Agents Tripled, Content Did Not
Four research streams asked different questions this summer and found the same three limits inside the company: fragmented budgets, unusable content, and expectations set against the wrong clock. Deployment is accelerating anyway.
Four research streams published in the last eight months asked different questions. Salesforce asked what makes an internal AI agent reach return. A measurement study asked what makes an answer engine cite a brand. Forrester asked where customer experience will break and where 2027 budgets should go. Gartner asked what enterprise applications need to change.
They found the same three limits, and all three sit inside the company. None of them is about model quality. And organisations are deploying faster anyway: the average number of agents per organisation roughly tripled over the last fiscal year, and the average agent went from acting on two skills to six.
This piece names the three limits and says what to do about each.
Limit one: the budget is split across three owners
The same asset is funded three times. Forrester’s 2027 Budget Planning Guides tell technology leaders to increase spend on machine-readable enterprise context that agents can act on, and tell marketers to increase spend on brand visibility in answer engines. Those are two line items in one document. In most companies they are also two departments, and a third department funds the contact centre knowledge base.
All three are paying for the same thing. I will call it the corpus: the structured, current set of content and data that describes what a company sells, how it works, and what rules it follows, held in a form a machine can read. Split across three budgets, the corpus gets funded to three partial standards and meets none of them.
What to do: inventory the corpus before approving any of the three budget lines, and give it a single owner. That owner is not marketing, which owns the words but not the systems, and not IT, which owns the systems but not the accuracy.
Limit two: the corpus is not usable by a machine
This is the limit the evidence supports most strongly, from three directions.
Salesforce surveyed 2,025 agentic AI decision makers across twenty countries. For their most autonomous agents, deployers named clean and accessible data at the moment the agent acts, plus a tightly bounded use case, as the two things that mattered most. Both ranked above model quality and above orchestration tooling. Only 31 percent had put the relevant data in order before launch, and those who did reached meaningful return in 7.3 months against 8.8 months for the rest.
The measurement work reaches the same asset from outside, with a correction that matters. Across more than one hundred thousand prompt responses covering a hundred brands, about 78 percent of citations go to corporate websites. The study is explicit that most of that share is third-party pages from companies in the same space rather than the cited brand’s own pages. Corporate content is what the machines read. It is not necessarily yours.
The customer-facing evidence is the third direction. Escalation from AI agents to human agents held steady at 32 percent even as service volumes rose sharply. An escalation means the agent reached a question the content could not answer. That is a content gap, not a model limit.
Gartner’s January note lists the work involved, and it is engineering rather than marketing: atomic and reusable content models instead of monolithic pages, automated schema.org markup, dedicated question-and-answer content types, and semantic metadata carrying entity relationships rather than keywords.
What to do: run three tests on the corpus. Can an internal agent act on it at the moment it needs it. Can an answer engine quote it without filling the gaps itself. Can a service agent close a conversation with it. Pages that fail one test usually fail all three, which makes a single remediation programme easier to defend than three separate ones.
Limit three: the expectations are set against the wrong clock
Boards are asking whether the company is moving fast enough. The data says speed of deployment does not predict speed of return.
Professional and business services was among the slowest sectors to deploy agents and among the fastest to meaningful return, at 6.5 months. High tech is one of the largest deployers and one of the slowest to return, at 10.1 months. Across the survey, average time to meaningful return was about eight months, and 47 percent of respondents were still piloting after two years of availability.
Forrester draws the budget conclusion directly: cut AI pilots that scale activity without governance, ownership or a defined path to scale. Its customer-side prediction is the same error seen from the outside, with a third of companies expected to damage customer experience in 2026 by pushing conversational agents into contexts where they were never likely to work.
What to do: set the internal expectation at roughly eight months to return, and measure progress on the corpus rather than on the number of agents deployed.
The gap is widening
None of these limits stops deployment. Agents per organisation tripled, creation time fell by more than half, and over 80 percent of leaders expect larger budgets in the next twelve months. Gartner expects 80 percent of customer interactions to move to agentic interfaces by 2028.
So the volume of agents is rising against content that is not ready, budgets that are split, and timelines that are wrong. The three limits are cheap to fix early and expensive to fix once agents are running against them in production.
Zurich and Milan
Around six in ten Swiss companies report using AI in some form, from a UBS survey of roughly 2,500 firms. In those companies the three limits are already live and remediation means changing systems people depend on.
In Italy the Politecnico di Milano observatory on digital innovation in SMEs found that 76 percent have neither invested in AI nor plan to, and 7 percent run structured training. Those companies still have the cheap version of all three fixes available. They also have a real disadvantage to work against: in the visibility measurement, small and niche brands appeared in 11 percent of relevant AI answers against 73 percent for household names.
Next: which content structures answer engines actually read, which they ignore, and why the most-cited format in AI answers is a ranked list somebody else wrote about you.
Stay close.
— Stefano
Sources
- Salesforce, New Study of 2,025 Agentic AI Leaders: First To Launch Isn’t Fastest to ROI (State of Agentic AI in the Enterprise), 27 August 2026. https://www.salesforce.com/news/stories/agentic-ai-leaders-survey-on-roi/
- Salesforce, Agentic Enterprise Index: Agent Deployments More Than Double Year over Year, 7 August 2026. https://www.salesforce.com/news/stories/agentic-enterprise-index-insights-2026/
- Pratyush Kumar, Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines, arXiv:2606.20065, 18 June 2026. https://arxiv.org/abs/2606.20065
- Irina Guseva, Optimize Enterprise Apps for Agentic AI’s GEO and AEO, Gartner, note G00841018, 7 January 2026.
- Forrester, 2027 Budget Planning Guides, press release, 15 July 2026. https://investor.forrester.com/news-releases/news-release-details/forresters-2027-budget-planning-guides-after-year-caution
- Forrester, 2026 B2C Marketing, CX & Digital Business Predictions. https://www.forrester.com/press-newsroom/forrester-b2c-marketing-cx-digital-2026-predictions/
- UBS, AI in Switzerland: Pragmatic adoption, growing potential. https://www.ubs.com/global/en/wealthmanagement/insights/marketnews/article.3430283.html
- Osservatorio Innovazione Digitale nelle PMI, Politecnico di Milano, research results, May 2026. https://www.osservatori.net/