First, here is a number that is more concrete than most claims about the benefits of AI.
TechWolf co-founder Jeroen Van Hautte published an experiment in June: in a 250-person organization, an agent that directly searches underlying records costs 50% more per question answered than an agent that uses a "context graph"; when the organization reached 1,000 people, the gap widened to 129%. The experiment involved 30 person-lookup tasks, each run three times.
On November 16, SAP announced the completion of the acquisition of this Belgian company; the agreement was announced on October 6, and the price was not disclosed. SAP plans to integrate its data model into SuccessFactors and Joule to answer “who can fill this position,” “which team has a skills gap,” and “what capabilities should be developed.”
Putting these together, there is one judgment worth making clear: a large part of what enterprises pay for AI is not spent on "thinking", but on "getting to know things all over again." The larger the scale, the more conspicuous this spending becomes.
1. First, clarify the boundaries of this number
The two TechWolf organizations are simulated, not real customers; SAP also did not provide any comparable customer deployment results in the completion announcement. Manoj Swaminathan (President and Chief Product Officer of SAP Autonomous Suite) said in October that this graph "makes token usage more efficient" and "lowers the cost of deploying workforce agents," again only directional, with no customer figures.
So the correct reading is: 50% and 129% cannot be taken as promises, but can be taken as a directional hint—whether the context is organized well will be reflected in the cost of each call, and as the organization becomes more complex, the gap will grow.
When citing these kinds of figures, please include a qualifier.The benefit figures in vendor announcements roughly fall into three types: customer-measured, vendor internal-measured, and simulated experiments. Their credibility varies widely; mix them even once, and all the figures that follow will be called into question.
II. What Is Actually Stored in the Graph
What TechWolf does sounds plain enough: it links together roles, tasks, employees, and skills, and stores them in a structure it calls a "context graph".
Its data comes from two types of sources. One is the HR and business systems that customers connect themselves; the other is work traces, such as records from tools like Jira and Confluence, which are used to infer what skills a person actually uses. It also integrates labor market data to see how in-demand a given skill is externally. Employees themselves can review and correct the inferred skills—this is crucial, because the biggest fear for a skills graph is "guessing wrong and no one knowing."
With this network in place, things are different. When a project lead needs to find "people who have done a certain type of task," the agent does not need to reread everyone's resumes and project records; instead, it follows the already-established links among employees, projects, roles, and skills. It may also surface some people whose job titles do not reveal relevant experience—someone with the title "after-sales engineer" may have done more than a dozen on-site debugging jobs over the past two years, and this kind of information is often absent from resumes.
A side effect is saving materials. Fewer raw records need to be read for each query, so tokens naturally go down. That is where the 50% and 129% come from.
For an already implemented scenario, look at Italian pharmaceutical group Chiesi. It used TechWolf's job architecture restructuring agent to consolidate roles with similar skill requirements, draft job descriptions at different seniority levels, have them reviewed and finalized by internal teams, and then implement them in SAP's Job Profile Builder through implementation partner TalenTeam, including job remapping and employee remapping. According to TechWolf's case description, this work, together with the remapping, took several weeks.
The weight of this case lies in the four characters "review and finalize". The agent does consolidation and drafting; the final decision rests with humans. Job adjustments involve everyone's vital interests. If machines were allowed to decide such matters directly, it would be impossible to implement.
III. The three figures on SAP's side show that what is missing is not the model
In mid-November, CCIDnet and e-works separately published interviews with SAP China Research Institute. Cai Yanbin, SAP global vice president and president of SAP China Research Institute, described a very specific failure case: a friend who does OEM bicycle processing for export asked an agent to generate a 'pricing strategy for the next three months.' The output was structurally complete and seemed very convincing, but was completely unusable. The agent did not know the product's specific processes and costs, did not understand overseas customers' bargaining habits, and had no historical order data to support it.
He calls this the 'iceberg dilemma.' At the proof-of-concept stage, relying only on general capabilities to solve a single-point problem tends to produce results easily; once it goes deep into core business, a model lacking context immediately struggles to adapt.
Three numbers are worth remembering:
- More than 7 million fields, more than 500,000 data tables, and over 15,000 business processes. These are the structured assets SAP has accumulated over fifty years, now being organized into a knowledge graph so that models understand not only general industry knowledge but also "your company."
- 88% versus 6%.SAP's 2026 AI Value Report shows that 88% of surveyed Chinese enterprises are optimistic about the transformation brought by AI agents, while only 6% believe they are fully prepared. This gap is not a matter of willingness, but of readiness.
- More than three hundred. The number of agents published in the SAP agent layer. But Cai Yanbin also made clear that it will not replace the application layer; the data layer will still be responsible for organizing and supplying.
He used an analogy that I think is very apt: people need both knowledge and experience to make judgments. With only knowledge and no experience, judgments are inaccurate; with only experience and no knowledge, judgments are unsystematic. For AI to understand an enterprise, it likewise needs both.
Many people think that "experience" naturally emerges once a model's scale reaches a certain level. Actually, it does not. Experience is the company's own records, the pitfalls it has encountered, and the exceptions it has defined; no matter how large the model is, it cannot grow this.
