Production-Ready AI Engineering · forward deployed
Data first, then AI. We make SAP, MES and PLM reachable, then build the operating layer on top. Eight weeks to production, inside your own tenant — and we operate it from there: monitoring, evals, observability and CI/CD for models.
Models, interchangeable behind one gateway
Questions this market actually asks
Pexon embeds senior forward deployed engineers inside industrial companies with €50 million to €2 billion in revenue. We start at the data layer, connecting SAP, MES, PLM and decades of documents, then build AI systems that respect your permission model, run inside your own cloud tenant, and are contracted under German law.
Why now
Between 2024 and 2026 most industrial companies proved that AI works. The demo answered the question. Then it stopped — at the data. Six reasons it stops, and all six sit below the model.
01 · Reachability
SAP, MES, PLM, thirty years of SharePoint, SQL nobody has touched in a decade. No prompt reaches into that.
The pilot never sees real data
02 · Entitlements
Your real permission model has to travel with the context. A system that shows everyone everything cannot go to production.
Security blocks the go-live
03 · Capacity
The one engineer who would push it through is committed to three other things, and not hireable in months on this market.
The project waits on one calendar
04 · Cost
Consumption pricing produces an invoice finance never approved.
Month four is four times the estimate
05 · Duplication
Four proofs of concept, four separate stacks, no shared governance.
No compounding, no scale effect
06 · Ownership
The pilot team moved on. No evals, no monitoring, no accountable name.
Quality drifts, silently
The consensus view
In the last twelve months every major AI organisation moved capital into embedded engineering. This is the strongest available evidence that the delivery model works.
May 2026
Stood up a dedicated deployment company, backed by over $4 bn, and acquired an applied-AI consultancy of roughly 150 forward deployed engineers to staff it.
2026
Built embedded engineering into its enterprise business on the same forward deployed model.
2026
$1 bn behind forward deployed engineering, including a programme to build the capability inside partners.
Q1 2026
Revenue +85 %, US commercial +133 %. A decade of embedded engineering, finally read as the proof.
12 months
Job postings for forward deployed engineers rose more than sevenfold, from 643 to 5,330.
All of it is aimed at the largest enterprises on earth. We are aimed somewhere else: industrial companies between €50 million and €2 billion in revenue, with the same deployment gap and nobody embedded to close it.
Sources, each re-verified before launch: OpenAI, 11 May 2026 (deployment company; Tomoro acquisition, ~150 FDEs; >$4 bn initial investment) · Microsoft, Frontier Company · AWS Partner Network, Forward Deployed Engineering · Forbes, 28 May 2026 (Palantir Q1 2026) · Indeed data via Business Insider (FDE postings, April 2025 → April 2026).
What we do
The asset in the middle is yours: your entities, your rules, your process knowhow. Raw systems on the left, agents doing real work on the right, and every one of them stops at a person who decides.
01
Extract, clean, mask, and carry your permission model through.
02
Frontier, open-weight or fully on-premise. No licence to defend.
03
Never leaves the tenant, never trains an outside model. Code included.
How it works
The failure pattern in enterprise AI is a wide pilot that touches everything and ships nothing. We invert it: one real data source, one real use case, all the way through to something a person uses, then widen.
01 · ACCESS
Access, data reality, permission model. Sprint goals in writing.
02 · BUILD
One source end to end. PII masked, access rights carried through.
03 · PROVE
Evals against your own questions, judged by your specialists.
04 · SHIP
Monitoring, cost control, runbooks. Handover starts here.
05 · ONGOING
More sources and business units, or transition into operations.
Each engagement makes the next one cheaper for you, not just for us.
That is the point of doing this repeatedly in one industry. The connectors, masking pipelines and eval harnesses we bring with us were paid for by earlier work, and you inherit them on day one rather than funding them again.
Maintenance lead
Which torque spec applies to the flange on the 2019 revision, and has it changed since?
Assistant
420 Nm, unchanged for the 2019 revision. The 2021 revision raised it to 460 Nm, so check the plate before you work.
Assembly_Manual_R2019.pdf · p. 148
ECR-2021-0442 · approved 14 Mar 2021
Maintenance lead
Show me the change request.
Service engineer · Plant 04
What did we quote this customer on the last three orders?
