tools used unsafely | The tool list for each agent, with limits |
| ASI03 Identity and Privilege Abuse | Inherited or misused credentials | Each agent’s own identity and scope |
| ASI04 Agentic Supply Chain Vulnerabilities | Tampered models, tools or third-party agents | Model and tool versions in use |
| ASI05 Unexpected Code Execution | Agents manipulated into running code | Which agents can execute code |
| ASI06 Memory and Context Poisoning | Corrupted stored memory biases future behavior | Which agents keep memory, and where |
| ASI07 Insecure Inter-Agent Communication | Unauthenticated or altered agent messages | A map of which agents talk to which |
| ASI08 Cascading Failures | One fault spreads across agents | Dependencies between agents |
| ASI09 Human-Agent Trust Exploitation | Abuse of confidence in agents | An owner and an approval path |
| ASI10 Rogue Agents | Drift while appearing authorized | A baseline and an owner to compare against |
MCP and the three-circle no-go zone
The Model Context Protocol has become a common way for agents to reach tools and data. Gartner’s April 2026 guidance for software engineering leaders is specific. Create a formal security review for MCP use cases, prioritize low-risk patterns, exclude high-risk combinations, and establish domain-oriented ownership of MCP servers, so that the people who know the data predefine the guardrails before an MCP client touches it.
Gartner also offers a simple filter. Any use case that combines three factors, an agent with access to sensitive data, exposure to untrusted content and a way to communicate externally, should be treated as a no-go zone because of the risk of data exfiltration.
MCP was built for interoperability, ease of use and flexibility first.Aaron Lord, Sr Director Analyst, Gartner, April 2026
The register makes the filter usable. Three fields, what the agent can read, what content it ingests and where it can send data, are enough to test any agent against the circle. Without them the filter is an idea. With them it is a query.
Six questions an auditor will ask
| The question | What a good answer looks like | |
|---|---|---|
| 01 | How many are there | A number and the date it was true |
| 02 | Who owns each one | A name, in the register |
| 03 | What data can each reach | A scope per agent, not per team |
| 04 | What can each one do | Actions and limits, written down |
| 05 | How do you know it is working | Logs, reviews and tests with dates |
| 06 | How is one retired | A process and a recent example |
The auditor’s first question
A mid-sized financial services firm is preparing for a review of its AI controls. The auditor skips the policy documents and starts with one request. A list of every AI agent in production, with the owner of each and the date the list was last reconciled. The firm’s security team has a list from a previous project. The business units have three others. Nobody can say which is right. The review continues, and the list becomes the first finding.
- Map your agents to the EU timetable using the register. Tag each one with the use case it serves and whether it touches a regulated function.
- Add the three MCP fields to the register, what it can read, what untrusted content it ingests and where it can send data, and run the filter.
- Assign an owner to every MCP server, preferably in the domain that owns the data.
- Further changes to the EU AI Act dates, and the first enforcement actions
- Auditors and insurers adopting the OWASP agentic risks as checklists
- MCP security features landing in gateways and registries
The census
One card per agent, five ways of looking.
The method has four parts. A definition of what counts. A register that holds one record per agent. A set of lenses that find agents. And a loop that repeats the exercise and compares the counts.
The register, in nine fields
The fields below are enough to start. Each comes with a test that it passes, so a reviewer can tell a complete record from a hopeful one. Add fields as your risk appetite requires, and see Appendix B for a template.
| Field | What to record | The test it passes |
|---|---|---|
| Name | A stable name and where it runs | Two people mean the same agent |
| Owner | One named person | They can be reached today |
| Purpose | One sentence | A reviewer could judge it |
| Model | Provider, model and version | A retirement notice can be matched |
| Data | What it can read and write | Scope is narrower than the builder’s |
| Tools | What it can call or change, including MCP servers | Each tool has a limit |
| Identity | Its own credentials | Logs show it as itself |
| Limit | Cap on spend and actions | A mistake has a ceiling |
| Status | Live, paused or retired, with a review date | Nothing goes unreviewed |
Five lenses, one list
No single source finds every agent. Each lens sees a different part of the estate and each has a blind spot. Use five and merge the results into one list. Where the sources disagree, you have found something worth investigating.
| Lens | What it finds | What it tends to miss |
|---|---|---|
| Cloud accounts | Agents running on cloud infrastructure | Assistants inside SaaS products |
| SaaS platforms | Built-in assistants and connections | Code that runs elsewhere |
| Code repositories | Agent code and embedded keys | No-code builds |
| Gateway logs | Live model traffic | Calls that bypass the gateway |
| Ask every team | Intent and purpose | What people have forgotten |
Eight rules for a register that survives contact with reality
If a record needs a paragraph, split it. A register is a database before it is a document.
