Cobalt ERP Intelligence Core

ERP that already did the thinking.

Eleven engines, and not one of them allowed to overstate what it found.

Eleven reasoning engines and eleven platform services, native to every ERP object, permission and workflow you already run. Not one model asked to do everything — eleven, each with one job and one boundary, composed by a runtime that will not let any of them overstate what it found.

Eleven engines · Eleven platform services · Evidence on every claim
The engines

Eleven engines, because the guarantees are different.

One general-purpose model asked to do all of this would give you the same shrug for an arithmetic question and a causal one. Cobalt ERP separates them, because what you are entitled to trust differs: arithmetic is reproducible, a mined pattern is an association, a forecast carries an interval, and an optimiser has to tell you whether a plan is merely feasible or provably best.

Eleven engines, each with one job

Not one model asked to do everything. Each engine has a single responsibility and a single kind of output, and the runtime composes them — they never call each other.

Reasoning engine

Engine 01

Deterministic and policy intelligence

Arithmetic, tax rules, eligibility, hard constraints

Finding

Engine 02

Analytical and data-mining intelligence

Cohorts, associations, outliers, change points

Mining insight

Engine 03

Predictive machine learning

Forecasts, probabilities, calibrated intervals

Prediction

Engine 04

Process and task intelligence

Discovery, conformance, bottlenecks, rework

Process insight

Engine 05

Computational intelligence and optimisation

Constraint solving, scheduling, multi-objective choice

Optimisation proposal

Engine 06

Simulation and scenario intelligence

Monte Carlo, discrete-event, sensitivity, stress

Simulation result

Engine 07

Knowledge graph and semantic intelligence

Typed relationships, lineage, dependency traversal

Graph traversal

Engine 08

Retrieval and enterprise memory

Permission-aware lexical, vector and prior-case search

Retrieval evidence

Engine 09

Document and multimodal intelligence

Text recognition, layout, tables, extraction, comparison

Extraction candidate

Engine 10

Language and conversational intelligence

Intent, clarification, grounded explanation

Explanation

Engine 11

Decision, recommendation and causal intelligence

Safe-candidate ranking, arbitration, uplift, causal analysis

Recommendation

Boundaries

What each one is not allowed to claim.

Every engine ships with a boundary as well as a job. The boundaries are the reason the output is usable on a ledger.

  • Deterministic and policy is the only engine that may block or permit a transition, and only through rules that already exist in the domain.
  • Data mining reports association. It is never rendered as causation — and no language model is permitted to reword it into one.
  • Predictive machine learning is advisory derived data. It never carries permission, ledger, tax or approval authority.
  • Process mining observes what happened. Observed practice does not redefine the permitted process.
  • Optimisation must distinguish feasible, best-known and proven optimal, and may never quietly relax a hard constraint into a soft penalty.
  • Simulation produces a scenario. A scenario is not a forecast, a commitment or an authorisation.
  • Knowledge graph is derived and rebuildable — never the authority for a transaction. Inferred links stay separate from asserted ones.
  • Retrieval returns evidence, and permissions are re-checked after reranking, not before.
  • Document intelligence produces candidates. They stay candidates until validated or accepted by a person.
  • Language may not write database queries of its own, write to the database, invent a calculation, or state a claim it cannot support.
  • Decision and causal ranks only what deterministic filtering has already declared safe. A score is not permission, and a causal claim needs a defensible identification strategy.
Composition

Engines never call each other. The runtime composes them.

A compiled recipe decides which engines run, in what order, and what happens when one abstains. That indirection is what keeps each engine's guarantee intact all the way to the screen.

How Cobalt Intelligence thinks

Every question is answered by the combination of engines that can actually answer it. This is which ones each of eight real questions calls, which one leads, and what it hands to next.

