The Agentic Enterprise AK · Wednesday Edition · 7 min read | Wednesday, September 2, 2026 Governance is the gating item now. SAP and ServiceNow shipped control planes to inventory every agent an enterprise runs, and both call governance a board-level issue. Microsoft, Okta, Thoughtworks, and a wave of startups are selling the same leash. The survey data underneath explains why. Yesterday the story was that a model is becoming the front door to enterprise software. This is the other half of that trade. When you talk to your systems and they act on your behalf, what you deploy is delegated authority, and right now most of it runs without an inventory, an owner, or an audit trail. SAP put a name on the problem this week, agent sprawl, and a price on the fix. The question for a CIO is no longer whether to adopt agents. It is whether you can say, on any given Tuesday, exactly what all of yours are doing. | The vendors that sold you agents are now selling the leash | S | omething predictable arrived on schedule this week. For two years enterprises raced agents into production, and the counts are now hard to look away from. IBM's latest survey puts most large companies on track for a digital workforce of more than 1,600 AI agents by the end of 2026, and Gartner projects the average global Fortune 500 will run north of 150,000 by 2028. Adoption is settled. Ninety-six percent of organizations report using agents in some capacity. The question that replaced it is harder and less flattering: who governs the ones you already have? |
SAP answered by shipping a product. It launched the AI Agent Hub, a governance layer housed in LeanIX that inventories, monitors, and retires both SAP and non-SAP agents, with agent identity management through SAP Cloud Identity Services and session-level observability tied to business KPIs. The framing was blunt: agent governance is now a board-level issue, not an IT one. SAP is not alone. At Knowledge 2026 ServiceNow expanded its AI Control Tower to discover, govern, observe, secure, and measure every agent, model, and workflow across the enterprise regardless of origin, Microsoft named launch partners for its Agent 365 governance stack, Thoughtworks shipped Agent/works as a governed runtime that runs on any cloud, and a wave of startups, AccuKnox's AgentZ and Aziro's Aziron among them, arrived promising sandboxes, role-based access, and audit traces. The category has a name now. It is agent sprawl, which is what you get when agents are created faster than the organization can say who owns them. The reason the vendors moved is the survey data underneath. Seven in ten executives say the AI governance they have today is not fit for purpose. Only 18 percent maintain a complete inventory of the agents running inside their walls, and just 12 percent have a centralized platform to manage them. The exposure is already real: in a 2026 survey of security teams, 88 percent of organizations reported at least one agent-related security incident in the past year, most often data leakage or manipulation through untrusted inputs. Mean monitoring coverage sits near half, which means roughly one in two production agents runs unwatched. A forgotten SaaS subscription leaks. A forgotten agent with live credentials acts. |
There is a pattern here worth naming for anyone reading this in a CIO's office. Enterprises adopted agents the way they adopted SaaS a decade ago: bottom-up, team by team, faster than central IT could inventory. SaaS sprawl gave us shadow IT and a cleanup that ran for years. Agent sprawl is the same story with higher stakes, because an agent does not just hold data, it takes action. It moves money, edits systems of record, sends messages, and calls other systems on its own authority. The Spearhead Take Buy the control plane if it helps you see the sprawl, but do not confuse visibility with governance. Before you approve the next agent, answer three questions in writing: who owns it, what can it touch, and how would you know if it misbehaved. If you cannot answer all three, you do not have an agent, you have a liability with a login. Start with the inventory, because you cannot govern what you cannot see, and make named ownership a condition of deployment rather than a cleanup task for next year. And notice who is selling you the leash. The platform that makes it trivial to spin up a hundred agents is often the one now offering to count them for free, which is how the switching cost accrues quietly. The enterprises that win the agent era will not be the ones with the most agents. They will be the ones that can tell you exactly what all of theirs are doing. |
| The Obvious & The OverlookedWhat everyone sees, and what they miss. The Obvious Adoption is settled; governance is the fight. Ninety-six percent of organizations run agents, and the open question has moved to who controls them. innobu Vendors are productizing governance. SAP's AI Agent Hub, ServiceNow's AI Control Tower, and Microsoft's Agent 365 all shipped or expanded as control planes this week. ServiceNow The problem finally has a name. SAP put "agent sprawl" on the record, usually the sign a problem has become too expensive to ignore. Yahoo Finance | The Overlooked The gap is accountability, not tooling. Only 7.2 percent of organizations have named a single person responsible for what their agents do. Gravitee Half of production agents run unwatched. Mean monitoring coverage sits near 52 percent, so roughly one in two agents acts with no oversight. Gravitee Identity is becoming the control point. Okta and SailPoint now govern agents as non-human identities, the way they govern employees, and SAP's Hub runs on Cloud Identity Services. Okta The leash is sold by the hand that caused the sprawl. SAP folds governance into its platform at no extra charge, counting your agents and deepening the dependency at once. erp.today |
