ARBA Nexus
Nexus Enterprise AI
One AI across the whole company: its data, documents, projects and rules
- ARBA International B.V.
- Delivery — Telecon and Impace Group
ARBA Nexus
What we mean by enterprise AI
Nexus is one place where an employee states a task in plain words and the system carries it out on the company’s data, its documents, its projects and its rules. One way in, instead of a shelf of assistants that each know one corner of the business.
It reads the systems you already run, puts the data in order where the task requires it, and works on top of them — with the authority of the person who started it, evidence under every conclusion, and a human decision wherever something material changes.
One set of permissions, one audit log, one body of evidence — across the whole company
Boundaries
Where the product’s boundaries run
In an enterprise conversation this is the first question: what are we buying, and what are we not throwing away
What Nexus is
- A place where enterprise AI scenarios actually execute
- Managed data planes of its own: master data, documents, projects
- One layer for authority, evidence, checkpoints and cost accounting
What it is not
- Not a replacement for your ERP, your document management, Microsoft 365 or your project systems
- Not a chatbot that answers beside the process and owns none of it
- Not a black box that first demands perfect data
Nexus reads the estate you already have, puts data in order where the scenario needs it, and runs the work on top
Differences
How this differs from what you have already tried
Each of these approaches does part of the job, and each can be extended with integrations. The question is where it stops in a standard deployment
| Approach | Where it stops | What Nexus does |
|---|---|---|
| Office-suite assistant | Out of the box it knows one person’s mail and files, not your master data or your rules | Runs on governed company data and acts in your systems — within that person’s authority |
| Document search | Without further integration it returns a quote and stops there: no task, no owner | The answer is a step of a task, with an owner, a checkpoint, a trace and a cost |
| Master-data programme | A separate project on a long horizon: one budget pays, another function sees the result | Order is created inside the work itself and pays for itself on the first domain |
| Bespoke development | In a typical point solution every further process is a new project, a new stack and a new contract | Your own people describe the scenario; the library of capabilities grows on your side |
Mechanics
How one answer destroys trust in the whole technology
Automation amplifies what is already there. A language model will not warn you that the data contradicts itself — it will produce a fluent, confident and wrong answer
- The data contradicts itselfThree records for one counterparty, mixed units, dead codes still in use
- The model answers confidentlyFluent text, precise figures, no hedging
- The figure enters a decisionIt is shown to the board
- The error surfacesIn public, during review
- The subject is closedTrust is lost not in the data quality but in artificial intelligence as such
There is nothing left to defend the next initiative with — and that costs more than the error itself
The price
Bad data costs money long before any AI
AI creates none of this — it only makes the consequences visible and fast
- $12.9Maverage annual loss per organisation from poor data quality4
- 63%of organisations either do not have AI-ready data management practices or are not sure they do3
The answer
Putting data in order is the first job, not the entry ticket
Most AI products open with a condition: get your data in order and then we will talk. Nexus is built the other way round — creating that order is the first piece of work it takes on.
The intelligence sits on the hardest part rather than on the surface: matching records, resolving duplicates, normalising descriptions, discovering rules and testing quality. Nobody works through thousands of rows by hand — people decide on prepared candidates that come with evidence.
