74% plan agentic AI by 2028 – why "off-the-shelf" isn't enough

74 percent. This figure from the new Deloitte report "The State of AI in the Enterprise" is currently making the rounds in boardrooms and executive suites. It sounds like a clear mandate: AI agents are coming, and soon. But anyone who reads the study closely and compares it with the reality of German Mittelstand companies runs into a second, less comfortable finding - and a major announcement from Accenture and Google Cloud that shows just how differently "Mittelstand" can be defined, depending on who is selling a product.

For managing directors and executives at German Mittelstand companies, a very concrete question arises: should you act now, wait, or book an off-the-shelf package directly from one of the major providers? The answer, as so often, is more nuanced than the headline - but it can be derived surprisingly clearly from the available figures.

What Deloitte actually measured

The Deloitte report is based on a survey of 3,235 business and IT executives in 24 countries and six industries, conducted between August and September 2025. The core finding: 74 percent of respondents expect their company to use AI agents at least "moderately" within roughly the next two years. For comparison: today, only 23 percent already use agents moderately to extensively.

The working-title horizon "by 2028" is an approximation. Deloitte speaks of a two-year window from the time of the survey, meaning closer to 2027. This nuance matters, because it shows how quickly a precise study figure turns into a catchy but slightly imprecise headline in public discussion. It changes little about the substance of the claim: the jump from 23 to 74 percent within two years describes one of the fastest expected technology adoptions that current enterprise studies show at all.

It is also notable that Deloitte published a second, often-confused 74-percent figure in the same report: 74 percent of companies hope to achieve future revenue growth through AI initiatives - but only 20 percent already achieve that today. Two different questions, two different answers, coincidentally the same percentage. Anyone citing "Deloitte's 74 percent" in presentations or articles should therefore always specify which of the two figures is meant. This risk of confusion is not a minor detail: it shows how easily two methodologically sound but substantively very different metrics turn into a single, seemingly unambiguous statement when quoted - and how important it is to ask exactly what any AI statistic in a board presentation actually refers to.

The governance gap and the need for adaptation: two figures that belong together

The truly explosive figure in the report lies elsewhere: only 21 percent of companies say they already have a mature governance model for agentic AI. Adoption speed clearly outpaces control mechanisms. Companies want to roll out agents before they have clarified who monitors these agents, who is liable for wrong decisions, and how autonomous actions are documented.

The pace of agentic AI adoption clearly outstrips control mechanisms - only 21 percent of companies, according to Deloitte, already have a mature governance model for it.

This gap is especially relevant for Mittelstand companies, which rarely have their own compliance department or a dedicated AI officer. If you introduce an agent that pre-screens job applications, prioritizes customer inquiries, or automatically calculates quotes, you need clear answers from the outset to a few questions: Who spot-checks the agent's decisions? Who is notified when the agent makes an exception outside its defined scope? And how do you document this in case of an audit or complaint?

Important here: the governance question does not resolve itself automatically just because you buy a ready-made solution from a large provider instead of building it yourself. Even a purchased system must be embedded in your internal approval processes, monitored, and documented. "Buy" shifts the technical complexity, but not the entrepreneurial responsibility.

Perhaps the most important individual figure for this article's headline is a third number from the Deloitte report: 85 percent of companies expect they will need to adapt AI agents to their own business requirements. Conversely, this means only a small minority believes a standard product can be adopted and run productively unchanged.

That matches practical experience from Mittelstand projects. A manufacturing company with €20 million in annual revenue has different approval processes, different ERP structures, and different edge cases than a service company with €200 million in revenue - even if both supposedly want to solve the same use case, such as automatically processing incoming service requests. A standard agent built for the average of all of a software provider's customers rarely fits a single company's process landscape exactly. It has to be connected to the existing inventory system, adapted to industry-specific wording, and aligned with internal escalation rules.

This adaptation work is no small matter. It often decides whether an agent is actually used after rollout or disappears into a drawer as a "nice pilot project" after a few weeks. This is exactly the core of the argument for why "off-the-shelf" isn't enough for most small and medium-sized businesses: not because the underlying technology is bad, but because the last 20 percent of adaptation - the specific process logic, the industry-typical exceptions, the connection to existing systems - usually determines success or failure.

