Cobalt ERPAI that understands the work.
Built into the ERP your team uses every day.
Ask about an order, investigate a variance or hand an agent a task. Cobalt uses the relevant records and methods to prepare a useful answer, with the next decision kept clear.
Start with the question your team needs to answer.
Can we deliver on Friday?
Check stock, existing commitments and incoming supply. See what can be promised now and what depends on a future receipt.
Why is this invoice on hold?
Compare the supplier invoice with its purchase order and receipts. Show the quantity or price difference and who needs to review it.
What changed in our margin?
Compare revenue and cost with the previous period or budget. Follow the difference into the transactions behind the report.
One Intelligence Core across the business.
Sales, finance and operations share the same records. The Core selects the methods a question needs, checks the result and connects it to the next step in the workflow.
Find the relevant records
The question is scoped to the company, period and records your role can access. Retrieval finds the evidence; a knowledge graph helps when relationships between records matter.
Calculate and explain
Cobalt applies the relevant calculations and business rules. Forecasts and scenarios keep their assumptions visible. A connected language model can help interpret the question or draft an explanation; it does not decide the accounting figures.
Prepare the next step
The result can become a report, a proposed order change or a draft for approval. Configured permissions and approval rules determine what can happen next, with an activity record of the action.
It works on your real orders, invoices and stock, follows the same permissions and approvals as your team, appears in the record you already have open, and checks what actually happened after you act.
What AI-native means at Cobalt ERPEleven specialist methods.
Each method answers a different kind of question and has a firm limit it will not cross. A task only uses the methods that help it, so a screen that needs a total gets one calculated exactly by Cobalt ERP.
Accounting sums, tax rules, checks, eligibility and hard limits, each with a stated reason, plus exact what-if sums.
Your books' own maths. Never an AI's opinion.
Customer groups, items bought together, repeat sequences, outliers, sudden changes and data-quality problems.
Shows what happens together. Never claims one caused the other.
Predicts outcomes, how long things will take and demand, with honest probabilities and ranges.
Advice only. It cannot approve, post or grant access.
Shows how orders and tasks really flow: the variations, bottlenecks, waiting time, rework and skipped controls.
Shows the process as you really run it.
Plans routes, schedules, pick runs and loads within your real limits, balancing more than one goal.
Says whether a plan works, is the best found or is proven best. Never quietly bends a limit.
Tests what-if questions, plays out many possible outcomes and shows how sensitive a plan is to change.
A scenario is not a forecast, a promise or an approval.
Knows how your customers, orders, products and suppliers link together, and where each figure came from.
Adds context to your books. Never a second copy of them.
Finds the right records, policies, procedures and past cases, cites them, and only searches what you may see.
Permissions are checked before it searches.
Reads text, layouts and tables from documents, sorts them, and picks up barcodes, images and signatures.
Values it reads stay suggestions until checked or accepted.
Understands what you are asking, checks when something is unclear, summarises, and explains answers from your data.
Explains the answer. Never decides it.
Ranks safe options, balances competing goals, and tests whether a change really made a difference.
A high score is not approval. Cause and effect needs solid evidence.
Where every method hands its work: prepared, approved, carried out and recorded.
The only way anything gets changed.
Eleven services that keep it safe.
Getting an answer is only part of the job. You also need to know where it came from, what it cost and how to stop it. Every method above relies on all eleven of these.
Every result follows an agreed, checked format, so nothing produces output nobody expected.
A checked list of what it can do. If a feature is listed, it works.
Keeps every answer within your company, user, legal entity, books, country, currency, site, warehouse, period and permissions.
Keeps track of what was true when and what happened, and checks data quality before it is used.
Plans the steps for each question, runs some side by side, has a fallback, and can decline to answer when unsure.
Runs scheduled and triggered jobs, large batches and quick on-screen requests, and logs each one.
Keeps every model, dataset, process map, playbook and prompt under version control.