The first bounded domain is scoped to go live in weeks
Honest about maturity
Where the product stands today
Three live installations at different stages of industrial use: at a customer — a major copper producer in Central Asia, at a delivery partner in Kazakhstan, and at ARBA itself, where the vendor runs its own processes
Running on real data
- Master data — golden records from the customer’s ERP: counterparties and units of measure
- Document store — a corporate corpus with recognition, cited search, retention schedules and legal holds
- Project control — the owner’s assurance picture over contractors’ documents on a capital project
Rolling out now
- Master data at full domain scale — validation continues
- Service desk — a pilot for named groups in Microsoft Teams with governed status write-back
Next stage
- Service desk beyond the pilot groups
- Further master-data domains on the same foundation, without rebuilding it
As of 3 August 2026. Writing into ERP and finance systems stays behind a human decision — by design, not by technical limitation. Which capabilities are switched on is decided per installation
Execution spine
Every task travels the same path
- RequestFrom chat, a channel or a ticket
- TaskA tracked object, not a one-off answer
- AuthorityThe agent acts for whoever started it; rights are checked at every step
- ToolsA reasoning loop inside a permitted toolset
- CheckpointMaterial records wait for a human decision
- TraceEvery step recorded, every cost counted
The same path for a procurement scenario, a master-data cleanup and a service request — governance is not bolted on top, it is built into execution
Platform
One foundation, four components
Core
- Scenarios — processes the AI carries out by reasoning, not by a rigid script
- Tools and connectors — the platform’s governed hands
- One execution spine: authority, evidence, checkpoints, tracing, cost accounting
Data planes of its own
- Master data — golden records, quality, hierarchies
- Document store — recognition, classification, cited search, retention
- Project control — an assurance picture across projects, on top of your systems
Your estate
- SAP · Oracle · 1C · Microsoft · Primavera · in-house systems
- Read as sources, governed on top, never replaced
The components run on one core: access, logging, evidence and cost accounting are configured once, not again for every task
Level 1
A master-data domain in disarray
An illustrative master-data scenario: the same engineer’s request, the same company, the same people and systems. The only difference from the next frame is the maturity of the domain
- RequestAn engineer needs a bearing for a stopped assembly
- SearchThree similar records, none an exact match: different descriptions, old supplier codes, different units
- DecisionEasier to create a new item than to prove the old one is the same thing
- OrderOrdered with a margin: lead times are long and nobody wants to repeat the cycle
- DeliveryMonths later the item reaches the warehouse
- OutcomeSome goes into the job, the rest settles on a shelf. The same bearing is already in stock at the neighbouring site — but it cannot be found
And round again, at every site
Cash is frozen in warehouses, procurement is loaded with repeat cycles, and the stock still is not where it is needed
Level 2
The same domain, assembled
- RequestThe same engineer, the same assembly
- SearchOne record: synonyms, old supplier codes and other sites’ designations all resolve to it
- AvailabilityStock at the neighbouring site and open requisitions for the item are visible at once
- DecisionAn internal transfer instead of a purchase — the assembly restarts within days
- ReplenishmentThe issue is booked, the minimum-stock rule fires, procurement automatically receives a replenishment request
- ControlNo new item is created: the system will not accept a duplicate without review and an owner’s decision
The same process, the same people, the same systems. Only the state of the master-data domain changed
A domain is not one reference list: a material lives together with its classification, units of measure, manufacturers and supplier codes. The unit of work is the whole domain — a material list cleaned on its own changes nothing
Interface
The conversational console
What the screen shows
- 1The task plan and the steps already taken
- 2Sources and tools used
- 3The checkpoint and the human decision
- 4The trace and the cost incurred
Trust
What makes such an AI fit for industrial operation
Evidence
Every fact carries a link to a document, a record and a version. «Not enough data» is a complete answer, not a failure
Full trace
Who asked, which sources and tools were used, what the human decided. The work can be replayed
Human confirmation
A material change passes a checkpoint: the work stops and waits for an authorised decision
On the user’s behalf
Nexus checks the starter’s authority before reading data and before any action — the agent holds no rights of its own
Sovereignty
The data plane stays inside your perimeter, up to full isolation from external networks
Your model
Inference goes through your endpoints; the egress latch rejects calls to addresses outside the allow-list
Economics
AI whose running cost is visible and can be capped
Every call is counted
Cost is attached to a task, a scenario and a person — you see what a specific process costs, not «AI» in general
A ceiling as a fuse
A daily spend limit stops the work before the bill becomes an unpleasant surprise
The model fits the job
Different tasks route to different models: some need the strongest, some are fine on a fast and cheap one
First stage
Test the machinery on one uncomfortable domain
Not a year-long programme but a bounded test on real data — with the decision agreed in advance
- From you
Input
- An extract of the chosen domain
- A data owner and subject-matter experts
- Quality criteria and constraints
- An agreed perimeter
- Together
Process
- Profiling
- Candidates with evidence
- Decisions by your specialists
- Rules and trace recorded
- Output
Result
- A measured picture of quality
- A reviewed set of candidates
- Golden records and rules
- Architectural conclusions for scale
The stage ends in a joint decision: continue, adjust or stop. Scope, criteria and calendar are fixed once we have seen the data
Who does what
Where responsibility sits
A baseline for discussion: exact boundaries and the partner line-up are fixed per project
- Product
ARBA Agency
- Core architecture and development
- Releases and product support
- Methodology for scenarios and data planes
- Delivery
Telecon and Impace Group
- Local delivery and project management
- Integrations and environment preparation
- Training and change support
- Ownership
Customer
- Domain owner and success criteria
- Access, policies and perimeter decisions
- Approval of material changes
Responsibility for a material change to data stays with the customer in every configuration
Next step
Start with one domain
There is no need to start with a year-long programme. Take the single most painful domain — materials together with their classification, units of measure and supplier codes, for example — and watch the machinery work on it: extract, profiling, candidates with evidence, decisions by your specialists, a golden record.