A person who can be reached today, not a team or a distribution list.
Record whether an agent was declared by its builder or found by a scan. The difference is your shadow rate.
Every field has a last confirmed date. Old data is a risk that looks like information.
Where two lenses disagree, log it. The disagreement is the work.
Start by observing. Add enforcement once people trust the list.
Pick a business unit, finish it, then widen. A finished partial census beats an abandoned complete one.
A first census is a photograph. A second is a film.
What comes after the census
Once the register exists, automation becomes possible. Gartner’s guardian agent concept is the clearest description. It expects guardian agents to make up 10% to 15% of the agentic AI market by 2030, in three roles. Reviewers check output for accuracy and acceptable use. Monitors track actions for follow-up. Protectors adjust or block actions automatically. Each role needs a list of agents to oversee, a scope for each and an owner to escalate to. The register is that list.
A first census is a photograph. A second one is a film. The photograph shows what you have. The film shows whether the estate is growing faster than your ability to govern it.
Counting what you cannot see
A census has an odd property. You can estimate how many agents you have not found, using the same idea ecologists use to count animals they cannot see all at once. Mark the animals you catch, release them, catch a second sample and see how many were already marked. The overlap tells you how complete each sample was.
Apply it to lenses. Suppose lens A, a cloud scan, finds 40 agents and lens B, a team survey, finds 30, and 12 turn up in both. The estimate of the total is the product of the two counts divided by the overlap, 40 times 30 over 12, or 100 agents. You have found 58 and can expect about 42 more.
A fisheries biologist who tags 50 fish and later catches 100 in which 10 are tagged estimates a population of 500. She never counted 500 fish. She counted two samples and the overlap. The overlap between two scans of your estate says the same thing about agents. A small overlap means a large unseen estate.
A maturity ladder
We use five steps. They are ordered, and each makes the next possible. In our reading, most organizations sit at the first or second step. Do not skip to enforcement before you can count.
A ten-question self-assessment
Score one point for each statement you could defend to an auditor today, with a document you could hand over.
Tick each statement you could defend to an auditor today, with a document you could hand over.
Step 1 or 2 on the ladder. Start with the definition and a first scan.
Step 3 or 4. You can count and are building ownership. Finish the owners and run the second census.
Step 4 or 5. You are close to enforcement. Check that you can prove it works.
The first ninety days
The whole method fits in a quarter and needs no new platform. A definition, a register, five lenses, a named owner on each record, and a second count. Everything beyond that builds on it.
- Registries and gateways adding agent discovery as a default feature
- A standard schema for an agent record emerging from standards bodies
- Procurement asking vendors for the agent register fields as a contractual deliverable
Guidance by role
One register, six users.
The register serves several masters. It is the record a security team defends, a finance team attributes cost against, a lawyer maps obligations to, and a board asks about. Each role asks it a different question.
Chief Information Officer
Chief Information Security Officer
Chief Financial Officer
General Counsel
Head of AI or Engineering
The Board or Audit Committee
A single page for the board is enough. It should show three numbers, their dates and the change since the last quarter. If the count of agents is rising, ask whether the count of owners is rising at the same rate.
Ask whether the count of owners is rising at the same rate as the count of agents.
What comes next
Five calls for 2027, held loosely.
Research houses make predictions because planners need something to plan against. We offer ours in the same spirit, with the confidence we hold and the evidence that would change our minds. Where a call overlaps with a published prediction we cite it.
By the end of 2027, the agent register is a standard audit item.
Why we think so. Every framework and every regulator asks what you run. IDC and Gartner both expect agent counts to rise sharply, which makes an audit question about counts inevitable. Auditors will ask for the count before they ask for anything else.