Assistant
You can see two of three. Order 8841 sits under a contract your role does not cover.
8839 · €41,200
8840 · €38,750
8841 · restricted
Entra ID group: SERVICE-P04
Row filter applied before retrieval, not after
Agent · invoice inspection
Freight invoice 77-2291 checked against the rate card. Two lines deviate.
Agent
Line 4 overcharged €280 against the agreed tariff. Line 7 lacks a matching delivery note.
Ratecard_2026_Q2.xlsx · row 219
Awaiting human approval before the debit note goes out
Accounts payable
Approved. Send the debit note.
EVERY ANSWER CARRIES ITS SOURCE · YOUR TENANT
ROW FILTER APPLIED BEFORE RETRIEVAL · YOUR TENANT
AGENT PREPARED · HUMAN APPROVED · YOUR TENANT
In production
The assistant prepares, cites and stops. A person approves. No extra licence per head, no context retyped, and nothing leaves your tenant.
Pick a claim to see the exchange behind it
Every response cites the document and page it came from, so your specialists verify in seconds instead of trusting. When the corpus cannot support an answer, it says so rather than inventing one.
Rights come from Entra ID and are enforced before retrieval, not filtered afterwards. A user sees exactly what their role already allows, down to the row.
Routine inspection and preparation run unattended. Anything with consequence stops at a named person, with the full trail attached, ready to approve or reject.
The outcome
Every enterprise runs four infrastructure layers. ERP for operations. CRM for customers. Identity for access. Cloud for compute. A fifth is being built now, and unlike the other four it cannot be bought off a shelf, because it is made of your own knowledge.
Your engineers stop searching for the drawing. Your service desk stops re-answering the same ticket. Your quality team stops reconstructing why a decision was made in 2019. The knowledge was always there. It was just unreachable.
Y Combinator named “Company Brain” one of fifteen categories in its Summer 2026 Requests for Startups. The through-line: the missing piece is company context, not model capability — and what gets sold is the finished work, not the tool.
Data-to-AI use cases
Eight patterns, each starting from a system you already own. Each begins as a two-week blueprint — none begins as a twelve-month programme.
1 answer
Head of Sales · SAP
Governed semantic layer
hours
Plant Manager · MES
Machine context joined to QA
no rework
Head of Engineering · PLM
Part and revision aware search
L1 + L2
Head of Service · Tickets
Confidence threshold and review
no surprises
Legal and Procurement
Clause-level obligation tracking
−1 FTE
Head of Quality
Inspection against specification
unblocked
CIO · Legacy
Business rules extracted, tests generated
condition
Head of Maintenance
Feature pipelines on failure history
Eight patterns, each starting from a system you already own. Each begins as a two-week blueprint, none as a twelve-month programme.
Delivered
Each figure is taken from the case study linked below, where its scope and measurement basis are stated. None of them is a market study or an industry average.
References
Clients are described rather than named. Each card links to the case study, which states the scope of the work and how the figure was measured.
800+
−90 %
+40 %
700 h
−80 %
3 tracks
Sovereignty as a spectrum
PragmaticMaximally sovereign
Level 1
Level 2
Level 3
Level 4
You do not have to decide today. The platform is built so models and location of operation are interchangeable: one gateway in front, one governance model behind. What runs in the EU cloud today runs on your own GPUs tomorrow, without rebuilding the applications.
Two buyers, one project
AI agencies deliver only the application. System integrators deliver only the infrastructure. We deliver both, because one without the other does not reach production.
The path
Levels 2 and 3 in detail
Level 2
Inside your network, in production in 6 to 8 weeks, at a fixed price. Every answer carries a source citation back to the original document, so your specialists can verify instead of trusting.
Typical first projects: knowledge assistant on technical documentation, bid and tender analysis, standards and regulation research.
Level 3
SAP, ERP, PLM, file shares, ticket systems. Order status, bills of material and price levels where they originate rather than in a copy. From here AI replaces manual work instead of simulating it.
Typical extensions: document inspection against specification, freight invoice checking, order intake capture, service reports.
The second use case is cheaper than the first. The fifth costs a fraction.
The architecture
Pro-code, not prompt writing. This is the reference architecture your installation is derived from, and the choice per layer follows your data classification rather than our preference.