The domain is taken whole not out of perfectionism: the effect comes from the links, not from one cleaned record. From there the same foundation carries the next domains, documents, projects and scenarios — with no second build.
Discuss a review of your master-data domainOn request
Appendix
The material usually needed after the meeting rather than during it: the evidence base, the mechanics and the depth of each plane
- The market and the research: three figures from 2026 and two field experiments
- Master data in depth: from extract to golden record and beyond the first cleanup
- The platform’s planes: documents, records management, project control
- The research core and the effect of every new connector
- Operation inside your perimeter, and the source list
Where the market stands
Three figures that describe 2026
One observation across a sample of public deployments and two Gartner forecasts. None of the three is about model quality
- 95%of generative-AI pilots in a sample of public deployments produced no measurable effect on financial results1
- 40%+of agentic-AI projects will be cancelled by the end of 2027 — Gartner’s forecast: cost, unclear value, inadequate risk control2
- 60%of AI projects unsupported by AI-ready data will be abandoned by the end of 2026 — Gartner’s forecast3
The trap
How companies arrive at those figures
AI for its own sake
A solution is chosen because it has artificial intelligence in it, not because of the job it closes
Sprawl
Every pilot brings its own infrastructure, its own access model, its own log and its own contract. The estate grows faster than the benefit
One pilot, one task
A narrow context: the tool knows nothing about the neighbouring process and does not remember yesterday’s conversation
Stopping at the edge of production
The demo passes. The test against real data, real rights and real risk does not
Diagnosis
The divide does not run along the models
The MIT NANDA study examined 300 public deployments, executive surveys and a series of interviews — and found no link between success and the choice of model. What separates the successful 5% is different: they embed AI in a real working process, give it memory and company context, and carry it through to production.
The rest buy a general-purpose tool that is convincing in a demo and helpless in the flow of work.
The question is not which model, but what it stands on
And yet
Opting out of AI is not an option
The need is real. AI is becoming the difference between a company that decides in hours and a company that spends weeks assembling the data for that decision.
So the question is not whether to adopt AI but what to build it on — because the foundation is what decides whether adoption reaches production.
Two field experiments
How AI changes work: two measured cases
Two field experiments: different tools, different tasks, different metrics
An assistant trained on company data
A purpose-built support assistant trained on the history of customer conversations · 5,172 agents, The Quarterly Journal of Economics, 2025
- Issues resolved per hour, on average+15%
- Among less experienced agents+30%
- Fine-tuned on customer conversations weighted by outcome — that is, on the practice of the best agents — and it surfaces links into internal documentation
- Those practices transfer to less experienced staff and lift them towards the average; for the most experienced the gain is speed alone, and quality dips slightly
A general model in an expert’s hands
General-purpose GPT-4 on realistic consulting tasks · 758 BCG consultants, 18 assignments, Harvard Business School
- Tasks completed+12.2%
- Time per task−25.1%
- Assessed quality of output+40%
- That is inside the model’s frontier. Beyond it the same consultants were 19 pp less likely to reach the correct solution
The two studies cannot be compared directly by effect size. Together they show that the result depends on the task, the data, the model’s frontier and the human’s role
The metrics differ between the studies and the bars are normalised within each column: the columns are compared by meaning, not by height
The condition
What augmenting an expert is made of
Remove any one term and augmentation becomes the accelerated production of confident errors
The expert works faster and stays right
The root
Bad data is a bad foundation for automation
Automation amplifies what is already there. If reference lists disagree and one counterparty lives in three systems under three names, automation will amplify exactly that.
