# Edwin Trebels

Edwin Trebels writes about enterprise AI, knowledge graphs, and turning business data into useful, traceable systems. Their posts connect technical topics to practical concerns such as governance, regulated industries, manufacturing, and sales.

- Profile: https://b2binfluencer.marketing/creators/edwin-trebels/
- LinkedIn: https://www.linkedin.com/in/etrebels
- LinkedIn followers: 6,283
- YouTube: https://www.youtube.com/channel/UCFtaReiR4JxuJqB1BNCFJCw
- Topics: artificial-intelligence, leadership, sales, physical-ai
- Based on 22 dated posts, latest 2026-10-06

## Summary

Most useful for translating enterprise AI and knowledge graph work into buyer-facing language for regulated industries.

- Best for: Enterprise AI or knowledge graph vendors selling into regulated industries.
- Not for: Mass reach or consumer campaigns needing guaranteed impressions and neutral endorsements.
- Check first: Ask for available formats, terms, and recent per-post impressions and engagement.

### Takeaways

- Enterprise AI fails on unreadable data, not weak models.
- A system you cannot trace, you cannot trust.
- Buyers describe the problem, not the category.

## Main ideas

- **Enterprise AI fails on unreadable data, not weak models**: This is the argument Trebels returns to most: decades of institutional knowledge sit locked in documents no system can read, so the missing piece is a semantic layer, not another model. Their standard examples are regulatory filings, clinical reports, product manuals and compliance records. They describe a compliance question whose answer existed in six document formats across three languages, with no system able to connect them. Trebels makes the point using other people's platforms, too: Dexory, which has scanned over a billion warehouse locations, says most physical AI projects die because the underlying data is bad; GKN Aerospace's automation technology director Andy Francis says the next factory gains will come from the terabytes of data its machines already throw off. Piers Linney's keynote claim that treating AI as an IT project is doomed, and his finding that only 30% of enquiries get answered properly because the data is not connected, get the same treatment.
- **A system you cannot trace, you cannot trust**: Trebels's recurring argument is that traceability is the price of trust in AI agents, and graph structures provide it. They report Jakob Freund at CamundaLocal London arguing for deterministic checkpoints inside otherwise autonomous processes, and object graphs as structured maps of what an agent touches, decides and connects to. Freund built his own agent with access to sales pipeline, travel and account data, and the first thing he added was a human approval gate. In Trebels's account, a live demo of an e-commerce returns agent that started issuing refunds was fixed not with better prompts but by separating concerns across purpose-built agents controlled by process models. They cite figures from Camunda's 2026 report: 84% of organisations cite business risk when IT lacks controls over AI in daily processes, 80% flag transparency gaps, and only 11% of agentic AI use cases reached production last year. The same reasoning underpins their commercial case for knowledge graphs, which they say give compliance teams auditability they can defend.
- **Buyers describe the problem, not the category**: Trebels argues that vendors' pitch language differs from buyers' language. After a climate intelligence forum, they note that nobody asked for a knowledge graph or help with "AI readiness": one attendee had years of safety logs archived in a form nobody could search, while another was matching claims against documents whose street names had changed since they were written. Both described a problem they were already paying for or had failed to solve. Trebels advises selling into those problems rather than technology categories, applying the same test to their own business: "If your AI strategy depends on data your systems can't actually read, that's the problem we solve."
- **Sovereignty and capability are not a trade-off**: Trebels unpacks the headline that Spain, France and Germany "banned" Palantir: Spain told state-owned firms not to sign new contracts, France's intelligence agency switched to a European alternative, and Germany's Constitutional Court ruled the legal basis for Palantir-powered police profiling unconstitutional. A procurement freeze, a vendor swap and a court limit are three different things. Beneath them, Trebels identifies two concerns: dependence on software governed by another country's law, and whether a state should let any vendor's system profile citizens before a crime exists. They cite Switzerland's defence agency rejecting Palantir because it concluded the risk was built into the architecture rather than fixable by contract, while noting the UK holds over £900M in Palantir contracts and Denmark has expanded its use into Greenland. Trebels concludes that where data lives and whether it can be audited are design choices, not fixed properties of the technology.
