On June 2, at Workday's enterprise conference in Las Vegas, a product called Agent Passport was released—issuing a "passport" to each AI Agent, recording its test scores, security verification, and continuous monitoring data. The next day, Bain & Company released a report that left the entire industry in silence: over $1 trillion in global capital has poured into AI, but the core ROI indicators are concerning. The report's title was just four words—"Technology works, value hasn't arrived." Also on June 3, domestic company Lanxin released a fully localized government and enterprise intelligent office solution. At the press conference, the deputy district mayor of Xicheng District made a weighty statement: Digital transformation has shifted from an "optional question" to a "mandatory question."
Three things have come crashing down in rapid succession within three days. I've read them over several times and feel that, strung together, they actually point to the same issue: AI's technology validation period is over; the real battlefield lies in the "last mile"—and most small and medium-sized enterprises haven't even entered the game yet.
In this article, I won’t talk about the beautiful visions in big companies’ PPTs, but only about one thing: how small and medium-sized enterprises should actually take action now.
$1 Trillion Cold Water: What the Bain Report Is Really Saying
Let's start with Bain's report. Released on June 3, 2026, the study covers the global return on AI investment. In plain English, the core conclusion is: large enterprises have poured $1 trillion into AI, proving it "works" on a technical level—but commercial value? Sorry, it's still a long way off.
The original report roughly states: the technical PoC (Proof of Concept) is mature, but the realization of commercial value is still stuck at the "last mile." Let me translate: AI works fine in the lab, in demos, and in PPTs, but once it needs to be truly embedded into a company's business processes, replace human decision-making, and run alongside existing ERP systems—it falls flat.
This judgment is completely consistent with what I have seen in actual projects. I have seen a case of a medium-sized manufacturing enterprise—the IT team spent six months developing an intelligent procurement prediction system, and the prediction accuracy did improve by 18 percentage points, but it was ultimately shelved. The reason is simple: the system's output could not directly enter the existing ERP procurement module, requiring another three months for system integration; the procurement department did not trust the AI recommendations and still placed orders manually. The $1 trillion dilemma is epitomized in this one project: There is a huge gap between technical capability and business value.
Significance for Small and Medium Enterprises
Don't rush to be pessimistic. The subtext of the Bain report is actually: You shouldn't spend big money on PoC first, because others have already done it for you. Large enterprises have spent $1 trillion verifying that "AI technology works," and you just need to skip the verification and go straight to value realization.
| Stage | Current status of large enterprises | SME opportunities |
|---|---|---|
| Technical Verification | Completed (funded by 1 trillion US dollars) | Skip directly, no need to invest repeatedly |
| System Integration | Stuck (last mile) | Choosing the right tool allows for a low-cost entry point |
| Value Realization | Less than 1/3 of projects met the target | Small-scale guerrilla warfare is, on the contrary, flexible. |
There is another set of data in Bain's report that surprised me a bit: even among "the most advanced companies in AI implementation," less than one-third of AI projects have truly realized value. The remaining two-thirds are still stuck in the stage of single-point experiments and localized optimization.
For small and medium-sized enterprises, this is actually good news. It shows that the "guerrilla warfare" approach—small scenarios, quick validation, replicability—still has plenty of room.
Workday issued a "passport" for AI agents: Should small and medium-sized enterprises follow suit
Regarding Workday, I believe its symbolic significance outweighs the product itself.
Released on June 2, Agent Passport is essentially a "driver's license + annual inspection" system for AI Agents. Each Agent must pass tests before going online—whether the most severe risks have been cleared; leave traces during operation—verification records are fully traceable; and undergo continuous monitoring after going live. The first partner is Cisco AI Defense, which is convincing in the security field.
My judgment is: This is the first time an enterprise management software vendor has turned "Agent governance" into a platform-level product. Previously, discussions about AI governance were all about white papers, frameworks, and principles—sounding good but not implementable. This time, Workday has made it a tool: you buy my Agent, and it comes with built-in compliance certification.
What is the industry trend behind this?
In the next three years, enterprises purchasing AI agents will be just like buying SaaS today—the first question won't be "can it work," but "do you have governance certification."
For small and medium-sized enterprises, most haven't yet thought about this level. I know that what SME owners usually think is: Can AI help me save two people? Can AI be used to screen customer service inquiries first? Can inventory alerts be automatically sent? — These are all correct, but they still remain at the "efficiency" level.
No one has thought about this: if an AI Agent handles customer refunds for you and misjudges an order, refunding an extra 5,000 yuan, who do you hold accountable? Without a complete record of the Agent's decision-making, you can't even trace responsibility. Not to mention that when industry regulations come into effect, if you can't produce compliance proof, you won't even qualify for bidding.
