Agentic AI for SAP
Ask your live SAP.
Check every number.
Today a question becomes a ticket, a report, an extract, a spreadsheet, and by the time it arrives the business has moved on. AnrilX answers it from your live SAP, and prints the entity, field, filter and period under every figure. So the number can be checked rather than believed.
- Every figure carries its Source Line
- Reads live SAP, never a copy
- No write without a named approver
-
Connects to SAP
ECC and S/4HANA
-
Inside your landscape
Deployed on-premise
-
Runs on your model
Bedrock, Vertex, self-hosted
-
Your SAP authorisations
It widens nobody's access
-
Eight SAP modules
FI, CO, SD, MM and more
-
Every write approved
A named person releases it
How it is built
0model calls
to render a saved dashboard or page a report
A stored view holds a compiled query, so re-running it is database work plus an SAP read. Nothing is re-reasoned, so nothing drifts.
100%of writes
pause for a named human
The write path has no automatic branch. The exact payload and the exact SAP call are shown, and approval is the only thing that releases them.
Everyfigure
carries its SAP source line
Entity set, field, aggregation, filter and period, printed under the number, not buried in a log. Where the question left something open, the assumption it made is printed with it.
On-premiseonly
deployed inside your landscape
Your SAP data, your documents and your policy sit in your own environment. There is no shared multi-customer store.
What checking a number looks like
How much discount did we give last month, by condition type?
412.8m SAR in August 2026, across four condition types, 6.2% of gross order value. ZK02 customer rebate is the largest at 178.4m and is contractual. ZD01 is the one worth a look: it is the manual header discount, entered by hand, and it accounts for 94.1m across 1,208 orders.
SAR
Source: zsd_sales_order / A_SalesOrderItemPrcgElmnt · discount = sum of ConditionAmount · grouped by ConditionType · joined to header for the document date · August 2026
Assumed: discount means the condition types your pricing procedure configures as discounts: ZK02, ZD01, ZD02, ZK05. Freight, surcharge and tax conditions are excluded. Item-level rows carry no date of their own, so the period came from the header.
The problem
No standard report answers these
These are not lookups. Each one needs two windows and the difference between them, or a figure that is not a field, or the rows that are not there.
Sales
“Which products are declining fastest, quarter on quarter?”
Discount is not a field on the order
Procurement
“Which suppliers have quietly become more expensive this year?”
Spend is on the item, not the header
Finance
“Which cost centres are over budget, and what drove it?”
The number you need is a calculation, not a column
Production
“Which orders are at risk from a material that has not arrived?”
Risk is a question about two modules at once
Who builds it
SAP people, not just AI people
AnrilX comes out of a working SAP practice. The questions on this site were written by consultants who spent years answering them by hand, which is why the platform prints its workings rather than asking you to trust them.
SAP Silver
partner: KO Innovation
AnrilX is built by the SAP practice of its parent organisation, inside the SAP partner programme rather than alongside it.
50+
SAP consultants, certified
Covering each module the platform reads (SD, MM, FI/CO, PP, EWM, QM, PM and HCM), so the questions were written by people who have answered them by hand.
One
AI team behind them
The engineers building the platform sit with the consultants who know the data. That is why a figure carries its derivation: somebody in the room knew it would be asked for.
30 min
demo, run by a certified consultant
Not a sales call with a slide deck. An SAP-certified consultant takes your questions and shows the platform answering them, with the source line under every figure.
Book the 30-minute demo and an SAP-certified consultant runs it: bring your own questions.
How we deliver
It is implemented, not signed up for
Nobody ships you a generic assistant and leaves. An SAP team reads how your system was actually built (the blueprint, the configuration, the customisations somebody made years ago) and the platform is tuned to that before it answers a single question.
- 1
We look at your landscape first
Which modules are live, which releases you are on, and what has been built on top of them. The shape of the work is agreed before anything is gathered.
- 2
Your configuration is read, not guessed
Business Blueprint documents, the configuration itself, and the customisations and Z-objects your system actually carries.
- 3
The platform is tuned to that
Your catalogue, your fields, your naming, your conventions. It answers from how your SAP was built rather than from how SAP ships.
