Last week, a report updated the data, and I stared at two numbers for a long time: 94% versus 2%.
These are core data points from KORE1's "2026 Mid-Market ERP AI Adoption Report" updated on August 5. The survey covered 100 mid-market executives who truly hold decision-making power over AI budgets. The results show that 94% of companies are already using generative AI, and 83% have moved beyond the experimentation stage into purposeful pilots or process integration. However, only 2% have actually achieved scaled operations.
A gap of 92 percentage points. Those who buy and those who use it are almost never the same people.
There is a sentence in the report that captures this gap perfectly: "The money is already approved. The plumbing is not."
I. Three independent surveys point to the same "cliff"
To address concerns that this might be a fluke from a single survey, KORE1 cross-referenced five independent studies from 2026, and the results were strikingly consistent:
| Research source | 样本 | "Used AI" | 「跑出规模」 |
|---|---|---|---|
| Kaufman Rossin × NewtonX | 100 mid-market decision-makers | 94% are using generative AI | 2% Scaled Operations |
| RSM Mid-Market AI Research | 1,030 executives from the U.S. and Canada | 86% partially/fully integrated | Only 17% are driving enterprise-level transformation. |
| Netrio IT Leader Survey | 401 seats (200-5000 person enterprise) | 82% already have AI running on production lines | 26% achieve scaled governance |
Different samples, different phrasings, but the pattern is almost identical: the adoption rate curve shoots straight to the ceiling, while the scaling curve hugs the floor.
Putting it all together, the conclusion is clear: the mid-market doesn't lack AI budgets or enthusiasm for trials—what it lacks is that "last mile" of pushing AI from pilot projects to company-wide adoption.
II. The bottleneck is not the model, but the "pipeline" layer beneath the ERP
So where exactly is this path stuck? The answer given in the report may surprise many managers who blame the problem on "AI products not being good enough":
- 53% of enterprises rank data quality as the top obstacle to scaling up;
- 47% mentioned system integration challenges;
- Manufacturing is even more concerning: 55% list legacy system integration as the top AI obstacle (industry average 41%), 73% are still in the pilot phase, and 45% of data remains in silos.
In plain terms: for AI to scale, the first step is not training larger models, but getting the "customer records," "material codes," and "closing standards" in the four systems aligned. If one customer is called A in the CRM, B in the ERP, and C in the reports, no matter how smart the model is, it cannot provide a consistent basis for decision-making.
Many companies ask, "Why is AI implementation not effective?" The answer often lies not in whether the model is chosen correctly, but in whether the underlying data pipeline is clear. The model is an amplifier, not an engine—if the pipeline is dirty, what gets amplified is even greater noise.
KORE1 also saw the same signal from the recruiting side: previously, clients looking for ERP consultants wanted people who "understood modules," but now they are looking for people who "can get four systems to agree on customer records and material codes, and carry that consistency through the quarter-end close." Only with unified master data can you even begin to point the model at it.
3. The industry trend is also shifting toward "practical results": this wave is not a cooldown, but a gear change
Data is saying "hard to land," while the market is saying "stop talking about concepts." On August 7, Securities Times published an observation piece with a very straightforward headline—the AI track is shifting direction, moving toward a new stage of pursuing practical results. Several signals are worth connecting:
Alibaba's enterprise-level intelligent agent product "Qianwen Office" has launched public beta, integrating multiple office products and targeting the enterprise market comprehensively. At ByteDance's mid-year all-hands meeting on August 6, the company clarified that its AI focus is shifting to ToB, explaining the background behind the integration of Doubao, Feishu, and Volcano Engine. Tencent is also consolidating its AI product lines simultaneously. Leading companies are unanimously redirecting resources from "competing on parameters and showcasing demos" to "competing on products and real-world implementation."
Securities Times' summary is spot on: the crude era of gaining attention through flashy demonstrations and concept hype has come to an end, and the enterprise AI track has entered a new phase of quality competition that emphasizes implementation, safety, and practical results. The yardstick in the capital market has also changed—technical parameters and concept stories no longer hold value, while implementation results, conversion quality, compliance standards, and customer retention have become hard metrics.
Is this a good thing or pressure for business managers? My view is: it's a good thing, but the challenges have come earlier. In the past, buying AI was like "getting on the bus first and buying the ticket later" — a nice demo was enough to close the deal; now buyers are increasingly rational and will directly ask: which process can your AI run on my ERP? Where does the data come from? Who is responsible if something goes wrong? Products that can't answer these questions won't survive this round of screening.
Fourth, fix the pipeline before discussing the model: the correct implementation order
After all this talk about data and trends, let's get down to practical application. If you remember just one thing, I suggest this: AI implementation is pipeline engineering, not model engineering. Get the order wrong, and you'll waste both money and time.
Step 1: Conduct an "AI Readiness Assessment" instead of directly selecting AI products. First, quantify the duplication rate, consistency rate, and differences in data standards across systems for master data (customers, materials, suppliers, accounts). If the assessment results are poor, fix the data first rather than forcing AI implementation.
Step 2: Start with scenarios where the "pipeline is shortest." Choose high-frequency actions where data is already relatively clean and processes are standardized—such as financial reconciliation, inventory alerts, order entry, and invoice matching. The approach of Juneyao Group can serve as a reference: Juneyao Airlines has advanced the systematic development of 64 scenarios around capabilities like action, decision-making, and learning. This was achieved not through one large, all-encompassing super application, but through the accumulation of numerous high-frequency, small-scale scenarios.
Step 3: Follow the rhythm of "Embed → Reconstruct → Integrate → Native" — don't expect to get it right in one go. First, let AI embed into existing workflows to improve efficiency. Once it runs smoothly, consider reconstructing the workflows. Only then can you talk about Agent-native. If you move through the stages too quickly, the organization won't keep up, and the system won't hold up either.
Another often overlooked point: this "pipeline-first" logic is especially important for small and medium-sized enterprises with limited budgets. A cloud ERP with an annual fee of a few thousand yuan, plus clean inventory and sales data, can support 80% of AI scenarios; conversely, with a pile of systems and data that each tell a different story, even the most expensive AI solution is useless. First get your numbers in order, then talk about intelligent transformation.
At the end, I want to say: the 92 percentage points between 94% and 2% are not because AI is inadequate, but because the data pipeline hasn't kept up. This shift from "showcasing skills to delivering results" is essentially making up for years of overdue master data governance. Whoever completes this first will have their AI running first—this sequence cannot be bypassed, nor should it be.