Accenture Edge and Google Cloud: who is the off-the-shelf package really for?

Fittingly, Accenture - through its newly founded unit "Accenture Edge" - and Google Cloud announced a package of pre-built agentic AI solutions specifically for mid-market companies on July 7, 2026. Technologically, the offering is based on Gemini Enterprise, the Gemini Enterprise Agent Platform, and the Agentic Data Cloud, combined with Accenture's own "Forward Deployed Engineers" - consultants who implement directly on-site with the client. Six areas are covered: Customer Intelligence & Growth, Customer Experience, Cybersecurity, Agentic Business Operations, industry solutions, and Agentic Workforce Enablement. In parallel, Accenture and Google Cloud announced an expanded partnership for large global enterprises around Gemini Enterprise - a separate but related offering that should not be confused with the mid-market initiative.

A crucial detail for German readers easily gets lost in the coverage: the "mid-market" target group defined by Accenture and Google sits at $300 million to $3 billion in annual revenue. That's a completely different order of magnitude from what is typically understood as "Mittelstand" in German-speaking countries. The EU SME definition draws the line below €50 million in revenue, and even broader German Mittelstand definitions rarely exceed €500 million.

In other words: the Accenture Edge offering targets companies that, by German standards, sit at the upper end of the Mittelstand up through entry-level corporate size - not the typical owner-managed business with €30 million, €80 million, or even €150 million in revenue. If you're the managing director of a company with, say, €80 million in annual revenue reading about this offering, it's worth a sober check before making contact: does your company even fall into the intended target group - structurally, in terms of pricing, and with respect to the internal resources available for working with a consulting firm of this scale?

There's also a structural aspect: the model explicitly ties Google Cloud technology to Accenture's own Forward Deployed Engineers. That means a dependency on the consulting partner not just for the rollout, but also for operation, adaptation, and scaling. Before signing any contract, it's therefore worth clarifying exactly how easily you could later switch providers or insource the solution - questions that easily get lost in the first excitement over a shiny product announcement.

The German Mittelstand: different numbers, different reality

While Deloitte and Accenture report from a global and mid-market perspective respectively, German surveys paint their own picture. The "AI Index Mittelstand 2026" from Salesforce, published in March 2026, finds: 16.6 percent of surveyed German companies already use AI agents - nearly double the 8.7 percent in 2024. A further 37 percent plan to introduce or expand use in 2026. Overall, 51.2 percent of companies already use or test AI in the broader sense.

Further German figures add to this picture but are not directly comparable methodologically, as they use different samples and definitions of "AI use": the Bitkom study from April 2026 puts active AI use in companies at 41 percent, up from 17 percent in 2024. The ifo Institute, in a survey from May/June 2026, arrives at as much as 54.5 percent of companies using AI. These three German sources - Salesforce, Bitkom, and ifo - cannot be overlaid one-to-one, because they ask different questions and survey differently sized, differently composed samples. Together, however, they paint a consistent pattern: German companies now have a solid level of AI use overall, but for the specific use of AI agents - as distinct from simple chatbot or assistant use - they remain well below the global expectation that Deloitte describes with its 74 percent.

In practice, this means: the Deloitte figure describes a global expectation across all industries and company sizes, while the Salesforce figure reflects the current, agent-specific reality in the German Mittelstand. Between "16.6 percent already use AI agents today" and "74 percent expect at least moderate use within two years" lies a substantial gap - and this gap is the real call to action for everyone responsible in the Mittelstand. Anyone not yet working with agents today, but who doesn't want to fall completely behind in two to three years, should start now instead of waiting for the perfect off-the-shelf solution.

Build vs. buy: the cost trap and how to avoid it

Regardless of whether you choose an off-the-shelf package from a major provider, a specialized partner, or an in-house build - the cost question is almost always underestimated in the first calculation. Several 2026 industry analyses on the total cost of ownership of agentic AI projects show a recurring pattern: companies frequently underestimate the actual total cost of self-built agents by 40 to 60 percent. Over a three-year period, the real total cost is often 1.5 to 2 times the originally calculated build cost. Responsible for this are ongoing maintenance, so-called prompt and model drift, necessary monitoring, and compliance effort, which are rarely fully priced into the initial business case.