- **Technical capability outruns the commercial story**: Across two manufacturing events, Trebels reaches the same judgement: the engineering is ahead of the case made for it. At a Manufacturing Technology Centre robotics and automation event, they report many advanced implementations alongside repeated vendor grumbling about hesitation towards automation, delivered with a grim smile and the phrase "perhaps especially in the UK." The selling points, Trebels says, range from survival to filling the labour gap, but could be far more specific and effective. At a later physical AI day at the same centre, they record GKN Aerospace and Dexory making essentially the same argument back to them, and the launch of an MTC and Plug and Play partnership for advanced manufacturing.
- **Deep tech sells through sales enablement, not engineering detail**: Trebels treats translating engineering into buying language as a discipline in its own right. They describe a four-week TeamWorks programme with the University of Warwick and Monash University, in which a student team built an enterprise sales playbook for neuro-symbolic AI covering financial institutions, professional services, manufacturing, energy, logistics, public sector, defence, retail, e-commerce, consumer goods, pharmaceutical and life sciences, telecom, media, news and non-profits. Trebels says the team was seeded with research so it could spend its time finding gaps and validating rather than starting cold. They later brought in Rodrigo Corradi as a fractional sales enablement lead for three months to build that foundation: problem framing, repeatable messaging, objection handling and substantiation mapped to real processes.
- **AI adoption is a people-and-process problem**: Trebels collects accounts of adoption failing on the human side. From Elite Business Live, they report Julie Holmes naming "leading by proclamation"—telling teams to adopt AI without giving them time to learn it—and Dave Birss's finding that 93% of AI budget goes to technology against 7% to people, with 80% of adoption failures non-technical. From the climate intelligence forum, they report Andrew Ogden-Swift saying optimisers get switched off if operators do not accept them, and Chris Dean saying founders underestimate "the time it takes to get pilot to real." Trebels relates this to their own Toyota Kata training.
- **Sustainability sells as waste and margin, not ESG**: Trebels argues that the environmental pitch needs to be restated in business terms to land. They cite The Conference Board's tracking of how many of the 100 largest US companies put "ESG" in their sustainability report title—around 40% in 2023, 25% in 2024 and roughly 6% on a partial-year count—and quote Andrew Ngigi of Klimani saying businesses were not asking for help to reduce emissions but to reduce waste and protect margins. Trebels concludes that the focus should be business-compatible, with environmental benefits to follow.
- **Practitioner tooling, shared in full**: Trebels writes about the tools they use as an operator, not a reviewer. Two posts document a CRM setup in which Lightfield updates the CRM automatically from email and meeting transcripts, and Kondo does the same for LinkedIn direct messages. A Granola notetaker integration came later. They frame the change as moving from sales that relies on heroic remembering to a system that remembers for you, and publish the build: a webhook trigger, JSON parsing, an agent request step and the exact prompt they use. They also note that a bulk delete feature they requested was built quickly and call the vendor's support top-notch.

## Work to read

- [TextDistil and the enterprise AI data problem](https://www.linkedin.com/posts/etrebels_knowledgegraphs-enterpriseai-datapreparation-activity-7446892020893315072-y4Hd)
- [AI agents need traceable governance](https://www.linkedin.com/posts/etrebels_camundalocal-activity-7436735233066012672-8Z51)
- [A close reading of the Palantir sovereignty debate](https://www.linkedin.com/posts/etrebels_digitalsovereignty-datagovernance-aigovernance-activity-7479878175058132992-PtwS)
- [What a climate and enterprise AI forum actually revealed](https://www.linkedin.com/posts/etrebels_knowledgegraphs-growth-fractionalleadership-activity-7487624177357766656-QPVP)
- [Robotics, manufacturing, and the gap between engineering and growth](https://www.linkedin.com/posts/etrebels_mtc-manufacturing-robotics-activity-7473100179156578304-4C9F)
- [GKN Aerospace and the data problem behind physical AI](https://www.linkedin.com/posts/etrebels_physicalai-manufacturing-aerospace-activity-7476268077303869440-kSt5)
- [Building an enterprise sales playbook for knowledge graphs](https://www.linkedin.com/posts/etrebels_teamworks-mentoring-langoptima-activity-7486343145748725761-G2do)
- [A practical CRM and LinkedIn DM workflow](https://www.linkedin.com/posts/etrebels_gone-is-the-weekly-sales-meeting-of-shame-activity-7424475068665204736-3ucq)
- [AI adoption is a people and data problem](https://www.linkedin.com/posts/etrebels_elitebusinesslive-eblive2026-eb100-activity-7437620946200551425-cWZ-)