And a language model will not flag the contradiction. It will produce a fluent, confident, wrong answer.
Why it still is not done
The classical route to governed data costs too much
An enterprise master-data programme is a project in its own right: its own team, its own catalogue, its own standards and a long horizon. The result is visible to someone other than the budget holder, and not soon.
So it is deferred, again and again. Order in data is needed by everyone at once and by no one in particular — no single function will fund it whole and wait longer than its own planning horizon.
A data foundation everyone needs and nobody builds
How it works
From extract to golden record
- 1SourceAn ERP extract or a customer file is fixed as an immutable source version — you can return to it and repeat everything
- 2ProfilingNexus shows what is actually in the data: completeness, formats, anomalies, disagreements in units of measure
- 3RulesIt proposes normalisation and matching rules that account for legal forms, Cyrillic and Latin script and legacy codes — and tests them against history
- 4CandidatesIt assembles duplicate pairs and groups with an explanation: which attributes matched, which diverged, what is missing
- 5Human decisionThe owner confirms or rejects. The AI proposes; the person disposes
- 6Golden recordThe surviving record is assembled: you can see which value won each field and why. The merge is reversible
- 7PublicationGoverned distribution of the golden record to consuming systems. Writing back into the source stays behind a human decision
Master data in depth
A domain lives on past the first cleanup
A one-off cleanup degrades within months. So the platform carries what holds quality continuously
Quality under observation
Rule-based profiling, statistical anomaly detection, validation against classifiers and normalisation proposals — as an explained queue of work, not as silent edits
Hierarchies and reorganisations
Groups, nodes and roll-ups across a tree; moving a node does not erase history, it opens a new period of validity
Memory over time
Bitemporality: «how it actually was» and «how we knew it» are kept apart. Correcting an error does not rewrite the past, it records new knowledge
Publication and divergence
The golden record goes to consumers as a governed package; a separate view shows where master and source have drifted apart
Interface
Resolving duplicates
What the screen shows
- 1The source records and their systems of origin
- 2Attributes that matched and diverged
- 3The reasoning behind the recommendation
- 4The data owner’s decision
Research core
What separates the core from an «AI construction kit»
Nexus is not a wrapper around someone else’s model or an assembly of off-the-shelf blocks. Underneath are ARBA’s own research results, shared by every component of the platform
Global context
The system holds a picture of the whole corporate corpus rather than the window of one document: an answer rests on the links between sources
Parallel execution
A complex task is decomposed and run in parallel by several agents, then reconciled
Knowledge on a live graph
Knowledge is attached to a model of the organisation — entities, roles, ownership, hierarchy — not to flat folders
Adversarial verification
A result passes self-checking and counter-verification before it reaches the user
Memory over time
Bitemporality: the system remembers both how things are and how they were known at any point in the past — an answer is reproducible to a point in time
Hybrid methodology
Deterministic, semantic and probabilistic methods work together — engineering rigour instead of prompt magic
Surfaces
Nexus lives where people already work
Conversational console
One way in: state the task in words, watch the work, the evidence and the decisions
Assistant over the screen
Answers in the context of the page you have open and points at the very field in question
Microsoft Teams
Requests, answers and approvals in the channel people already use, with no separate portal
API
The same capabilities for your own systems — with the same rights checks and the same tracing
Scenarios and Studio
A new capability is described by your employee, not by a contractor
A scenario describes a process: what result is needed, which tools may be used, where a checkpoint is mandatory and whose authority is required. The agent does not follow a rigid script; it reasons inside those bounds.