These figures come from several industry sources and consulting analyses, not from a single, independently verified primary study - they should therefore be understood as reference points, not as an exact formula. Still, they provide an important guardrail for your own decision: if a provider, whether offering an off-the-shelf package or custom development, only names the initial build costs, ask specifically about ongoing costs for maintenance, model updates, monitoring, and - for consulting-heavy models like Accenture Edge - the expected ongoing consulting fees. These follow-on costs are frequently missing from initial proposals, precisely because they can barely be seriously quantified before actual operation begins.

A rough rule of thumb derived from several TCO analyses: the smaller and more clearly scoped a use case, the more an existing solution tends to pay off, because development costs are spread across many customers. The higher the interaction volume, the more specific the process logic, and the tighter the required integration with company-specific systems, the more a tailored build tends to pay off - because generic solutions would then have to be adapted so heavily that the difference from building in-house shrinks anyway. These thresholds come from general industry sources and are not a formula specifically validated for Germany or the Mittelstand, but they provide useful initial orientation.

The cost question is compounded by another, widely cited forecast: Gartner expects more than 40 percent of all agentic AI projects to be abandoned by the end of 2027. This figure comes from a June 2025 press release, based on a survey of 3,412 webinar attendees in January 2025 - a self-selected sample, not a representative full survey. As the main reasons for failed projects, Gartner cites escalating costs, unclear business value, and inadequate governance or risk controls. Gartner also points to the phenomenon of "agent washing": of the thousands of vendors marketing themselves today as "agentic AI," Gartner estimates only about 130 are truly substantially agentic - the rest are often renamed chatbot or RPA products with polished marketing.

This combination of cost trap and failure risk is no reason to stay away from agentic AI. It is a strong argument for starting small: a single, clearly scoped use case - such as pre-qualifying incoming job applications, automatically prioritizing customer inquiries, or continuous market monitoring for sales - can be tested with manageable risk, measured cleanly, and scaled up deliberately on success. An immediate, company-wide rollout of a complete off-the-shelf package, by contrast, increases exactly the risk factors Gartner names as reasons for failure: high upfront costs, unclear value across many departments, and governance structures that still need to be built.

What this means for your decision today

Taken together, the various data points paint a picture that is far more nuanced than the punchy headline "74 percent will soon rely on AI agents." Yes, global expectations are high, and yes, adoption speed is noticeably increasing. But several of the underlying figures - 21 percent mature governance, 85 percent expected need for adaptation, more than 40 percent projected project cancellations - show that the path there is neither simple nor achievable for most companies with a single off-the-shelf product.

For German Mittelstand companies, there's also the size question: large, prominently advertised offers like Accenture Edge in cooperation with Google Cloud target a segment that is simply oversized for most owner-managed businesses. This doesn't mean the underlying technology - Gemini Enterprise and the Agentic Data Cloud - is unsuitable for smaller companies. It means the actual sales model, consulting intensity, and presumably the pricing structure are tailored to significantly larger companies than what German usage understands by "Mittelstand."

Realistically, the right solution for most German SMEs sits between two extremes: neither a blanket consulting package designed for billion-dollar corporations, nor the attempt to solve all governance, monitoring, and integration questions entirely alone from scratch. A more sensible approach builds on existing, mature technology building blocks, but adapts them specifically to your own process landscape, company size, and existing IT infrastructure - with a partner who understands both sides: the possibilities of modern language and agent models, and the actual workflows in a company with 50, 100, or 250 employees.

Concretely, for the coming months this means: first, check internally which single process in your company suffers most from recurring, rule-based tasks - whether in recruiting, customer service, quote preparation, or market monitoring. At the same time, clarify who in the company will be responsible for monitoring and approving automated decisions going forward, even if that person initially spends only a few hours a week on it. And before deciding on an off-the-shelf package or an in-house build, get an honest assessment of the actual three-year costs - not just the costs of introduction.

Deloitte's 74-percent figure is therefore less a countdown that should push you into rushed action than a confirmation that the direction is right. The real entrepreneurial task lies in shaping your own path there so that it fits your own company size, your own risk appetite, and your own internal resources - rather than letting an impressive percentage or a shiny announcement push you into a pace that doesn't fit your business.