## Brand relationships

- **LangOptima** (owned_business)
- **Lead Semantics** (disclosed_partnership)
- **Lightfield** (affiliate)
- **Kondo** (affiliate)
- **Granola** (affiliate)

## Partnerships

The documented commercial arrangement is Edwin Trebels's promotion of third-party software through personal referral links and named discount codes. The recommendations read as personal product tests, not vendor copy. Conference attendance, event write-ups, university mentoring, and named collaborators are presented as observation or ongoing work, not sponsored or paid formats. For buyers, any advisory, content, or product-collaboration arrangement would need to be negotiated directly. The closest precedent for a vendor tie-in is a specific reader benefit, a plain statement of Trebels's interest, and an explanation of how the tool changed the workflow.

## Alternatives

- Viveta Gene, PhD: Choose them when the vendor is selling against a regulated-industry problem that hinges on precision in multilingual documentation, and needs an authority who can frame it as a knowledge-architecture problem rather than a translation-accuracy one. Their post on the July 31, 2026 fragrance-allergen rule walks through a single term expanding into 81 individual declaration requirements across 27 member states. It explains that sequential translation systems render "linalool" correctly but cannot query concentration thresholds by product type or jurisdiction. That makes the same compliance-auditability argument a knowledge-graph vendor needs, anchored in a concrete regulation and date. They also make the case to two audiences: enterprise localization practitioners at the TAUS EU Localization Championship in Rome, and academic peer reviewers at Convergence 2026 at the University of Surrey, where a co-authored paper with Prof. Vilelmini Sosoni sits in a track on LLMs supporting multilingual communication. Choose them when buyers want to see the approach withstand scrutiny from regulatory affairs and research reviewers, and the target vertical is cosmetics, life sciences or another regulated sector where naming and declaration rules drive the purchase.
- Keith Peiris: Choose them when the vendor wants a founder-operator who has built a company on the same premise and can explain to enterprise software buyers how the data layer is actually fixed. The Series A announcement describes creating Lightfield as a world model of a business, with $47M led by Andreessen Horowitz, more than 5,000 companies signed up since November, and the argument that Salesforce-era records are the wrong substrate for agentic work because the underlying data is incomplete and lacks structure. A second post details the mechanics: schema rewritten programmatically through APIs, migrations running in a persistent Python sandbox, a chronological timeline per customer to encode causality, and the claim that MCP access is not understanding. Choose them when the campaign goal is category education and product credibility with GTM and SaaS buyers, and the vendor benefits from an operator who can answer technical architecture questions in public rather than repeat positioning.
- Matt Serna: Choose them when the deliverable is adoption and customer-story content rather than methodology. A post describing a customer conversation gives specific, checkable details: accounts and research triggered by app installs, trust-centre visits used as an intent signal, product usage data from PostHog modelled in the CRM, about thirty automations running, and a sign-up flow that started as a fifty-block Zapier tree and was rebuilt in under a day. A second post reports asking 100 GTM leaders on Salesforce or HubSpot whether they loved their CRM and getting zero hands, then identifies fear of recommending something new and retraining a team as the dominant reaction. Choose them when the vendor wants field-level stories about what customers actually do with the product, or a speaker who can open a conference session on why adoption stalls. That maps directly to a moderated customer-interview series or an adoption-focused event format.

---
AI-assisted summary of public posts. Audience demographics are not available. Engagement appears only where we checked recent posts. Method: https://b2binfluencer.marketing/methodology/