| Moving PiecesFive developments worth a CIO's attention. GovernanceSAP shipped a control plane for every agent you run, SAP or not SAP launched its AI Agent Hub this week, a governance layer housed in LeanIX that tracks, monitors, and retires agents across the enterprise, including non-SAP ones, with agent identity management via Cloud Identity Services and session-level observability tied to business KPIs. It targets general availability in the third quarter and is bundled into the Business AI Platform at no additional charge. The move positions SAP as the governance layer of record for the broader agent ecosystem. The signal for buyers is that governance has become a product you can purchase, and that the vendor offering to count your agents is frequently the one whose platform makes them easy to create. Useful, but price the dependency. GovernanceServiceNow pitches its Control Tower as the governance layer of record ServiceNow used Knowledge 2026 to push AI Control Tower into a system that discovers, governs, observes, secures, and measures every AI agent, model, and workflow across the enterprise regardless of origin, adding 30 integrations spanning AWS, Google Cloud, Azure, SAP, Oracle, and Workday. It deepened its Microsoft partnership so the Control Tower governs agents built in Copilot Studio and the Agent 365 ecosystem, and tied in NVIDIA to extend oversight to large model workloads through Project Arc and the Enterprise AI Factory design. The tell for buyers is that the governance layer is consolidating into the platforms that already sit at the center of enterprise workflow, ServiceNow and Microsoft chief among them. That is convenient, and it concentrates the control point in the same hands that hold the workflow. ResearchThe security data behind the governance rush is ugly The reason vendors are racing to sell control planes is that the numbers on unmanaged agents are bad. In a 2026 survey of security teams, 88 percent of organizations reported at least one agent-related security incident in the past year, most often data leakage or manipulation through untrusted inputs, while mean monitoring coverage sat near half. IBM projects most large enterprises will run more than 1,600 agents by year end. The gap between deployment velocity and oversight is the whole story. For any enterprise, the first governance deliverable is unglamorous and unavoidable: a current inventory of what is running and who owns it. PolicyEU enforcement against model providers is now live Since August 2 the European Commission's AI Office can act directly on general-purpose AI models: request documentation, run technical evaluations, demand risk mitigations, restrict or withdraw a model from the EU market, and levy fines up to 3 percent of global annual turnover or 15 million euros, whichever is higher. Its preferred opening move is a "technical compliance dialogue" rather than a formal proceeding. For enterprises, the division of labor matters. If you deploy a foundation model to EU users, the provider owns general-purpose compliance, but the logging, human oversight, and documentation are the parts you own, and the parts an agent that acts on its own makes harder to produce after the fact. DealsMicrosoft bundles Copilot into a sovereign-AI push with HUMAIN At LEAP 2026 on August 31, Microsoft and Saudi Arabia's HUMAIN expanded their partnership, pairing HUMAIN ONE with Microsoft 365 and Copilot into a new productivity bundle aimed at one million enterprise users across the Middle East and Africa, with HUMAIN's first enterprise AI PC due September 20. The enterprise read is that sovereign-AI programs, once positioned as alternatives to Western suites, are increasingly distributed on top of them. For multinationals in the region, the default productivity stack and the default AI assistant are converging into a single procurement decision, and with it, a single governance surface. | On the RadarEight signals, sharpened. | Leadership | John Ternus became Apple's CEO on September 1, with Tim Cook moving to executive chairman. Bloomberg frames the handoff around fixing Apple's lagging AI strategy ahead of an ambitious fall hardware slate. Bloomberg | | Product | Thoughtworks launched Agent/works, a control plane and governed runtime for enterprise agents on any cloud. It joins the same governance wave SAP and Microsoft moved on this week. Thoughtworks | | Governance | OpenAI will end Cursor's access to its models on November 12 after SpaceX bought its parent, Anysphere. Model supply is now a lever a vendor can pull over a corporate event that has nothing to do with you. CNBC | | Infrastructure | OpenAI detailed the PORTS-Pike data center in Ohio with SB Energy, Nvidia, and the Department of Energy. The projected 35,000 construction jobs are another sign the model buildout is now a heavy-infrastructure business. PR Newswire | | Governance | Okta made Okta for AI Agents generally available, a control plane to govern agents as non-human identities. It works across frameworks, hyperscalers, and SaaS, including Salesforce Agentforce and the ServiceNow AI Platform. Okta | | Deals | Nvidia finalized its 700 million dollar Run:ai acquisition after EU approval. Adding GPU-orchestration software previews the regulatory path a far larger Hugging Face deal would face. Barchart | | Product | GPT-5.6 is now the preferred model in Microsoft 365 Copilot. Wired into Word, Excel, and a persistent workspace, it is another sign the assistant is becoming the way people enter their software. Microsoft | | Policy | Georgia Power's 3.2-gigawatt contract to supply an OpenAI data center hit its regulatory deadline. It is a live test of who pays when AI load lands on the public grid. The Current |