01
Chat surface, business apps, Office integration
02
Orchestration of tasks and processes
03
Standardised access to your systems over MCP
04
One entry point: cost, limits, filters, logging
05
Interchangeable behind the gateway, per sovereignty level
06
Vector search plus our own ETL for indexing and embedding
07
EU cloud, your data centre, air-gapped
Model versions are deliberately not named here. The choice is made at project time by request class and data classification, with evaluation on your own data rather than a public benchmark.
The gateway layer
Model choice stops being a procurement decision and becomes a routing rule. Swap a provider without touching a single application.
Every provider behind a single contract and a single call. Your applications never learn a vendor name.
Frontier, open-weight, on-prem
When one provider degrades, the gateway routes to the next without an incident ticket.
No single point of failure
Cheap models for cheap questions, the expensive one only where it earns its cost. Measured, not guessed.
Caching, batching, right-sizing
Fine-grained rules per data class. A prompt carrying personal data never reaches a provider you have not cleared.
Enforced at the gateway, logged
Integrations
We connect to the systems your people already work in. No ripping, no replacing, and no parallel user directory. SAP and ERP read where the data lives, Entra ID carries the rights, and thirty years of SharePoint gets indexed with its permissions intact.
Security and compliance
Credentials
Held by Pexon Consulting GmbH
Pexon Group, across all five companies
The certification and the partner registrations above are held by Pexon Consulting GmbH, and both figures describe Pexon Group as a whole. This site is published by PEXON ROMANIA S.R.L., the group's AI engineering hub in Cluj-Napoca, registered in 2026: a separate legal entity, the party you contract with here, and the holder of none of them — which is what the list above means by “not a certificate this company holds”.
Why Pexon
Segment
The large deployment teams are built for global enterprises, and their published references show it. We are built for companies between €50 million and €2 billion in revenue — the segment with the same problem and none of the attention.
Neutrality
We are not selling licences underneath the engineering. Model and cloud decisions get made on your requirements and your data classification, including the answer that nothing leaves your building.
Accountability
One named senior architect, German-speaking, personally accountable for delivery. Contracts, SLAs and the data processing agreement under German law, with PEXON ROMANIA S.R.L. as the contracting party.
Entry
A two-week blueprint, not a programme approval. If it does not convince you, you have lost two weeks and a rounding error.
On determinism, the question most offerings avoid: almost every company-brain product optimises for retrieval over unstructured text. When an agent and the CFO calculate “revenue” differently, wrong answers arrive silently in board material. For regulated and industrial clients, traceability beats answer coverage — metrics come from defined sources with defined definitions, not from free-text similarity search.
How to start
2 weeks
€4,900 fixed price
Most chosen
6 to 8 weeks
from €15,000 fixed price
8 to 12 weeks
Price in conversation
No upfront investment in a platform. The foundation is built underneath the first use case, not before it.
Once the first use case is live, capacity is the only variable. Start with monitoring and add engineers when the backlog justifies them.
Monitoring
from €3,000
per month · no FTE
1 FTE
€7,500
per month · one engineer
2 FTE
€14,500
per month · pod of two
3 FTE
€21,000
per month · full pod
| What you get | Monitoring | 1 FTE | 2 FTE | 3 FTE |
|---|---|---|---|---|
| Availability, cost and eval monitoring | yes | yes | yes | yes |
| Pipeline fixes per month | up to 3 | unlimited | unlimited | unlimited |
| New sources connected per quarter | – | 1 | 2 to 3 | 4+ |
| Named forward deployed lead | shared | 20 % | 40 % | dedicated |
| On-site days per month | – | 1 | 2 | 4 |
| Parallel workstreams | – | 1 | 2 | 3 |
| Monthly, month to month after the first cycle | from €3,000 | €7,500 | €14,500 | €21,000 |
Month to month after the first cycle, and a tier change takes effect the following month. Consumption is never marked up.
Cloud consumption, model tokens and any specialised hardware are billed directly to your own accounts so you see them. No licence, no per-seat fee, no success fee, and no margin on your consumption. All prices are net and exclude VAT.
Operations
The operating model is switchable rather than built into the contract. You can decide again every year without touching the platform.
Full managed service with an SLA. Your IT does not carry another system.
Accountability with Pexon
Handover, training and on-call. Your team leads, we are reachable when something breaks.
