Only 6% of companies get value from AI, Forbes reports

Six percent. That is the share of companies getting value from AI, according to a Forbes piece by John Koetsier published Sept. 16, and it appears in the outlet's enterprise AI coverage curated as of 3 p.m. ET that day. The rest of that feed reads like a working diagnosis of why everyone else is stuck.

Forbes' curated overview, refreshed at 3 p.m. ET on Sept. 16, boils its recent reporting down to three points. Enterprises are moving off a single-model approach toward a combination of frontier, open and specialized models. Despite heavy spending, many companies cannot show a clear return, which Forbes describes as a productivity gap, and as AI gets embedded in business processes the emphasis is shifting away from what a model can do and toward governance, security and the practical work of running it.

For the CIO or head of enterprise AI signing the next platform contract, that third point is the one that changes the purchase order. The model is becoming the smaller decision.

One model is giving way to a portfolio

Tim Bajarin argued in a Sept. 1 Forbes column that the future of enterprise AI is a mix of models rather than one. Gartner made a compatible forecast at its IT Symposium/Xpo in Orlando in October 2025, Network World's Michael Cooney reported: by 2028, more than half of the generative AI models enterprises use will be domain-specific.

The spending baseline shows how far that has to travel. Gartner put worldwide end-user spending on generative AI models at $14.2 billion for 2025, with specialized models, a category that includes domain-specific language models, at $1.1 billion, according to Network World. Against that starting point, Gartner is forecasting a large swing in a little over two years.

Gartner's argument for domain-specific models, as Network World reported it, is that models trained or fine-tuned on data for a particular industry, function or process deliver higher accuracy, lower cost and better compliance than general-purpose ones. Tori Paulman, a vice president analyst at Gartner, told the symposium that agents built on these models can read industry-specific context and make sound decisions in unfamiliar situations, with better explainability.

For an enterprise whose AI spend is concentrated in a single general-purpose model contract, the practical question this raises is whether the current architecture can route different tasks to different models without a rebuild. Gartner has a related number for the infrastructure side: by 2028 it expects more than 40% of leading enterprises to have adopted hybrid computing architectures, combining CPUs, GPUs, AI ASICs and other processors, in critical business workflows, up from 8% today, per Network World.

The bill is arriving before the return

Dev Patnaik's Sept. 15 Forbes column carries the title "When the AI Bill Comes Due." Kara Dennison's Aug. 24 piece calls AI productivity corporate America's biggest self-deception. Both are listed as sources for Forbes' current highlights, and together they are the basis for Forbes' summary that a significant number of companies cannot demonstrate ROI on their AI investment.

Forbes' own brief on closing the AI implementation gap says the move from pilots to enterprise-wide scale hinges on infrastructure, specialized talent and a focus on workflows. A companion brief frames execution as the next frontier: adoption is rising, but results require a shift from strategy to delivery.

That squares with what Thomas H. Davenport and Randy Bean wrote in their MIT Sloan Management Review column on five AI trends for 2026. They expect greater focus on generative AI as an organizational resource rather than an individual one, and continued progress toward value from agentic AI despite the hype. Individual productivity tools are easy to buy and hard to book as return, which is one reading of why the gap exists.

A 6% benchmark is uncomfortable, but it is useful. It means the CFO who can point to one AI deployment with a measured return in the ledger, rather than a survey of employee sentiment, is already running ahead of most peers. It also suggests the renewal conversation this fall could turn on evidence of return rather than on feature lists.

Governance and security move into the spec

Forbes' third highlight, the shift from model capability to governance, security and operational practicalities, has a Gartner forecast behind it. Gartner predicts more than 50% of enterprises will use AI security platforms by 2028 to protect their AI investments, Network World reported. Gartner describes these platforms as a single place to secure both third-party and custom-built AI applications, centralize visibility, enforce usage policies and guard against prompt injection, data leakage and rogue agent actions.

The reason that matters for agents specifically, per Gartner's remarks reported by Network World, is that agents act on probabilistic models and are therefore less predictable, which makes risk management less straightforward than it is for conventional software. Custom-built agents are opening new attack surfaces that demand secure development and runtime practices.

Who owns all this is still open. Daryl Plummer, Gartner's chief of research, said in the symposium keynote that enterprises need a dedicated AI leader, according to Network World. Davenport and Bean independently list ongoing questions about who should manage data and AI among their 2026 trends. Two separate reads landing on the same blank box in the org chart indicates the question has not been settled.

Data governance is part of the same conversation. Gary Drenik's Aug. 25 Forbes piece, featured in the outlet's spotlight section, argues unstructured data is becoming critical to enterprise AI.

InformationWeek's Pam Baker added another set of constraints in her January predictions for 2026: AI would run into energy ceilings and platform lockouts, organizations would develop a focus on performance per watt, and machine identities would outnumber humans by orders of magnitude. If those hold, the power budget and the identity directory belong in the same governance review as prompt-injection controls. For a facilities director already negotiating rack power, that is a planning input, not an abstraction.

Vendors are sending engineers into customers' buildings

The Information's Jyoti Mann reported on May 27 that Meta plans to place engineers and product managers inside large corporate customers through a new unit called Enterprise Solutions. The plan was laid out in an internal memo from Naomi Gleit, Meta's head of product, and its purpose is to get businesses using Meta's AI tools and services.

Sandy Carter's Aug. 11 Forbes piece frames a race among Microsoft, UnifyApps and others toward an "enterprise brain" that replaces stand-alone AI agents and loops. Read together with Meta's memo, the pattern suggests vendors now see the implementation gap as their own problem to close, not the customer's.

For a procurement director evaluating a platform this quarter, that turns embedded staff and services into a negotiable line item. If the gap between pilots and production is what separates the 6% from everyone else, the engineers a vendor is willing to put on site could be worth more than a discount on usage.

Gartner has left three 2028 checkpoints on the table: more than half of enterprise generative AI models domain-specific, more than half of enterprises on AI security platforms, and hybrid computing in critical workflows at more than 40% of leading enterprises. Only the last one comes with a current reading, 8%, which makes it the first number worth checking against next year.