Compares results with what actually happened, spots when accuracy slips, and promotes or retires models.
Records sources, citations, how each figure was worked out and its known limits, so any result can be replayed.
Keeps workspaces apart, controls what leaves Cobalt ERP, enforces separation of duties and consent, and has an off switch.
The panels, work queues, assistant, alerts, feedback and cost controls you use day to day.
It tries the simple way first.
A typed question, an uploaded invoice or an agent's next step all start with the simplest option and stop as soon as there is a confident answer. That keeps Cobalt ERP Intelligence quick and cheap, and it keeps working if an outside AI provider goes down.
| Route | Capability | How it works |
|---|---|---|
| Layer 1 | Built-in rules | Recognises common questions in your own words, fills in details from the record you have open and runs the search. |
| Layer 2 | Small built-in model | Handles supported questions inside Cobalt without sending them to an outside model. Confidence checks determine when another method or clarification is needed. |
| L2b | Remembered questions | Remembers how it handled a type of question before. It reruns the search each time, so the figures are always current. |
| Layer 3 | Outside AI model | Used when a task needs an external model and the configured policy allows it. Only the permitted context needed for that request is sent; relevant record fields may be included. |
| Layer 4 | Reading hard documents | For documents the simpler readers could not handle cleanly, such as complex layouts, tables, handwriting and poor scans. |
Use the simplest option that meets the required standard for quality, safety and speed.
The rule Cobalt ERP Intelligence followsAgents that see a task through to the end.
An agent can work on a task over several days. It plans from an approved list of skills, does the work through Cobalt ERP's normal processes, and then checks separately that the result really happened. You decide how much it may do alone, one action and one workspace at a time.
Chooses from an approved list of skills and builds a plan with limits. It cannot use any tool directly.
Carries out each step under your current rules, checks permissions again at every step, and records the result.
Checks separately whether the task really succeeded, using exact checks where it can. It is never the part that did the work.
Enforces limits on steps, time, cost and scope, and stops the task if it is cancelled, switched off or goes out of bounds.
Agents work towards a goal using approved skills. Cobalt ERP Intelligence gathers the evidence, your normal business rules make any change, and people keep the final say on anything that matters.
How agents work in Cobalt ERPThe same steps, every time
Confirm who is asking and what they may do before anything runs.
Only the ones this workspace and this person may use.
Check every step before any of them runs.
Read the records, save progress and ask when unsure.
Anything that matters becomes a proposal to approve.
A person approves, or an automation rule you set up applies.
Permissions and the record are checked again, then Cobalt ERP makes the change.
The result is checked separately and a record is kept.
Eight levels of independence
You grant each level one action and one workspace at a time. Most work happens in the middle levels. The top level is a deliberate choice for each action, with limits and an off switch.
| Level | Capability | How it works |
|---|---|---|
| Level 0 | Explain | Answers questions about the product and your processes. Changes nothing. |
| Level 1 | Diagnose | Looks at live data you are allowed to see and reports what it found, with evidence. |
| Level 2 | Navigate | Takes you straight to the screen where you can sort the problem out. |
| Level 3 | Prepare | Prepares a complete draft you can edit and review. |
| Level 4 | Coordinate | Sends questions and normal approvals to the right people and waits for them. |
| Level 5 | Apply when you confirm | You confirm, then Cobalt ERP checks it again and applies it under your login. |
| Level 6 | Prepare in advance | When something happens in the background, it prepares the work and assigns it before anyone has noticed the problem. |
| Level 7 | Limited automation · opt-in | One action you have chosen, which can be undone, runs on its own within set rules, limits and a live off switch. |
It understands seventeen areas of your business.
Each module adds its own knowledge of what a warehouse, a supplier or an accounting period is. Every area works the same way, so adding a new one does not mean rebuilding anything.