Cognitive gap of small and medium-sized enterprises
Now is the cheapest time to make up for costs. If you wait until industry standards are truly established and regulations are actually implemented, then passively comply, the cost could be five times what it is now.
In an open-source ERP like Odoo, you can do three things (no need to buy Workday, no need to integrate with Cisco):
- Audit Log: Records the input, output, and manual confirmation results of each AI-assisted decision
- Threshold circuit breaking: For operations exceeding the preset threshold, mandatory manual review is enforced.
- Monthly AI Decision Accuracy Report: Monthly statistics on the adoption rate and accuracy rate of AI suggestions, serving as foundational data for governance
These three things do not require any additional procurement; Odoo's audit module + custom reports can achieve them. Record the decisions first, then talk about efficiency optimization. The order cannot be reversed.
Lanxin's "Full-Stack Domestic Production": A Signal for the Government and Enterprise Market
I carefully reviewed the full-stack domestic government-enterprise intelligent office solution released by Lanxin on June 3. The press conference was of a high standard—the Deputy District Mayor of Xicheng District attended and delivered a speech, saying: "Artificial intelligence is the core driving force of the new round of technological revolution, and the digital transformation of government and enterprises has shifted from an 'optional question' to a 'mandatory question.'"
The phrase "must-answer question" is more convincing than any industry report. It means: in the government and enterprise market, localization is not a bonus point, but a ticket to entry.
The "full-stack secure intelligent office solution" launched by BlueMail this time is fully localized from the operating system to the application layer, with two core logics: first, to avoid being "choked," and second, to meet compliance requirements. These two logics are hard constraints in the government and enterprise market, not optional.
| Dimension | Government/State-owned enterprise clients | Private enterprise clients |
|---|---|---|
| Domestication pressure | Hard requirements (driven by the Xinchuang catalog) | No hard requirements, but the trend is upward |
| Data sovereignty | Must be deployed locally or on a private cloud | SaaS is acceptable, but data security awareness is on the rise |
| AI governance requirements | Regulation-driven, compliance is survival | Business-driven, still in early stage |
| Procurement decision cycle | 6-12 months (including approval process) | 1-3 months (short decision chain) |
For small and medium-sized enterprises and Odoo implementers, the reference value of this trend lies in:
If your clients are government/state-owned enterprises—the degree of localization directly impacts procurement decisions. As an open-source solution, Odoo inherently offers the advantage of "controllable source code and independent deployment." Your servers are under your control, with full data sovereignty, making it easier to pass government security reviews compared to SaaS vendors that claim "your data is stored on our cloud."
If your client is a private enterprise—there is no immediate pressure for localization, but the topic of AI data security is shifting from a compliance requirement to a commercial competitive advantage, and this may happen within just the next two years. Laying the groundwork early is not about following trends; it's about securing leverage for the future.
To put it bluntly, Lanxin has taught me this: The combination of "open source + private deployment + data sovereignty" is a real demand in the government and enterprise market, not just a marketing gimmick. Odoo implementers can turn this combination into a differentiated selling point for government and enterprise clients—and this selling point is truly backed by a technical foundation.
SMEs' "Guerrilla Warfare" Path: 90-Day Validation, No ERP Replacement
After discussing the signals of the three major events, let's return to the most practical question: What should small and medium-sized enterprises actually do now?
I've thought about it repeatedly and believe that the biggest taboo for small and medium-sized enterprises implementing AI is to imitate large enterprises by adopting a "big and comprehensive" AI strategy. Large enterprises can set up an AI committee, spend six months on planning, and then another year on PoC—SMEs cannot afford to wait, nor should they do this.
I call this approach the "guerrilla warfare path", which boils down to three sentences:
- Don't Replace ERP——Add AI capabilities on top of existing systems without touching the underlying layer
- Start with a small scenario—find a scenario with the clearest pain point and most complete data to run first
- 90-day verification — see results within 3 months, change direction if not
Why Not Switch ERP
I have seen too many enterprises that, the moment they start digitalization, want to replace their ERP. The cost of replacing ERP is not just the software expense—data migration, employee training, and business disruption each cost real money. For small and medium-sized enterprises, replacing ERP once basically reduces operational efficiency to zero for nearly half a year.
More importantly: Switching ERP systems won't solve the AI problem. AI is not a module of ERP; it is a layer of capability parallel to ERP. The right approach is to integrate AI capabilities into the existing ERP system, and Odoo's modular + open-source features precisely support this low-cost integration.
