Enterprises run a dozen agents each, with no one system in charge
Maven counts an average enterprise fleet of 12 agents, Levelpath says procurement agents stall across siloed systems, and Harvey and Dataiku show what happens when agents get real work with clear ownership.
- Maven says the average enterprise runs 12 AI agents, half of them in isolated silos, and the count is expected to grow 67% within two years. Self-reported m:…7076
- Maven says 94% of IT leaders worry sprawl raises complexity, technical debt, and security risk, while 44% of customers are frustrated by repeating their issue. Self-reported m:…7076
- Levelpath says procurement AI agents often fail because they work across siloed systems with different permissions, and argues agents should run under the same permissions as the human who started the session. m:…6230
- Harvey says SUSE ran an entire due diligence process in-house in one month and cut manual review time by 73% on a recent acquisition. Self-reported m:…4737
- Dataiku says ENGIE's support agent delivered a 20% productivity gain for a three-person team, coordinating four specialized agents for troubleshooting, guidance, knowledge retrieval, and escalation. Self-reported m:…8501
The fleet nobody manages
The average enterprise is running a fleet of 12 AI agents, according to Maven, and half of them operate in isolated silos. The number is expected to grow 67% within two years. Only 12% of organizations have a centralized platform to manage AI agents, which means most fleets are growing without an owner. m:…7076
The cost shows up in customer experience first. Ninety-four percent of IT leaders are concerned that sprawl increases complexity, technical debt, and security risk, and 44% of customers are frustrated by having to repeat their issue or re-enter information. Only 15% of executives believe repeating issues is a problem for customers, a perception gap that explains why the problem persists: the people buying the tools do not feel the friction the customers do. m:…7076
Maven's answer is a single customer agent working across chat, voice, email, SMS, WhatsApp, Slack, and Teams from one engine, with one context layer. The company cites Clio reaching 80% autonomous resolution and making live support four times faster, and Mastermind reaching 93% autonomous resolution within six weeks. Those are vendor customer stories, but they illustrate the pitch: consolidation is sold on resolution rates, not on architecture diagrams. m:…7076
Why agents stall in real systems
Levelpath's analysis of procurement explains why consolidation is hard. Procurement AI agents often fail because they operate across siloed systems with different permissions and failure modes, acting as fast interns with limited access rather than autonomous agents. Procurement AI adoption sits between 1% and 5%, the company says, despite 68% of organizations building AI solutions currently building agents. m:…6230
Levelpath's prescription is to treat agents as first-class users: they should operate under the same permissions as the human user who initiated the session, with explicit, admin-defined policies for human-in-the-loop approvals, and audit logs covering both successful actions and denied attempts. The company also notes that Gartner retired autonomous procurement from its Hype Cycle in 2026, a sign that the industry's expectations have cooled to match the implementation reality. m:…6230
Harvey's legal deployments show the same lesson from the other side. Bitfury uses Harvey to prepare transaction documents internally to 80% completion, reduced reliance on outside counsel by more than 50%, and says its lawyers save at least 10 hours per week. ICL Group coordinated its rollout across Legal, Data Privacy, IT, IT Security, and Procurement, and completed a cross-border restructuring matter internally in five weeks. The common thread is cross-functional ownership, not just a better model. m:…4737
Small teams, measured gains
ENGIE Supply and Energy Management is the small-scale version of the same story. Its support agent delivered a 20% productivity gain for a three-person support team, coordinating four specialized agents for troubleshooting, guidance, knowledge retrieval, and human escalation across a distributed environment of 25 servers. The system uses roughly 24 purpose-built tools, and the team's own account is that routine issues resolve automatically while humans keep judgment-based tasks. m:…8501
ENGIE has over 250 Dataiku users, about 80% of them in non-technical roles, and uses Dataiku to replace manual spreadsheet and business-intelligence processes. The detail matters because it shows where agents land first: not in a grand transformation, but in a queue that was already a bottleneck, with a team small enough to feel a 20% gain directly. m:…8501
The pattern across all four accounts is consistent. Agents work when they get a defined job, a permission model that matches a human's, and an owner who measures the result: SUSE saved nearly 1,500 hours over four months and delivers legal advice to the business 32% faster. Agents sprawl when they are bought team by team, with no inventory and no shared context. The fleet is already here; the management layer is what is missing. m:…4737
What is still unresolved
- Whether Maven's sprawl figures, drawn from its own customer research, hold across industries.
- How many procurement deployments actually implement first-class agent permissions versus stopping at pilots.
- Whether ENGIE-style gains scale beyond small, well-bounded support queues.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.