WORKWork queue
master data managementMaster data
customer relationshipsCustomer relationships
quote to cashQuote to cash
SUBSubscriptions
procure to payProcure to pay
INVInventory
warehouse managementWarehouse
DISTDistribution
MFGManufacturing
FSVCField service
SDService desk
HRPeople
FINFinance
FAFixed assets
PLATPlatform
PROJProjects
What that looks like in your day
Quote to cash
Fewer surprises when you sell
- How likely each quote is to win
- Where discounts and pricing eat into margin
- How sure you can be of a delivery date, with other options
- Common reasons behind returns
- What is holding an order up, found before dispatch
Finance
Close sooner and explain every figure
- Supplier invoices read, coded and matched to order and delivery
- Duplicate invoices and risky payments flagged
- Bank reconciliation and matching customer payments
- Month-end checks that run all month and remember last month
- Expected cash in, with a likely figure and a cautious one
Inventory & warehouse
Stock levels that keep up with demand
- Demand forecasts that suit how each item sells
- Safety stock, reorder points and order quantities that adjust
- Slow-moving, dead and ageing stock spotted
- Pick runs and stock counts put in priority order
- Short picks and stock-outs flagged before the shift starts
Procure to pay
Buy with the facts in front of you
- Supplier lead times learned from your history, typical and slow
- Supplier quotes compared, and spend outside contract terms
- Purchase orders that could be combined
- Whether paying to expedite beats running out
- Alerts when a supplier's bank details change
Distribution
Every van, every drop
- Vehicle and load plans that respect real limits
- Late-arrival risk and patterns in failed deliveries
- Cost to serve each customer and each route
- Carrier choice and the risk of missing the dispatch cut-off
Onboarding & support
Help getting started and staying running
- A guided interview about your business, then a proposed setup
- Help moving your data, with a test run and a way back
- Support that knows exactly how your system is set up
- A clear reason why something is blocked
Every figure shows where it came from.
Five labels, used everywhere in the product. A figure calculated from your books is never shown as if AI wrote it, and AI-written text is never shown as a calculation.
Behind each one is a record of the figures that went in, the rule or model version used, the evidence it relied on and any limits it noted. Predictions are saved before the result is known and later compared with what actually happened, so you can see how accurate they really are.
What does AI-native mean here?
The AI works on your real orders, invoices and stock, follows the same permissions and approvals as everyone else, and appears in the record you already have open. Every module shares the same Cobalt ERP Intelligence, so there is no separate AI tool for each area. It shows its evidence and how sure it is, and it checks what actually happened after you act.
What is an AI agent?
An agent takes a goal, plans the steps from an approved list of skills, gathers the facts, prepares the work and sees it through to a checked result. A task can run for days or weeks, wait for a person and pick up where it left off. It does not make up business rules, look through your data freely, or take on authority nobody gave it.
Can it act on its own?
Only as far as you allow, for each action and each workspace. Most teams let it prepare the work and then confirm it themselves. Limited automation is a deliberate opt-in for single actions that can be undone, with value limits, a cool-down period, a live off switch and an automatic step back if quality drops. There is no general switch that hands over control.
Which AI model does it use?
Cobalt ERP Intelligence can use Anthropic's Claude or OpenAI's GPT models through one controlled connection, and which one is used is a setting. Built-in tools can answer supported requests without calling an outside model. Requests that need an external model follow the configured policy and available allowance.
What happens to my data?
Built-in analysis and requests to an outside AI service are separate routes. External requests use a controlled connection with permissions and cost limits. Only the permitted context needed for the request is sent; relevant record fields may be included. Your records are read to answer your questions and are not used to train AI for anyone else. Review the configured provider and data-handling terms before enabling an external service.
Which plan includes Cobalt ERP Intelligence?
Growth and above. Growth includes 100 AI credits a month, Scale 500 and Premier 2,000. Built-in rules and tools do not use external-model credits. They remain available when that allowance is exhausted, subject to the feature and usage limits in your plan; paid-model requests need available allowance.