Which scene to start with
I suggest choosing one from these four scenarios—the criteria are "data available, rules clear, high fault tolerance":
| Scene | Why is it suitable to do first | Investment threshold | Expected Return Period | Risk Level |
|---|---|---|---|---|
| Financial Reconciliation | Clear rules, structured data, time-consuming manual work | Low | 30-60 days | Low |
| Order Processing | Highly repetitive, heavily templated, with controllable error costs | Low | 30-45 days | Low |
| Customer Service Initial Screening | FAQ-type questions account for a high proportion, with significant labor costs | 中 | 45-90 days | 中 |
| Inventory Warning | Data already exists in ERP, rules are configurable | Low | 60-90 days | 中 |
These four scenarios share one commonality: AI only makes "suggestions," while humans make "decisions." AI recommends reconciliation results for you to confirm; AI classifies orders for you to review; AI pre-screens customer issues and transfers them to human agents. This "AI-assisted + human-confirmed" model preserves human judgment while giving AI room to make mistakes — if problems arise, the chain of responsibility is clear; if no problems occur, efficiency is indeed improved.
How to split 90 days
Days 1-30: Data Collection + Process Mapping. After selecting the scenario, first clarify the data flow and business process of this scenario in Odoo—identify which fields are key inputs, which are decision nodes, and which are outputs. At the same time, export historical data from the past 6 months as a training set. Do not touch AI during this phase; only perform data cleaning and process documentation. Days 31-60: AI Tool Integration + Parallel Operation. Integrate AI capabilities into the Odoo environment (using Odoo's AI module or via API integration with third parties), but do not directly replace human work—let AI suggestions and human decisions run in parallel, comparing the differences. Your core task in this phase is to measure: What is the accuracy rate of AI suggestions? What is the average time spent on manual processing, and how much time is saved with AI assistance? Days 61-90: Effectiveness Evaluation + Expansion Decision. Three core metrics: Accuracy (the proportion of AI suggestions adopted by humans), Efficiency Improvement (the percentage reduction in time spent on the same task), and Employee Acceptance (whether frontline staff are willing to continue using it). If all three pass, the scenario validation is successful and the scope can be expanded; if any one fails, go back to check the issue or restart with a different scenario.
The advantage of 90 days is that even if the final result is not ideal, your sunk cost is limited—what you invest is human time, not millions in software procurement fees. Moreover, you gain three things: a true understanding of AI capabilities, a diagnosis of your own data quality, and a basis for judging the next scenario.
A Practical Checklist for SMEs and Odoo Implementers
I condensed the above content into an actionable checklist. Don't take on too much; take it one step at a time.
5-step action checklist
1. Conduct a data inventory this week. Open your Odoo system and see which module has the most complete data and the most standardized processes — that will be your first battlefield for AI implementation. Don’t choose the area with the most missing data, as that would be digging a hole for yourself.
2. Lock in a scenario and clearly define "what AI does and what humans do." For example, in financial reconciliation: AI handles matching and anomaly flagging, while humans perform final confirmation. If the boundaries are not clearly defined, disputes will inevitably arise after deployment.
3. Conduct rigorous validation for 90 days. If accuracy is below 80%, efficiency improvement is below 30%, or employees are unwilling to use it — if any one criterion is not met, do not expand the scope. First fix one, then consider the next.
4. Record AI decision logs from day one. Refer to the Workday Agent Passport approach, recording every AI-assisted decision in the Odoo audit module. This isn't about following trends; it's because this could become an industry entry requirement within the next three years. Doing it now costs the least.
5. If you serve government and enterprise clients, make "open source + private deployment + data sovereignty" your selling point. Lanxin's domestic full-stack solution has validated the real demand in the government and enterprise market. Odoo naturally possesses this foundation—controllable source code, independent deployment, and data staying within the domain. This is not marketing rhetoric; it is a technical fact.
Schedule Recommendations
| Time period | Things to do | Deliverables |
|---|---|---|
| June-July | Scene Selection + Data Cleaning | Business process documentation for the selected scenario + cleaned historical dataset |
| August-September | AI access + parallel operation | AI-assisted decision-making vs. manual decision-making comparison report |
| October | Effectiveness Evaluation + Expansion Decision | 90-day validation report (including three indicators: accuracy, efficiency, and acceptance) |
| Before the end of the year | The first scene officially goes live + launch of the second scene | Production environment AI feature launch + verification plan for scenario 2 |
Stop thinking about creating a "big and comprehensive" AI strategy first. Even trillion-dollar enterprises get lost in their strategies; the only thing small and medium-sized businesses should do is: find a scenario where you can survive, and survive first.
AI implementation is not a 100-meter sprint, but a guerrilla war. Find a hilltop, plant a flag, hold it, then find the next one.
The technology works, but the value hasn't arrived yet. So where is the value? It might be in your next 90-day validation.