What would prove us wrong. Platforms ship a built-in inventory that auditors accept, which would make a separate register unnecessary.
Watch. Audit requests that cite agent counts, insurer questionnaires, platform inventory features.
The first full census of a large estate finds more agents than any survey figure, and the board is surprised.
Why we think so. The surveys already show most organizations finding shadow agents. A first scan across clouds and SaaS finds more than any survey says, because surveys ask people and scans ask systems.
What would prove us wrong. Building agents becomes tightly gated by default, which would shrink the shadow.
Watch. Scan results published in aggregate, breach reports citing unknown agents, vendor gating features.
Retirement becomes a named control.
Why we think so. Only one in five organizations has a formal process, and old agents keep their access. It is a cheap control and an obvious audit finding.
What would prove us wrong. Vendors expire agent credentials automatically, so a manual process matters less.
Watch. Auto-expiry in identity tools, retirement language in frameworks, incidents traced to old agents.
Every production agent carries its own identity.
Why we think so. Mixed logs are the barrier to every investigation. 68% of respondents cannot tell agent actions from human ones, and Gartner already advises tailored authentication for agents.
What would prove us wrong. Tool-level audit trails become good enough that identity per agent is less urgent.
Watch. Identity vendors adding agent identities, logs that tag agent actions, incident reports naming the gap.
Spend follows the count, and finance asks for it.
Why we think so. Model spend lands on invoices with no owner. Once a register exists, finance will want every dollar mapped to a name. IDC expects the cost of an agent action to fall 87% by 2029, which makes unattributed volume easy to ignore and expensive in aggregate.
What would prove us wrong. Model prices keep falling steeply, easing the pressure to attribute spend.
Watch. FinOps tools adding per-agent views, budgets set per agent, price trends.
Where Lyzr fits
This is a vendor view and should be read as one. Lyzr builds infrastructure for enterprise AI agents. Opencontroller, its governance product, reads across clouds and platforms to find agents, records an owner for each, and can sit in the request path to stop calls that break policy or budget. It enforces rules you write. It does not write them for you. In one enterprise scan it found 227 shadow AI assets. That is one company and one scan, not an average, and we do not offer it as a benchmark. A broader agent census is on the Lyzr roadmap.
The method in this paper does not depend on any vendor. A definition, a register, five lenses and a named owner can be built with a spreadsheet and a few weeks of attention. Tools make it faster, repeatable and enforceable, and the first census is worth doing before you buy one.
227 shadow AI assets were found by a Lyzr Opencontroller scan of one enterprise in 2026. We include it as an example of what a first scan can turn up, not as a figure to plan against.
Count first.
There is a plain kind of courage in counting. It means being willing to learn that the number is larger than you said, that some of what you run belongs to nobody, and that the confident answer you gave last quarter was a guess.
Organizations that make that leap tend to find that the fear was larger than the finding. The agents were not hostile. They were useful, quick and unlabeled. Once labeled, they can be supported, secured, budgeted and, when their time comes, retired.
A census is not a project that ends. It is the first act of an organization that intends to stay in charge of the things it builds. Start with one business unit. Write the definition on a page. Put a name on every record. Do it again in ninety days.
The estate grows whether or not you count it. Only one of those choices lets you say how large it is.
The evidence base, and how much weight each source can bear.
Not all surveys are equal. We show the sample, the date and who ran each one, so a reader can judge. Vendor-run surveys are useful and are also run by firms with products to sell. We treat their figures as indications, and we have said so in the text.