Primary engine Supporting engine Not used
Eight questions asked on a live record, and which of the eleven reasoning engines each one calls. Every cell says whether that engine is primary, supporting or not used. The last column names the engine that leads and what it hands to.
Question asked on a live record Intelligence engines used Lead engine & chain
Question DeterministicData MiningMachine LearningProcess MiningOptimisationSimulationKnowledge GraphRetrievalDocumentLanguageCausal Inference Lead engine
Can we promise this order by Friday?Quote to cash · Sales order Deterministic: primary engineData Mining: not usedMachine Learning: primary engineProcess Mining: not usedOptimisation: primary engineSimulation: supporting engineKnowledge Graph: supporting engineRetrieval: primary engineDocument: not usedLanguage: not usedCausal Inference: supporting engine DeterministicDeterministic → Causal7 / 11
Why did this quote lose margin?Quote to cash · Pricing Deterministic: primary engineData Mining: primary engineMachine Learning: supporting engineProcess Mining: not usedOptimisation: not usedSimulation: not usedKnowledge Graph: primary engineRetrieval: not usedDocument: not usedLanguage: supporting engineCausal Inference: supporting engine Data MiningData Mining → Deterministic6 / 11
Which supplier should we buy this from?Procure to pay · Sourcing Deterministic: primary engineData Mining: supporting engineMachine Learning: primary engineProcess Mining: not usedOptimisation: primary engineSimulation: supporting engineKnowledge Graph: primary engineRetrieval: not usedDocument: not usedLanguage: supporting engineCausal Inference: supporting engine OptimisationOptimisation → Deterministic8 / 11
Where does our order process actually stall?Procure to pay · Process Deterministic: supporting engineData Mining: primary engineMachine Learning: not usedProcess Mining: primary engineOptimisation: not usedSimulation: not usedKnowledge Graph: not usedRetrieval: supporting engineDocument: not usedLanguage: supporting engineCausal Inference: not used Process MiningProcess Mining → Language5 / 11
What happens to cash if Redlands slips two weeks?Finance · Scenario Deterministic: primary engineData Mining: not usedMachine Learning: supporting engineProcess Mining: supporting engineOptimisation: not usedSimulation: primary engineKnowledge Graph: primary engineRetrieval: not usedDocument: not usedLanguage: supporting engineCausal Inference: supporting engine SimulationSimulation → Causal7 / 11
Pull the totals off this supplier PDFProcure to pay · Bill inbox Deterministic: primary engineData Mining: not usedMachine Learning: not usedProcess Mining: not usedOptimisation: not usedSimulation: not usedKnowledge Graph: supporting engineRetrieval: supporting engineDocument: primary engineLanguage: supporting engineCausal Inference: not used DocumentDocument → Deterministic5 / 11
What does our returns policy say about this?Service desk · Policy Deterministic: supporting engineData Mining: not usedMachine Learning: not usedProcess Mining: not usedOptimisation: not usedSimulation: not usedKnowledge Graph: supporting engineRetrieval: primary engineDocument: not usedLanguage: primary engineCausal Inference: not used RetrievalRetrieval → Language4 / 11
Which customers are about to churn?Subscriptions · Revenue Deterministic: not usedData Mining: primary engineMachine Learning: primary engineProcess Mining: supporting engineOptimisation: not usedSimulation: supporting engineKnowledge Graph: supporting engineRetrieval: supporting engineDocument: not usedLanguage: supporting engineCausal Inference: primary engine Machine LearningMachine Learning → Causal8 / 11
The right engines. Every time.

Cobalt Intelligence selects and orchestrates the combination of engines that can answer the question in front of it — a compiled recipe, not a routing guess, and never a ranking before the deterministic safety filter has run.

11Intelligence engines
DynamicEngine orchestration
100%Explainable outcomes
Working together

What each engine gives the others.

The value is in the composition, not the list. A forecast without graph context is guessing about the wrong entity; an optimiser without the forecast's uncertainty plans for a number that will not hold; a recommendation without deterministic filtering is a liability.

  • Deterministic supplies the hard constraints, validators and safe candidates every other engine works inside — and mined exceptions can propose new rules, but only for a person to ratify.
  • Knowledge graph supplies relationship context and prunes candidates, so prediction and optimisation start from the right entities.
  • Process mining supplies actual flow state, variants and where the delay really is — which is what optimisation needs for capacity.
  • Predictive machine learning supplies probabilities and intervals; simulation takes those distributions and stress-tests the plan built on them.
  • Retrieval supplies the cited policy and the prior case; document intelligence supplies candidate facts that deterministic validation confirms.
  • Language is last, not first — it presents what the others established, using typed tools and their evidence.
  • Decision and causal arbitrates between objectives, and records exposure so the effect of its own recommendations can be measured later.
Per module

One core, every module, the same eleven engines.

A module does not fork the intelligence core — it plugs into it, contributing what a subject is and what counts as an observed outcome. Everything else, from drift monitoring to the accuracy report, derives automatically from one registration.

Every module runs on the same eleven engines

A module does not fork the core or bring its own model. It plugs into the same eleven engines every other module uses, and the recipe for its questions decides which of them run.