- 4
Deployed where you decide
Inside your landscape, with the deployment shape and the model host chosen with you rather than for you.
The same people who implement SAP for a living. Who we are
How it works
A model reasons. Code queries.
The one decision everything else rests on: the AI is never allowed to reach your SAP itself. It works out what you meant, and ordinary software goes and gets it.
-
STEP 01
You ask
A question in plain language. No table names, no transaction codes, nothing to learn first.
-
STEP 02
It finds where the answer lives
It knows the layout of your SAP, so it goes to the part that can actually answer, not the one whose name sounds closest.
-
STEP 03
The AI decides what you meant
This is the only place the AI is used, and all it does is choose from a shortlist. It cannot reach your data itself.
-
STEP 04
Your rules are applied
Every request is checked before it goes anywhere. Anything that would change SAP stops and waits for a named person to approve it.
-
STEP 05
You get the answer, with its workings
The figures come back with where each one came from printed underneath, so the number can be checked rather than believed.
Which products are declining fastest, quarter on quarter?
Source: zsd_sales_order / ZA_SalesOrder net revenue = sum of TotalNetAmount Q3 2026 vs Q2 2026
Illustrative figures · the source line is what the product prints
What you get
One workspace for the whole question
Ask it, watch it, report on it, forecast it, and act on it, without leaving the answer to go and find a tool.
-
Ask it the way you would ask a person
No query language, no entity names, no transaction codes. A question in plain English comes back as figures, with the SAP field and the period they were computed from printed underneath.
-
Describe the view. It builds itself.
Say what you want to watch and the dashboard assembles. Each panel stores the question rather than the answer, so it runs again against live SAP every time it is opened.
-
A report you can page, sort and keep
Tables that page, sort and filter inside SAP rather than over an extract. Save one, come back tomorrow, and it is reading today's data.
-
It reports on your basis, not its own
Give it the documents you have already written (reporting standards, close policy, pricing guidelines), and it works out which figures each one governs.
-
Trained on your history. Tested on what it never saw.
History is pulled from SAP, a model is trained on it, and it is backtested on whole periods held out of training. Then it tells you what it thinks happens next, and what is pushing it.
-
However the work arrives, it becomes SAP
An email, a WhatsApp message, a text file, or one sentence you type. AnrilX reads the work, builds the SAP documents it implies, and holds the whole batch for you to check before any of it is written.
-
Nothing reaches SAP unapproved
When something must change, the exact change and the exact SAP call are put in front of a person first. Approve it and it runs. Reject it and the reason becomes a constraint for the rest of the conversation.
Who asks
The same SAP, five different questions
What makes a question hard is rarely the data. It is that the answer is a calculation across grains, modules or windows that nobody has time to assemble by hand, and which calculation depends entirely on whose job it is.
Sales operations
The sales question your report does not have a column for
“Which products are declining fastest, quarter on quarter?”
What this role asksProcurement
Which suppliers quietly became more expensive
“Which suppliers have quietly become more expensive this year?”
What this role asksSupply chain & operations
The question that crosses four modules
“Which orders are at risk from a material that has not arrived?”
What this role asksFinance
A figure that is not a field
“Which cost centres are over budget, and what drove it?”
What this role asksIT & SAP teams
Every report request is a question nobody could ask
“Which entity sets does a given question actually read?”
What this role asksForecasting
Trained on your history, and tested against it
Demand, revenue, consumption and lead times, modelled on your own SAP history rather than an industry curve. Every forecast is backtested on months held out of training, and the error on those months is printed with the number.
- The accuracy measured on held-out months, shown with the forecast
- A warning naming the months it is extrapolating into, not a silent line
- No figure at all where the history is too short to test against
- The same source line underneath: you can see what it learned from
Forecast net sales for the next six months, and tell me where you are least confident.
1.24bn SAR for September, 7.46bn across the six months, 4.2% below the last actual month, backtested at 2.5% MAPE on three months held out of training, so that error is measured rather than asserted. Least confident: organisation 1400, where three of the six months fall below anything in its history. Those are flagged below rather than averaged into the total.
