Gartner says 70% of enterprises will walk away from vendor-built agents by 2028

The consultancy's warning is aimed at what it calls forward-deployed engineering, or FDE: vendors embedding their own engineers with a customer to build and deploy software for that customer's requirements. Gartner argues the model can produce fast early progress while leaving the buyer dependent on expensive external expertise. The Register reported the prediction on 30 September, citing Gartner's forecast that 70 percent of enterprises will abandon vendor-assisted agentic systems by 2028.

Gartner senior director analyst Mukul Saha framed the problem as one of engagement design. "The best-scoped FDE engagements have clear guidelines on governance, business value delivery, IP ownership, project co-ownership, knowledge transfer, and an exit strategy from day one," he said in the research note. He also warned that "many providers now use 'forward deployed' as a label for implementation, professional services, solution engineering, or AI consulting," with some charging premium fees without the delivery depth to justify them.

That is a sharp contrast with the launch energy of the past two days.

OpenAI used its DevDay conference on 29 September to show Dots, always-on agents that act on a user's behalf. WIRED's reporter spent a night testing one and found it flagging an unfinished data request and an underplanned trip. CNBC noted the agents are limited at launch to OpenAI's Pro plan, which starts at $100 per month, though Sam Altman said: "You should, of course, expect us to do a mass-market thing for billions of people." The same CNBC piece recorded that Meta shares rose 29 percent in September after the Muse launch, the stock's best month since 2013.

Security and cost arrive with the demos

Muse, Meta's personal agent, is already under scrutiny. Tom's Hardware on 30 September reported that Inc's Jason Aten found Muse referring to a private Messages conversation it had never been granted permission to read; the agent claimed the Mac app had read incoming notifications and passed the context to the iPhone app. Meta disputes the characterisation, according to TechCrunch. Tailscale, meanwhile, published a technical post on 30 September describing how Muse connects to a user's tailnet as its own node, with outbound-only connections and explicit confirmation the first time it touches a device.

The infrastructure bill is not small either. A Robonomics estimate published on 30 September put the average power needed to serve 100 million daily Muse users at roughly 1 to 2 GW in its base case, with 3 to 4 GW plausible depending on how many reasoning-grade model calls each user generates. The same analysis reckons the sandbox layer alone needs 75 to 100 PB of physical DRAM, working from an observed configuration of 2 vCPUs, about 8 GB of RAM and roughly 100 GB of persistent storage per user.

Governance tooling is filling the gap. OpenClaw published OpenClaw Enterprise on 30 September, an open source control plane for persistent agents that the project says began at OpenAI and was donated to the OpenClaw Foundation, then developed with Red Hat and NVIDIA. It is being piloted internally at Red Hat and OpenAI, and is free to self-host. Archestra's OpenAPPA, posted the same day, claims no scored attack succeeded across 1,320 evaluations on Bench-Corp and AgentThreatBench, while completing 88 to 90 percent of tasks; Microsoft FIDES let 28 to 35 percent of attacks through in the same comparison. Separately, Cloudflare began an open beta of Issues, error monitoring for Workers that can send a grouped exception straight to a configured coding agent.

None of that resolves the FDE question. Gartner also predicts that through 2028, fewer than 20 percent of FDE engagements will turn recurring customer requirements into features in the vendor's core product, and warns of "FDE washing" as conventional consulting gets rebranded. Adoption figures point the other way: Vercel's 25 September State of agent skills report, cited by DevNavigator, counts more than one million listings and nearly 280 million recorded installs on skills.sh in seven months. Install counts are registry activity, not proof of business value, which is roughly the gap Gartner says buyers will keep falling into.