| Quick HitsTwelve more, worth knowing. | Clay is raising a new round led by Wellington Management at a 7 billion dollar pre-money valuation, up from 5 billion in January, for its sales and marketing AI.Axios | | Databricks closed a 5 billion dollar round at a 190 billion dollar valuation with a run rate above 7 billion, funding Lakebase, Genie, and its AI cost-control gateway.Bloomberg | | DeepSeek reopened a roughly 8 billion dollar round at about a 74 billion dollar valuation after its cost-efficient V4 Flash model, and is exploring a STAR Market IPO.Bloomberg | | Instinct raised a 250 million dollar Series B at a 2.5 billion dollar valuation for AI assistants, the largest AI round of the week.Crunchbase | | Gatik raised 200 million dollars in Series D from the Qatar Investment Authority, Koch, Millennium, and ARK Invest for autonomous middle-mile trucking.Tech Startups | | AccuKnox launched AgentZ, a model-agnostic platform that moves agents to production with role-based access, runtime credentials, and audit traces.AI Agent Store | | Aziro launched Aziron, an enterprise agent execution platform bundling agents, workflows, documents, and enterprise controls in one governed environment.AI Agent Store | | Emerald AI raised to help data centers manage their power draw on strained electric grids.Tech Startups | | NRG is nearing a roughly 1.2-gigawatt hyperscaler power deal even as Texas pauses some new data-center interconnections.Utility Dive | | Andreessen Horowitz closed a 1.1 billion dollar "Machine Age Fund" for the physical backbone of AI, from chips to robotics.Today's Startup News | | J.P. Morgan estimates hyperscaler capital spending will reach about 697 billion dollars in 2026, one of the year's defining capital themes.J.P. Morgan | | JLL projects global data-center capacity could roughly double to about 200 gigawatts by 2030, potentially requiring up to 3 trillion dollars in new spending.PR Newswire |
| The Number7.2% of organizations have named a single person accountable for what their AI agents do Everyone else describes accountability as unclear, shared but undefined, or simply never discussed. Set that against the deployment numbers, ninety-six percent adoption and more than 1,600 agents per large enterprise by year end, and the mismatch is the whole story. We have handed action to software at scale and, in more than nine cases out of ten, no one owns the outcome by name. This is what makes the governance-tool boom worth watching skeptically. A control plane shows you the agents you failed to control. It does not decide who is responsible when one moves the wrong number. Governance does not start with a purchase. It starts with a name next to a system that can act. | Counter-SignalMarketsYou cannot buy your way out of agent sprawl The tidy read on this week is that the governance problem finally has a solution: SAP, ServiceNow, Microsoft, Okta, Thoughtworks, and a dozen startups will sell you a control plane that inventories, monitors, and retires your agents. The read underneath is that most of these vendors also sell the thing that creates the sprawl. SAP will govern your agents at no extra charge inside its Business AI Platform, which is another way of saying the platform that makes it trivial to spin up a hundred agents is also the one that will count them. That is not a scandal. It is the SaaS playbook run one layer up: give away the governance, own the platform, and let the switching cost accrue while everyone admires the dashboard. The uncomfortable part is that a dashboard does not govern anything. It shows you what you already failed to control. The 7.2 percent of companies with a named owner for agent behavior did not get there by buying software. They got there by making someone accountable, scoping what each agent can touch before it acts, and deciding in advance how they would know if it misbehaved. Tooling helps you see the sprawl. It does not carry the responsibility. Buy the control plane if it earns its keep, but keep your agents describable and portable across more than one vendor's governance layer, or you will have solved lock-in at the model tier by recreating it at the governance tier. The registry is a spreadsheet with a logo until a person's name sits next to every agent in it. | From the FieldYesterday's close asked you to make the agent show its work, every action, every time. This week the industry turned that sentence into a product category. I have spent much of the last month doing the least glamorous work in applied AI: counting. Not building agents, counting them. And the counting keeps turning up surprises, the reconciliation agent someone spun up in March that still holds write access to the ledger, the support bot quietly calling three internal APIs nobody documented. None of it malicious. All of it ungoverned. So I understand the appeal of a hub that promises to inventory the whole mess for you, and some of these tools are genuinely useful. But a control plane is a mirror, not a conscience. It shows you the sprawl. It does not decide who is responsible for it. Buy the tooling if it helps you see. Just do not outsource the accountability to it. The enterprises treating an agent registry as bureaucracy today are the ones that will spend next year explaining an incident to a board that suddenly cares. The ones treating it as infrastructure, named owners, scoped permissions, monitored behavior, a documented way to switch an agent off, are buying the right to move faster, because they can approve the next ten agents without adding ten unknowns. When you talk to your software and it acts on your behalf, the thing you are really deploying is delegated authority, and delegated authority without an owner is not automation. It is hope. Put a name next to every agent before you need to. Let's get to production,
AK | | The Agentic Enterprise Know more about AI than 95% of your peers. By 7 AM. A daily AI intelligence briefing for enterprise leaders, published by Spearhead. We build AI systems that work. Strategy. Engineering. Production. Outcomes. © 2026 Spearhead. All rights reserved. |
|