Every business area in Cobalt ERP and which of the eleven reasoning engines it composes. The last column gives the count.
Business area Engines it runs on Engines
Business area DeterministicData miningMachine learningProcess miningOptimisationSimulationKnowledge graphRetrievalDocumentLanguageCausal Count
Finance & accountingClose, controls, reporting Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
Customer relationshipsOpportunities, accounts, activity Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
Quote to cashPricing, promise, fulfilment Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
Subscriptions & revenueRenewal, churn, recovery Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
Procure to paySourcing, matching, contracts Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
Inventory & planningDemand, safety stock, transfers Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: not used 10 of 11engines composed
WarehouseWaves, slotting, workload Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: not used 10 of 11engines composed
Distribution & fleetRoutes, arrival time, cost to serve Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: not used 10 of 11engines composed
ManufacturingShortage, sequence, yield Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
Field serviceScheduling, first-time fix Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
Service deskClassification, service level, resolution Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
People and payrollPolicy help, intake, checks Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
Master data managementDuplicates, resolution, gaps Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 10 of 11engines composed
Fixed assetsLifecycle, revaluation Deterministic: usedData mining: usedMachine learning: usedProcess mining: usedOptimisation: usedSimulation: usedKnowledge graph: usedRetrieval: usedDocument: usedLanguage: usedCausal: used 11 of 11engines composed
What that produces

The chain behind each module's headline capability.

  • Quote to cash — conversion, margin leakage, promise confidence and fulfilment alternatives, from mining and prediction through graph, then simulation and optimisation, filtered deterministically and ranked.
  • Procure to pay — supplier lead time, sourcing comparison, contract leakage, order consolidation and the expedite trade-off.
  • Inventory — demand class, safety stock, obsolescence, transfers and a service-level what-if, ending in a draft transfer or requisition.
  • Finance — close readiness, control breaks, anomaly evidence and run-rate scenarios, with the ledger untouched by any of it.
  • Manufacturing — material shortage, sequence, yield and work-in-progress variance, as a planner proposal validated against hard constraints.
  • Distribution — arrival-time risk, route and load plans, failed-delivery patterns and cost to serve, as a feasible dispatch proposal.
  • Service desk — classification, service-level risk, repeat clusters and a cited resolution drawn from prior cases.
  • Master data — duplicates, entity resolution, configuration gaps and change-impact analysis, into a maker-checker merge proposal.
The agentic layer

A bounded run, and a gate no step crosses on its own.

An agent in Cobalt ERP is not a model with tools and a long leash. It is a task with a budget, a plan compiled from registered skills, and a verifier that is not allowed to be the thing that did the work.

A bounded run, and the gate in the middle of it

Nine steps, and nothing crosses the middle of them on its own. The planner cannot reach a tool, the verifier is not allowed to be the thing that did the work, and between staging and applying everything is resolved again from scratch.

Bounded agent run Nothing crosses this gate on its own
  1. 1

    Bounded task

    one goal, one budget

  2. 2

    Live context

    actor, permission, entity resolved

  3. 3

    Plan compiled

    from registered skills only

  4. 4

    Reads execute

    typed results checkpointed

  5. 5

    Proposal staged

    preview and fingerprint

  6. 6

    Human approves

    or a named automation policy

  7. 7

    Re-resolved

    permission, duties, fingerprint again

  8. 8

    Domain service

    the same one a person would use

  9. 9

    Verified

    independently, against the success question

PLANNERcannot call tools
VERIFIERindependent and deterministic
WATCHDOGstep, time, cost, fan-out
ON ANY CHANGErefuse the whole request
Autonomy

Granted per skill, per tenant, one level at a time.

Eight levels, granted per skill and per tenant

Autonomy is not a switch. It is eight rungs, admitted one skill at a time for one workspace at a time, and the top three need a named policy or a person before anything is applied.

Autonomy level Granted per skill, per tenant
0Explainproduct and process concepts, no tenant data touchedNo tenant data
1Diagnoseinspects permitted live data and produces findingsReads live data
2Navigatedeep-links to the exact permitted resolution surfaceReads and deep-links
3Prepareproduces an editable draft or proposalWrites a draft
4Coordinateroutes questions and normal domain approvalsRoutes for approval

From here a person or a named policy is required

5Attended applyyou confirm; the domain service applies after revalidationApplies on your confirmation
6Event proposala background trigger prepares and assigns a proposalPrepares unprompted
7Bounded automationone admitted reversible action executes within policyApplies within policy
Platform services

Eleven services that make the engines governable.

An engine nobody measures is not an asset. These carry the contracts, the context, the evidence, the evaluation and the kill switches — and every capability registers once, so nothing can be scored while quietly escaping drift monitoring.

Eleven platform services, which is what makes them governable

An engine on its own is a library. These are the services that give it a contract, a scope, an audit trail and somewhere to be switched off.