Accountability shared
Full documentation, operations handbook, runbooks. No residual contract, no tie back to us.
Accountability with you
Inside the engagement
Weeks 1 to 2
Access, target picture, permission concept, alignment with information security.
Weeks 3 to 5
Open up the sources, preparation, retrieval, first usable version.
Weeks 6 to 7
Specialist evaluation with your test users, quality tuned against their verdict.
Week 8
Production release, documentation, cost model, decision paper for your committees.
The team
Not the proposal of a partner you never meet. The people in the first call are the people on the project, and they are still there in month nine.
Procurement and legal
The forward deployed lead works on site or hybrid by agreement. For regulated programmes we offer a dedicated engineer four days a week. The delivery team works in your time zone and inside your tools, whether that is Jira, GitHub or GitLab.
A two-week readiness blueprint is €4,900 at a fixed price. One use case in production is a fixed price from €15,000 over six to eight weeks. After that, operations are priced per engineer: from €3,000 a month for monitoring alone, €7,500 for one engineer, €14,500 for two and €21,000 for three. All prices are net and exclude VAT.
Keep it where it is enough. Copilot is a licence per head for standard cases. What we do is the work that has to run on your own data, inside your own governance, with a consumption cost you control. Azure AI Foundry is a construction kit rather than a finished platform, and we build on it rather than against it.
You do. Code, data pipelines, prompts and documentation are yours, and usage rights to sprint output transfer on payment of the first invoice. Generic building blocks we bring with us — connectors, masking pipelines, eval harnesses — remain ours and are licensed to you perpetually for this deployment.
Everything runs inside your Azure, AWS or GCP tenant, using your identities, in your regions, on your cost centre. Our engineers work with named accounts under your identity provider, scoped to the sources in the current sprint. Access is logged and revocable by you at any time.
That is the goal, and it is stated in the engagement plan. Handover begins in month two: runbook, architectural decision records, and shadowing until your team ships changes without us. There is no residual contract and no tie back to us.
Cloud consumption, model tokens and any specialised hardware, billed directly to your own accounts so you see them. Travel for on-site days is agreed in advance. No licence, no per-seat fee, no success fee, and no margin on your consumption.
Yes. Operating agreements run month to month after the first cycle, and a change of tier or operating model takes effect from the following month. That is possible because the runbook and the operations handbook are always current and always with you.
The contracting party is PEXON ROMANIA S.R.L., registered in Cluj-Napoca. Contracts, service levels and the data processing agreement are governed by German law, which is what procurement in this market is set up to review. Those are two separate facts: the counterparty is Romanian, the law is German.
Because you get the people who build it, not the proposal of a partner you never meet. And because the dependency risk is contractually excluded: full code and infrastructure transfer, a documented exit clause, operable without us at any time. For a nine-figure group programme that needs board-level air cover, we are the wrong partner, and we will say so in the first call.
Blog and resources
Engineering notes and case studies from the work above. Published in English.
Herbert James Draper, Ulysses and the Sirens, 1909 · public domain
Case study
A LangGraph multi-agent system answering DIN and ISO questions for 300+ engineers, with pseudonymised tracing and a published accuracy split by query type.
Case study
Custom CNNs on Databricks with dual detection for known defects and unknown anomalies, feeding a self-service WebGUI with real-time alerting on the line.
Case study
Three parallel model tracks with half-yearly upgrade windows, a Terraform module library, and Kubernetes right-sizing that keeps bank APIs off GPU nodes.
Private AI
What four-bit and three-bit quantisation cost in accuracy, the KV cache and power arithmetic behind the decision, and how offline retrieval stays permission-aware.
More engineering notes and the full case-study index:Browse the blog
Browse by topicClaudeData foundationPrivate AI
Data first.
Then AI.
Built inside your tenant, documented, operated. With a bill your CFO can defend and an audit trail your CISO signs.
How to start
Thirty minutes with the architect who would actually run the engagement. You leave knowing whether this fits and what the first source should be. If your pilot is still a pilot, the problem is probably reachable.
Noel Dinger
Administrator and Managing Director
PEXON ROMANIA S.R.L.
Calea Dorobanților 14–16, Etaj 2, Ap. 62
400117 Cluj-Napoca
Județul Cluj, Romania