An employee describes the process in words, Nexus proposes a governed scenario, the owner reviews and activates it. The library of capabilities grows on your side, not ours.
Customers create scenarios; they do not create authority for themselves — that is checked on activation
Global context
Why Nexus knows your company from the first question
- Level 1Map of the corpusWhat exists at all: collections, volumes, owners, links
- Level 2DigestsA summary of every document and of groups of them, recomputed as things change
- Level 3Cross-cutting observationsFindings visible in no single document on its own: contradictions, gaps, changes between revisions
- Level 4Standing backgroundAll of the above is mixed into every answer — within the rights of the individual asking
This is not loading every document into the prompt: the map, the digests and the observations go into the work, not the corpus itself
Documents
From a scan to an answer with a citation
- IntakeUpload, a folder, SharePoint or mail
- RecognitionRussian, Kazakh, English — paper stops being an obstacle
- ClassificationDocument type and attributes extracted automatically
- LinksThe document attaches to a counterparty, an asset, a project
- AnswerA quotation linked to the document and the page
The main reason corporate archives decay is manual attribute entry. Here an agent does it and a person confirms
Records management
What counsel and the auditor will ask for
Retention schedules
Governed rules: a document knows its retention period and the date it starts
Legal holds
A hold outranks retention: while a dispute runs, the document cannot be destroyed
eDiscovery
Selection by the subject of a dispute and a sealed export of the inventory with version and access history
Read log
You can see who opened a document and when — not only who changed it
Separation of duties is mandatory: whoever places a hold does not destroy, and whoever destroys does not place holds
Project control
Reported status against supported status
The project-control plane does not replace your execution systems — it reads them and shows the owner where the report and the evidence have diverged
The gap is what the plane exists for
Schematic: the ratio is shown for clarity. In the product both figures are computed for a specific project period, and every deviation enters an exception queue with an owner
The effect
Every new connector strengthens every scenario at once
This is where the logic of a zoo of pilots breaks: pilots add up, connectors multiply. Connect the warehouse and procurement, maintenance and project control all see it — not just a «warehouse assistant»
| Nexus core6 connectors · 5 scenarios | Procurement | Maintenance | Project control | Back office | Support |
|---|---|---|---|---|---|
| ERP | available | available | available | available | available |
| Warehouse | available | available | available | available | available |
| Documents | available | available | available | available | available |
| Mail and Teams | available | available | available | available | available |
| Projects | available | available | available | available | available |
| Service requests | available | available | available | available | available |
Instead of a set of assistants — one enterprise AI that holds both the operational and the strategic picture
Sovereignty and operation
How this lives inside your perimeter
Separate installations
One product image, separate independent installations: own database, own settings, own tenant account
Pull-only updates
The installation fetches a signed release when it chooses. Nothing is pushed into the perimeter from outside
Full isolation from external networks
Running with no external connections is supported as standard: the platform keeps working from packages already received
Release discipline
A weekly release cycle; a build is not accepted until it passes end-to-end validation against twenty business scenarios
Release cadence and the composition of end-to-end validation as of 3 August 2026
Evidence base
Sources
- [1]MIT NANDA. The GenAI Divide: State of AI in Business, 2025About the effect on financial results across a sample of public deployments — not about whether the pilots technically worked.
- [2]Gartner. Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25.06.2025A forecast, not a measurement. Gartner names the causes itself: escalating costs, unclear value, inadequate risk controls.
- [3]Gartner. Lack of AI-Ready Data Puts AI Projects at Risk, 26.02.2025The forecast is conditional: it covers projects not supported by AI-ready data, not AI projects in general.
- [4]Gartner. Data Quality — the cost of poor data
- [5]Brynjolfsson E., Li D., Raymond L. Generative AI at Work. The Quarterly Journal of Economics, 2025
- [6]Dell’Acqua F. et al. Navigating the Jagged Technological Frontier. Harvard Business School / BCG
Every quantitative claim in this deck rests on the public research listed here