SAR
Source: zsd_sales_order / ZA_SalesOrder · net sales = sum of TotalNetAmount · backtested on 3 held-out months (2026-06…2026-08), 12 figures scored
Outside the observed range: 3 of 6 forecast months for organisation 1400 fall below anything in the training history. Those months rest on extrapolation, not measurement; compare them with history, never with the other forecast months.
Apps
Not every job is a question
Half the day is work with a shape: orders to raise, quotations to confirm, requisitions to place. AnrilX takes that work in whatever form it arrived and turns a batch of it into SAP documents in one pass, and still stops for a named person before anything is written.
Work arrives as
A customer enquiry, a purchase order in the body, a confirmation thread.
How a lot of field and distributor ordering actually happens.
Text and documents
A pasted note, a text file, a PDF order form, a scanned sheet.
Spreadsheets
The order list somebody already keeps, in the shape they keep it.
Or you just say it
No file and no forwarded thread. The instruction is the input.
Read into a batch
One pass, whatever it arrived as
Customer, material, quantity and date lifted from each item, with the ones it could not read flagged rather than guessed at.
“Create sales orders for every confirmed quotation this week”
Reads the quotations, builds an order per confirmation with its positions, and presents the whole batch for approval before anything reaches SAP.
SD“Turn these forty WhatsApp orders into sales orders”
Extracts customer, material, quantity and date from each message, flags the ones it could not read rather than guessing, and lines the rest up as a single approval.
SD · MM“Raise purchase requisitions for everything below reorder point at plant 1000”
Checks stock against reorder levels, groups by supplier, and drafts the requisitions with the figures that justified each one attached.
MM · EWM“Block delivery on every order for customers over their credit limit”
Crosses open orders with credit exposure, lists who and why, and holds the block for a person to release: one approval, every line attributed.
SD · FIIt stops for a named person
The whole batch at once, every line showing what it was drawn from. Nothing has reached SAP yet, and nothing does until somebody says so.
Written to SAP
Documents, in the modules they belong to
The payload approved is the payload sent, not a summary of it, and not re-generated afterwards. Approved, refused or failed, the attempt is recorded.
And the screens themselves
The ones that ship with it
Task screens for the work every SAP shop repeats: raising a quotation, tracking an order through delivery and billing, inspecting discounts by condition type, checking stock before a promise.
Order-to-cash tracker · Quotation capture · Discount inspector · Stock check
The ones you describe
Say what the screen should do and it exists. An app is configuration rather than a deployment, so there is no release to wait for and no ticket to raise, and the same approval rules apply to it as to everything else.
"Build me a screen for returns approval across plants" → a working app
What it runs on
Your SAP, and your model
It connects to what you actually run
Not only the newest thing. The estate most SAP customers are on today is supported the same as the estate SAP would prefer they were on.
- S/4HANA on-premise In your own data centre, on your own hardware.
- S/4HANA private cloud RISE private edition, inside your tenant.
- S/4HANA public cloud Public edition, over the published OData services.
- SAP ECC The estate most SAP customers are still running today.
It runs on the model you choose
Whoever hosts the model is a sub-processor, unless it runs somewhere you already control. So it is your decision, not ours, and the platform is not tied to one vendor.
- AWS Bedrock Inference inside your own AWS account and region.
- Google Vertex AI Inference inside your own GCP project.
- Azure OpenAI Inference inside your own Azure tenant.
- Provider APIs Anthropic, OpenAI and others, called directly.
- Self-hosted models Open-weight models on your own hardware. Nothing leaves your network.
- Any OpenAI-compatible API Point it at an endpoint you already run, and it uses that.
Governance & trust
Nothing reaches SAP unapproved
An AI that can change your ERP is only useful if it cannot change it quietly. Every governed call passes a fail-closed policy gate, and every write stops for a person, showing the exact change and the exact SAP call before anything is sent.
- The payload a human approves is the payload SAP receives Not a summary of it, and not re-generated afterwards.
- A rejection is a constraint, not a retry Say no and the identical call is refused for the rest of the run.
- Your SAP authorisations still apply A person sees exactly what their own SAP account permits.
- Every attempt is recorded Approved, refused or failed, with what was asked and what came back.
Block delivery on sales order 5000099: the customer is over their credit limit.