| Source | Type | Sample and date | Note |
|---|---|---|---|
| Gartner CIO poll | Webinar poll | 147 CIOs and IT leaders, May 19, 2025 | Self-selected audience. Directional |
| Gartner cybersecurity leaders survey | Analyst survey | 302 leaders, Mar to May 2025, published Nov 2025 | Prohibited GenAI use, 69% suspected |
| Gartner press releases | Analyst predictions | Jun 2025, Jun 2025, Apr 2026 | Predictions, not measurements |
| IDC FutureScape 2026 and Agent Economics | Analyst predictions | Published Oct to Dec 2025 | Predictions. The APAC deck adds survey sample sizes |
| EY Responsible AI Pulse | Consultancy survey | 975 C-suite leaders, 21 countries, revenue above US$1B, 2025 | Release gives two survey windows |
| Accenture State of Cybersecurity Resilience | Consultancy survey | 2,286 executives, 17 countries, Oct to Dec 2024 | Correlational findings |
| a16z enterprise CIO survey | Investor survey | 100 CIOs, 15 industries, as of May 8, 2025 | Method detail limited |
| Cloud Security Alliance with Token Security | Vendor-assisted survey | Sample not stated, April 2026 | Treat as indicative |
| SpyCloud Identity Threat Report | Vendor survey | 750 security leaders, Sept 2026 | Field dates not stated |
| Permiso State of Identity Security | Vendor survey | 500+ professionals, Jan 2026 | Field dates not stated |
| Okta Businesses at Work 2026 | Vendor research | 261 executives and platform data, 2025 | Survey date not stated |
| IBM Cost of a Data Breach 2025 | Industry research | 600 breached organizations | Widely used benchmark |
| Entro Labs, Rubrik, Reco, LayerX, 1Password | Vendor telemetry and research | Various, 2025 to 2026 | Cited via the Cloud Security Alliance. Secondary |
How we handled conflicts
Where sources conflicted we used the one with the clearer method and said so. On the EU AI Act we used the Data Protection Report dated July 2026 over a secondary note that gave different high-risk dates. EY’s release gives two different survey windows and two different counts of roles, and we have reported its figures without relying on those details. Alatirok and the Cloud Security Alliance research notes are secondary sources for figures that originated with IBM, OutSystems, Gartner, Reco, LayerX, 1Password, Entro Labs and Rubrik. Where possible we traced these to primary releases, and where we could not, we label them.
What we calculated ourselves
Four figures in this paper are our own arithmetic from published numbers. The minimum overlap of 50 out of 100 between those confident in visibility and those surprised by shadow agents. The 26 and 40 out of 100 split of large companies that allow employee-built agents. The one-seventh ratio of the shadow AI premium to the average breach. The 450,000 and 1.44 million machine identities for a company of 10,000 people, and the 1,100 to 5,000 credentials range. Each is shown next to its inputs.
An agent register template.
Copy this into a spreadsheet, one row per agent. The example row is illustrative.
| Field | Example entry | Notes |
|---|---|---|
| Agent ID | AGT-0142 | Stable, never reused |
| Name | Supplier email matcher | What people call it |
| Where it runs | Internal automation platform, finance workspace | Environment and account |
| Owner | A named person in finance operations | Reachable today |
| Backup owner | A named person | For leave and role changes |
| Purpose | Matches supplier emails to purchase orders and drafts payment holds | One sentence |
| Model and version | Provider, model name, version | So a retirement notice can be matched |
| Data it can read | Supplier mailbox, purchase order table | Sensitivity label |
| Data it can write | Draft payment holds | Systems of record touched |
| Tools and MCP servers | Mailbox connector, ERP connector | Each with a limit |
| Reads untrusted content | Yes, external email | For the three-factor filter |
| Can send data externally | No | For the three-factor filter |
| Identity | Its own service account, not a person’s key | Credential type and rotation date |
| Spend or action limit | Monthly spend cap and daily action cap | A mistake has a ceiling |
| Declared or discovered | Discovered | Your shadow rate |
| Status and review date | Live, next review in 90 days | Nothing goes unreviewed |
| Retirement plan | Disable the account, revoke tokens, archive logs | Written before production |
A glossary
| Term | Meaning in this paper |
|---|---|
| Agent | Software that can decide and act toward a goal with some autonomy, using tools, data or other systems. Include embedded assistants if your definition says so |
| AI estate | Everything of this kind that runs in or for the organization, whether declared or not |
| Shadow agent | An agent running without the knowledge of IT or security |
| Non-human identity (NHI) | A credential, key, token, certificate or account that belongs to software and not to a person |
| Register | The list with one record per agent and a named owner |
| Census | The act of finding agents and filling the register, repeated on a cadence |
| MCP | Model Context Protocol, a common way for agents to reach tools and data |
| Guardian agent | In Gartner’s term, an agent that reviews, monitors or protects other agents |
| Orphaned agent | An agent that was counted but has lost its owner |
Sources.