IV. Implementing in Odoo: These relationships are already written in the tables
When they see 7 million fields, small and medium-sized enterprises can easily reach a wrong conclusion: this job is too big, we can't handle it. The opposite is true—your field count may be only one hundredth of that, but the few relationships that truly trip up AI will be present in every company.
What is even more coincidental is that Odoo's data structure itself is a relational network. Customers are linked to parent companies, and contacts are linked to customers; products are linked to supplier product codes, BOMs have hierarchies, and operations are linked under routings; employees are linked to skills, projects are linked to customers, and inventory moves also have upstream and downstream chains between them. These relationships are not something to be newly created—they are fields that are already there.
The problem lies in maintenance. In the projects I've seen, these types of fields have the highest vacancy rate:
| Relationship | Corresponding objects and fields | Common Current Situation | Who should maintain |
|---|---|---|---|
| Customer consolidation | res.partner's parent_id, vat | Three profiles for the same customer, with contacts attached separately to each | Sales Back Office |
| Supplier part number | product_code of product.supplierinfo | A single material can be written in more than a dozen ways, relying on people to recognize it. | Procurement |
| Material Level | mrp.bom and bom_line_ids | The modified BOM has no version trace. | Craft |
| Employee skills | hr.skill, hr.employee.skill | A table is built, but if nobody fills it in, it's as good as an empty table. | HR and Workshop Supervisor |
| Projects and Clients | partner_id of project.project and project.task | The client name is written in the project name, and the field is left blank. | Project Manager |
| Document traceability | move_orig_ids and move_dest_ids of stock.move | After manual merging and splitting, the chain broke. | Warehousing |
This table can be used directly as a worksheet. Fill it in row by row; whichever row you get stuck on is where future AI will not answer accurately.
Let's take a real example. One customer that does equipment maintenance has after-sales engineers who have to answer seven or eight questions a day like, "Who installed this equipment last time, and what parts were replaced afterward?" This information is scattered across four tables: equipment records, installation work orders, spare parts requisitions, and service records. A manual lookup takes about ten-plus minutes each time, and when things get busy they answer from memory, with a fairly high error rate.
What we did later was not complicated: we completed the linking fields between equipment records and service records, requiring every service to be attached to an equipment; then we linked spare parts requisitions to service orders. After the data was completed, we connected a read-only assistant; for the same type of question, it produces results in seconds, with document links. This project did not use any new model; all the benefits came from "the relationships being connected."
V. How Do We Settle This Account
Talking about context easily turns into mysticism, so in the end it comes down to algorithms. I usually have clients calculate three figures, and they are all quite crude:
- Lookup time. For the same type of question, how long does it take a person to look it up once, and how long does it take the assistant to look it up once? The difference in minutes multiplied by how many times it is asked in a year is the first number.
- Time spent on rework. Repeated communication, repeated orders, and repeated shipments caused by being unable to find it or finding the wrong one. This category is often larger than the first transaction, but no one usually tracks it.
- Cost of invocation. If billing is by token, multiply the average number of questions per day, the average consumption per question, and the unit price. After the graph is built, consumption per question will decrease, but you must measure how much it decreases yourself; do not copy the vendor's simulated values.
Add up the three items and divide by the labor invested in building relationships, supplementing data, and assigning maintainers; that gives the payback period for this investment. In most cases, this period is much shorter than the cycle for deploying a new system, because there is no downtime cutover and no impact on existing processes.
An actionable starting point: Pick one question from those "asked more than five times in the past month, and whose answer requires joining three tables to piece together." Fill in the join fields for these three tables, assign a maintainer, and use two weeks to record a baseline. If you can calculate the price difference, continue; if not, it means you picked the wrong scenario, so pick another.
VI. Closing with Three Sentences
First, context is an asset, not a switch. It requires people to maintain it, regular correction, and someone to be accountable for whether "this field is filled in correctly." A system you buy will not come with your company's relationships.
Second, establish relationships first, then discuss models. For the same set of problems, the gap before and after the relationships are completed is often larger than that from switching to a stronger model, and this gap will not disappear when the model is upgraded.
Third, don't take someone else's simulated values as your own targets.50% and 129% can be used to persuade the boss to approve the project, but they cannot be written into acceptance criteria. Acceptance should use baselines you have run yourself.
Cai Yanbin said the governance framework must exist on the first day of project launch, not be remedied after problems arise. This statement also holds in context: relationship fields cannot be completed in the week before launch; they must grow with the business day by day.
Models change every year, and suppliers change frequently too. What truly remains—and will only become more valuable—is that network showing "who is connected to whom."
References: ERP News, 'SAP Closes TechWolf Deal to Expand Workforce Intelligence in SuccessFactors' (2026-11-16); CCIDnet, 'When the Agent Wave Hits ERP: What Enterprise AI Needs Is Not “Smarter” but “Business-Savvy”' and the e-works reposted interview (mid-November 2026). TechWolf's cost comparison data comes from a simulation experiment published by the vendor, not actual customer testing; SAP's completion announcement did not disclose the price or customer-side benefits. The Odoo field names in this article are based on Odoo 19/20.
Zhongdi Information · Enterprise Management Digitalization Insights · Vendor-disclosed figures have not been independently verified; please refer to the original text when citing.
