AI that shows where its answer came from, and stops before deciding
AI-native does not mean a chat window in the corner. It means intelligence working inside the flow, over the document in front of you, through the same permissions a person has — and every conclusion carrying the data, the explanation and the path to the source document. What it cannot support, it does not claim.
- Answers with a source
- Proposals as drafts
- The same permissions as a person
- An append-only trail
Why this is not another chat window
The usual 'AI in the ERP' is a panel that paraphrases what is already on the screen. It is useful for about a week, which is how long the novelty lasts.
The problem with AI in a business system is not that it cannot answer — it is that it answers when it does not know. A sentence that sounds right with no data behind it is not an inconvenience in an ERP; it is a wrong posting, a wrong dunning letter and a decision made on a figure nobody checked. So here an answer without a source is not an answer: every conclusion carries the data, a short explanation and a drill-down to the line it came from.
The other half of the same problem is autonomy. A system that sends a dunning letter to the wrong customer does not save time, it spends it twice — once on the mistake and once on the apology. So everything the AI proposes is a draft: it enters the same workflow, the same approval limits and the same trail as a decision made by a person. Autonomy levels exist and they do rise — an agent proposes, acts on confirmation, or acts on its own within limits — but they are switched on per process and per amount, in the console, and they always come back. An agent's limits are the same approval limits a person has. Human-in-the-loop is the default, not the other way round — a decision about the product rather than a limitation.
The third is access. The AI has no privilege for 'being the system': it goes through the same permissions, the same data scope and the same tenant isolation as the person who invoked it. The RAG index is built per tenant, under the same database policy as the data itself. Every call stays in an append-only trail — input, model and prompt version, output, and who confirmed it.
Three things the AI does here, and one it does not
These are not values but constraints built into the model. Each is visible on screen and verifiable in the trail.
Evidence
Every conclusion arrives with the data, an explanation and the source documents. Not 'the customer owes 2.4M', but '2.4M RSD across seven open invoices, three of them more than 30 days late' — with the list and a drill-down to each. An answer that cannot be supported by a source is not shown.
Drafts
Dunning letters, purchase orders, postings and replies arrive as drafts for review, not as executed actions. A draft passes through the same workflow, the same limits and the same trail as a human decision — the AI has no side entrance into the system.
Context
The panel works on the screen you are on and the object you are looking at: explain this posting, check the VAT treatment of this line, find the goods receipt behind this invoice. Concrete actions on a concrete document, instead of an empty box asking you to think of a question.
And not — access outside the rules
The AI sees exactly what the user who invoked it sees: the same permissions, the same data scope, the same tenant isolation. Isolation and prompt-injection tests run in CI as a standing check rather than a one-off. When that check fails, the change does not enter the branch.
What the AI actually does
Five things the intelligence layer does in daily work. Each has a screen, a data source and a trail.
A copilot that answers with evidence
Not “this customer owes 2.4M”, but “this customer owes RSD 2.4M across seven open invoices, three of them more than 30 days late” — with those invoices listed and drill-down into each. An answer that cannot be backed by a source is not shown.
A daily brief that starts at the decision
AI Daily Brief on My Work: revenue above last month, three of your largest customers late, an item running out in six days — with two purchase orders proposed. Every conclusion carries an explanation and a path to the data it was derived from.
OCR that is allowed to say “not sure”
Extraction is measured per field type — supplier, invoice number, total, VAT and line items separately — and a field the system is unsure about is flagged rather than guessed. A confidently wrong field costs more than an empty one.
A suggestion is a draft, not an action
AI prepares reminders, purchase orders, tasks and replies as drafts that enter the same workflow as a human decision — with the same limits, approvals and audit trail. Autonomy is enabled per process and per amount, and can always be stepped back.
Anomalies and forecasts with a stated reason
NG One forecasts cash flow and demand, scores collection risk before a document is approved, and detects deviations in postings, duplicate invoices and unusual discounts. Every flag carries the reason the record was singled out — otherwise it is just an alert people mute within a week.
Twelve capabilities in the AI layer
Two groups. Evidence: the daily brief, the contextual panel, OCR with a posting proposal, proposals as drafts, an evaluation framework in CI, and a model gateway with an audit trail. Foundations: cash-flow and demand forecasts, anomaly detection, agents with autonomy levels, local models and RAG, an MCP server, and AI credits per tenant.
AI with evidence
Every answer carries the figure, the reasoning and the source document. Every suggestion is a draft.
6 capabilities
AI Daily Brief
On My Work, at sign-in: what changed, who is late, what is running out — with proposed actions. Every conclusion carries an explanation and drill-down to the data.
NG One AI — contextual panel
A right-hand panel on every screen, working on the current object: explain this posting, check the VAT treatment, find the related goods receipt, check for a duplicate, draft a message to the supplier.
OCR and posting proposal
PDF or scan → extraction → supplier recognised → duplicate detection → posting proposal → proposed link to the goods receipt; uncertain fields are flagged, not guessed.
Action suggestions as drafts
“Prepare reminders for customers more than 30 days late” produces drafts for review, not sent reminders. The draft enters the same workflow as a human decision.
AI evaluation framework in CI
Golden sets of queries and documents, regression tests, tenant isolation tests and prompt-injection tests; OCR is measured per field type and by the share of fields the system refuses to guess.
Model gateway and AI audit trail
Swappable provider, budget and quota per tenant, visible cost, personal-data redaction, and an append-only trail: input, model and prompt version, output, and who confirmed it.
Forecasting, agents and AI foundations
Autonomy is switched on per process and per limit — it never ships enabled.
6 capabilities
Forecasts: cash flow, demand, collection risk
Collection forecasting from each customer's payment history, demand forecasting for replenishment and MRP, and a collection-risk warning before a document is approved — not after it falls due.
Anomaly detection
Deviations in postings, duplicate and suspicious invoices, unusual margins and discounts — flagged as they occur, with the reason the record was singled out.
AI agents with autonomy levels
An AP agent, a sales-order agent and a collections agent operate in three modes: propose → act on confirmation → act autonomously within limits. The limits are the same approval limits a person has.
Local models and RAG
Hosted or local models for on-premise privacy; RAG over documents, master data and manuals, indexed per tenant under the same RLS isolation as the data itself.
MCP server
NG One resources and actions exposed as MCP tools: an external agent reads through permissions and data scope, and changes state only through state machines and workflow.
AI permissions, credits and consumption
The entitlement framework: AI credits, quotas and usage metering per tenant, with an entitlement check at the API layer alongside the permission check. AI cost is visible, not estimated at month end.
Automation has its own page: rules, approvals and the 'what did the ERP do on its own' measure live under Automation, because a workflow kernel carries them rather than a model — a rule is deterministic, and there is always an answer to which rule produced what.
See a decision with its evidence, not a slide about AI
We run the walkthrough on a demo tenant with twelve months of data: the panel answers with the source document, OCR flags the fields it is unsure of, and the proposal arrives as a draft that you confirm.