Platform service

Service 01

Contracts and schema registry

Versioned artifact schemas and adapters

Versioned schemas

Service 02

Capability catalogue and domain kit

Manifests, registration, reachability checks

Capability manifests

Service 03

Context, semantics and scope

Tenant, actor, entity, book, period, permission

Permission scope

Service 04

Fact, event and dataset plane

Point-in-time facts, events, frozen datasets

Frozen datasets

Service 05

Composition and decision runtime

Recipe compilation, arbitration, abstention

Compiled recipes

Service 06

Job, worker and serving runtime

Ledger, scheduled work, batch scoring

Job ledger

Service 07

Artifact, model and knowledge registries

Models, datasets, ontologies, playbooks, indexes

Versioned artifacts

Service 08

Outcomes, evaluation and lifecycle

Drift, promotion, demotion, retirement

Drift and promotion

Service 09

Evidence, provenance and explainability

Sources, lineage, citations, receipts

Lineage and citations

Service 10

Security, privacy, egress and actions

Isolation, egress, duties, consent, kill switches

Kill switches

Service 11

Native experience, feedback and operations

Panels, queues, alerts, cost, operations

Panels and queues

Cost

The cheapest layer that meets the bar.

The conversational surface has its own discipline on top of all this: a request walks a cascade and stops at the first layer that resolves it. Deterministic routing first, an on-device classifier next, a replay of something this tenant has asked before, and only then a paid model.

A request stops at the first layer that can answer it

Most questions never reach a model. They resolve on a deterministic router or an on-device classifier running on your own server, at no credit cost, and a model is asked only when the layers below it cannot answer.

Layer 1Deterministic router

Regex, slot-fill and conversation focus. No model involved.

62% free
Layer 2On-device classifier

A bag-of-words language model, trained at boot, running on your own server.

20% free
Layer 2bLearned replay

A phrase this tenant has asked before, replayed and re-validated live.

9% free
Layer 3Model router

Claude or OpenAI, answering into a strict JSON contract.

7%
Layer 4Extraction and vision

On-device text recognition first; a model only on documents that need it.

2%
LAYERS BEFORE A MODEL3
RUNS ON YOUR SERVERLayers 1 and 2
WORKS AT ZERO CREDITSEverything but 3 and 4
MODEL OUTPUTUntrusted until validated
Evidence

Every figure says what kind of thing it is.

The one thing that makes intelligence usable on a ledger is knowing which claims are arithmetic and which are inference.

Calculated by Cobalt ERP

Arithmetic over your own records. Reproducible, and the same every time you ask.

Rule-based

A deterministic rule fired. The explainer names which one and what it read.

Extracted from a document

Read off a PDF or an image, with a confidence score and the source page attached.

AI-generated

A model wrote it. Marked as such, every time, with no exception for convenience.

FAQ

Frequently asked questions.

Why eleven engines rather than one model?

Because what you are entitled to trust differs by method. Arithmetic is reproducible; a mined pattern is an association and never a cause; a forecast carries a calibrated interval; an optimiser must tell you whether a plan is feasible, best-known or provably optimal. One model asked to do all of it gives you the same confident tone for every one of them.

Do the engines call each other?

No. A compiled recipe composes them, and the compiler rejects anything unsafe — cycles, a missing input, ranking placed before the deterministic safety filter, more than one side effect, or a user-visible output with no evidence and no expiry.

Which engines does my module use?

All fourteen modules run on the same eleven; what differs is which do the primary work. Manufacturing, procure-to-pay and field service use all eleven. Master data barely needs simulation. HR is deliberately restricted.

What is restricted for HR?

Prediction, optimisation, simulation and causal methods are limited to privacy-thresholded aggregate planning and administrative quality. They must not rank individuals or drive an employment decision, and task mining uses application events only — no screen recording, no keystroke capture, no productivity ranking.

What can an agent actually do on its own?

Whatever level you have admitted for that one skill on your tenant, and nothing beyond it. There is no blanket autonomy switch. Above attended apply, an action needs either a person or a named automation policy with its own caps.

What can an agent never do?

Post or reverse a journal, approve a controlled transaction, release a payment, change bank details, determine or submit tax, close or reopen a period, override a credit hold, dispatch regulated goods, release a quality hold, raise its own permissions, delete evidence, or make any hiring, firing, pay or disciplinary decision. That boundary is permanent.

How do I know an answer is right?

Every output carries its provenance — calculated, rule-based, extracted or generated — along with the records it read, the version of the rule or model that produced it, and an expiry. The evidence and decision graph keeps the lineage from source fact through to your decision.

Is a model allowed to instruct the system?

No. Model output is untrusted data with the same standing as a message from a stranger. It enters only by validating against a schema that cannot express an action, entity identity is re-resolved against your database, and permissions are checked live rather than taken from anything the model supplied.

See the whole chain on your own records.

Join the waitlist and put a real question to it. Every answer names the engines it used, the records they read and the rule or model version behind each figure.

Join waitlist
No lock-in — export your data anytime.