Validated against SAP and ready for your approval. Nothing has been sent yet, and nothing will be until you release it.
SAP write requires your approval
- operation
- update
- entity set
zsd_sales_order / A_SalesOrder_2- record
SalesOrder = 5000099
Fields that change
DeliveryBlockReason - 01 · Credit limit
Source: zsd_sales_order / A_SalesOrder_2 · PATCH on key SalesOrder 5000099 · validated against the service contract before display
Approve or reject only. The payload cannot be edited here, because approving one object and sending another is not approval.
The difference
What changes on Monday
Today
With AnrilX
Raise a ticket
Ask the question
Wait for the report
Read the answer
Extract to a spreadsheet
Work from live SAP
A number with no context
A number with its source
Change SAP by hand
Approve a validated change
Insights
Measured, not assumed
Why an SAP AI has to be implemented
A generic assistant pointed at SAP knows how SAP ships. It does not know what your pricing procedure was extended to do in 2014, and no reasoning recovers that.
What an AI should say when it fails
Most of the risk in an AI system is not what it does when things work. It is what it reports when a read came back short or the system was unreachable.
The dashboard that froze, and nobody noticed
Eight saved dashboards carried hardcoded date ranges. Nothing errored, no alert fired, and the titles still said last 30 days. Here is why that happens.
Questions people ask first
What is AnrilX?
AnrilX is an agentic AI platform for SAP. You ask your live SAP system a business question in plain language and it returns the figures, with the SAP entity, field, filter and period they were computed from printed underneath. It also builds dashboards and reports, forecasts from your own history, and executes governed changes back into SAP after a person approves them.
Does AnrilX copy our SAP data out?
No. AnrilX reads SAP over OData at question time and does not maintain a copy of your transactional data. Reads are row-capped by design, and the platform is deployed on-premise: your data, your documents and your policies stay in your own environment.
Can it change data in SAP?
Only with explicit human approval. A change is compiled, validated against SAP's own contract, and then shown to a person as the exact payload and the exact call. Nothing is sent until that person approves it. A rejection becomes a hard constraint for the rest of the run rather than something the system quietly retries.
Which SAP systems does it work with?
AnrilX connects to SAP ECC and to S/4HANA (on-premise, private cloud or public cloud) over OData. It is deployed into your landscape and uses your existing SAP authorisations, so a person sees exactly what their SAP account already permits.
How is this different from asking ChatGPT about our SAP data?
A general chatbot has no connection to your system and no way to check what it says. AnrilX queries your live SAP directly, builds every query in code rather than letting a language model write one, and prints the provenance of every figure. A model in AnrilX never writes an SAP filter, never writes a date, and never picks an entity outside a list the platform built for it.
Why not just use SAP Joule?
If you are already fully on RISE or GROW with largely standard processes, that is often the right answer and we will say so. Joule is included, native, and spans SAP's whole product estate. Where it does not reach you is the interesting part: since May 2026 SAP extends it to ECC and on-premise S/4HANA, but only for customers who have already started on RISE, shifted at least half their maintenance spend to the cloud, and signed the Max Success Plan, and SAP describes that as an interim solution. AnrilX has none of those conditions. The durable differences are architectural rather than commercial: where the deployment sits, whether a figure carries its derivation, and whether a write can happen without a named person.
Doesn't SAP's API policy rule this out?
It restricts a specific shape rather than the category: an AI that plans, selects and executes sequences of API calls on its own. SAP's own CTO said at Sapphire that they are not keeping agents out, they are letting them in where the interface is defined. The controls SAP's ecosystem states as the requirement (named-user identity, least privilege, human approval before business-critical writes, allowlisted published APIs, and an audit log correlating prompt to tool call) are the design here rather than a configuration of it. What we will not do is tell you that makes you compliant; only SAP, your contract or SAP support can answer that for your landscape.
How long does it take to get running?
The platform ingests your SAP service catalogue on first run and is answering questions against it the same day. Loading your own reporting standards and business rules is a drag-and-drop step, and is what turns correct answers into answers on your basis.
Bring the question your reports cannot answer
Thirty minutes against a live SAP system we provide: no access to yours, nothing to set up. If it cannot answer, you find that out in half an hour rather than three months into a pilot.