- Cloud Security Alliance with Token Security, Autonomous but Not Controlled, April 20, 2026. Sample size not stated. cloudsecurityalliance.org/artifacts/autonomous-but-not-controlled-ai-agent-incidents-now-common-in-enterprises
- Cloud Security Alliance AI Safety Initiative, Shadow AI Apps research note, May 30, 2026. Source of the cited Gartner, IBM, Reco, LayerX, 1Password and Entro figures. labs.cloudsecurityalliance.org/research/csa-research-note-shadow-ai-apps-enterprise-20260530-csa-sty/
- Alatirok, How Many AI Agents Per Enterprise in 2026. Source of the cited OutSystems, IBM Think, Rubrik and survey figures. A secondary source. alatirok.com/how-many-ai-agents-per-enterprise-2026/
- Gartner, Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release, June 25, 2025. gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- Gartner, Guardian Agents will Capture 10 to 15% of the Agentic AI Market by 2030, press release, June 11, 2025. gartner.com/en/newsroom/press-releases/2025-06-11-gartner-predicts-that-guardian-agents-will-capture-10-15-percent-of-the-agentic-ai-market-by-2030
- Gartner, 25% of Enterprise GenAI Applications Will Experience at Least Five Minor Security Incidents per Year by 2028, press release, April 9, 2026. gartner.com/en/newsroom/press-releases/2026-04-09-gartner-predicts-25-percent-of-all-enterprise-gen-ai-applications-will-experience-at-least-five-minor-security-incidents-per-year-by-2028
- IDC, Agent Adoption, The IT Industry’s Next Great Inflection Point, Rick Villars, December 10, 2025. idc.com/resource-center/blog/agent-adoption-the-it-industrys-next-great-inflection-point/
- IDC, FutureScape 2026, Charting the Agentic Future, Asia/Pacific excerpt, November 2025. info.idc.com/rs/081-ATC-910/images/IDC-futurescape-2026-charting-the-agentic-future-excerpt-AP.pdf
- EY, Companies advancing responsible AI governance linked to better business outcomes, Responsible AI Pulse survey, October 2025. ey.com/en_gl/newsroom/2025/10/ey-survey-companies-advancing-responsible-ai-governance-linked-to-better-business-outcomes
- Accenture, Only One in 10 Organizations Globally Are Ready to Protect Against AI Augmented Cyber Threats, June 26, 2025. newsroom.accenture.com/news/2025/only-one-in-10-organizations-globally-are-ready-to-protect-against-ai-augmented-cyber-threats
- a16z, How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025. a16z.com/ai-enterprise-2025/
- SpyCloud, 2026 Identity Threat Report Finds Non-Human Identities Are Now the Leading Path into the Enterprise, September 9, 2026. spycloud.com/newsroom/spycloud-2026-identity-threat-report-finds-non-human-identities/
- Permiso, State of Identity Security 2026, January 21, 2026. permiso.io/blog/permiso-state-of-identity-security-2026-survey-report
- Okta, Businesses at Work 2026, April 30, 2026. okta.com/newsroom/articles/businesses-at-work-2026/
- IBM, Cost of a Data Breach Report 2025, as cited by the Cloud Security Alliance.
- Data Protection Report, EU AI Act update, July 2026, for application dates.
- NIST, AI Risk Management Framework 1.0, GOVERN 1.6. nist.gov/itl/ai-risk-management-framework
- OWASP, Top 10 for Agentic Applications 2026, published December 12, 2025, as summarized by the Non-Human Identity Management Group. nhimg.org/complete-guide-to-the-2026-owasp-top-10-risks-for-agentic-applications
- Lyzr Opencontroller scan of one enterprise, anonymized, 2026. One data point, not an average.
Infrastructure for the agents you will have to govern.
Lyzr builds infrastructure for enterprise AI agents. Lyzr Research publishes evidence-led work for the leaders who have to run, secure and govern them. This paper is part of that series.
The views here are those of Lyzr Research. They are offered to help leaders reason about their estates, and they are not legal, financial or security advice. Confirm regulatory dates and obligations with counsel.
LYZR.AI