Agentic AI Chapter 6 of 9

Six stages take you from scattered data to confirmed actions.

Chapter 5 followed one agent from a Friday sentence to confirmed cancellations on the Monday.

REGISTERMAPASKTEACHACTLEARNPeopleLima JunoYour systemsPick 23tablesA. Novak · Mon09:10Register inplacenames · keys ·sampleRows stayputSAP · Oracle · 3moreConfirm ·adjust · add338 · 33 · 41 ·TuePropose 371of 412ontology version1EKPO ·PO_LINES_ALLone object: POlineAsk in plainwordsL. Tan · Wed09:12Answer in4.2 squery on showRead inplace3 sources joinedAdd threetermsNCNR · LT fence· aliasIndexretrainedvocabulary v2 ·6 minDraft 29 ·route 9rule version 3 ·A2Approve eachdraftD. Okafor · Mon09:12ERP confirmswrite-backacknowledgedRecord everydecision203 of 212 aswrittenReview thelevelJ. Ridley · weeksixOutcome readbackday 14choosereadreviewone meaningaskqueryteachdraftwriteevidenceoutcomeshowcase estate · week one, 14 to 18 September 2026, and the Monday after · purple: the PO cancellation agent’s path
People, Juno and your systems across the six stages. The purple arrows trace the PO cancellation agent’s own path, from the tables it reads to the record it leaves. Showcase estate.
6
stages from registered tables to a confirmed write-back
product
23
tables from five systems, registered on the Monday and read where they live
showcase estate
338
of 371 proposed mappings confirmed in one action on the Tuesday
showcase estate
4.2 s
from L. Tan’s question to an answer read from three systems
showcase estate
So far
  1. The PO agentYou have seen one agent configured, run and released, reading three systems as one.
  2. Why nowYou have seen that spread data and different meanings are two of the four reasons earlier AI stalled.
The six stages

The machinery behind the agent you watched

The purple arrows in the picture above are this path; the rest of the chapter follows it stage by stage.

01
Register
SAP S/4HANA, Oracle Fusion and the planning suite: 23 tables, read where they live. Data registration
02
Map
EKPO in SAP and PO_LINES_ALL in Oracle map to one ontology object, the purchase order line: two systems read as one. Automatic mapping
03
Ask
On the Wednesday L. Tan asks which open lines are still inside their cancellation window; the answer, 31 lines worth $348,000, comes from the same object in 4.2 seconds. ChatLima
04
Teach
On the Thursday NCNR, the lead time fence and the strategic supplier alias join the vocabulary; the Friday policy uses them. Training
05
Act
On the Friday the policy becomes a five-row planning table, version 3 at A2; on the Monday the rules engine drafts 29 cancellations for D. Okafor’s queue. Rules engine
06
Learn
Every approve, edit and reject is recorded with the rule version and the facts; after six weeks, 203 of 212 approved as written. Learning loop
Stage one · Register

The data stays where it is, and Juno reads it there

One morning of registration gives every later stage the same tables.

FROM A CONNECTOR TO TABLES JUNO CAN QUERY01Pick a connectorSAP, Oracle, a planning suite, a warehousesystem, a lake or a file share02Choose the tablesbrowse the source and select the tables thispiece of work needs03Read in placeJuno queries the rows where they live, underthe connection’s rightsa sync scan lists what is new, changed or gone in the source
The rows stay in the system that owns them; the dashed line is the sync scan. Product.

Registration is how a table becomes something Juno can query. A. Novak, the data lead, picks a connector, gives the connection and chooses the tables; S. Iyer, the IT architect, sets up access and security. Juno keeps the names, types, keys and a sample of values; the rows stay in the source. Data registration

On the Monday of the first week, at 09:10, A. Novak opened read connections to SAP S/4HANA and Oracle Fusion. By noon the estate held 23 tables from five systems; the agent’s purchase order lines came from EKPO in SAP and PO_LINES_ALL in Oracle. Data in place

Stage two · Map

One shared map, and how the agent read two systems as one

The stage that answers the question chapter 5 left open.

Chapter 5 closed on the agent reading purchase order lines from SAP and from Oracle and treating them as one thing. This is where that was settled. The supply chain ontology is Lima’s built-in map of supply chain things and how they relate: materials, plants, suppliers, purchase orders and their lines, stock, demand and their measures.

Mapping attaches your columns to that map. EKPO.MENGE in SAP and PO_LINES_ALL.QUANTITY in Oracle both became the ordered quantity of one object, the purchase order line. From then on a question, a policy and an approval queue all name that object. The ontology model

YOUR SYSTEMSTHE SUPPLY CHAIN ONTOLOGYEVERYTHING THAT READS ITSAP S/4HANAEKKO · EKPO · EKETOracle FusionPO_HEADERS_ALL ·PO_LINES_ALLPlanning suitecoverage · exceptionsWarehouse systemstock · movementsSupplier portalconfirmationsChatLimaL. Tan’s questionAgentsthe cancellation policySmartAppsD. Okafor’s queueBI toolsthe same definitionsAPIsother systems read it01Objectspurchase order line, material,plant, supplier02Attributesordered quantity, cancellationwindow, requested date03Relationshipsa line belongs to an order, refersto a material at a plant04Measuresprojected excess, days of supplyone meaning · one version · one owner
Five registered systems feed one ontology, and ChatLima, the agents, SmartApps, BI tools and APIs all read it. Showcase estate.

The same word in three systems

Where it is heldWhat the field is calledWhat it means thereOntology attribute
SAP S/4HANAEKET.EINDTThe date the buyer asked for on the schedule lineRequested delivery date · schedule line
Oracle FusionPO_LINE_LOCATIONS_ALL.PROMISED_DATEThe date the supplier promised when the order was placedPromised delivery date · purchase order line
Supplier portalCONFIRMED_SHIP_DTThe date the goods leave the supplier’s dockSupplier confirmed ship date · supplier confirmation

One business word, three meanings. The ontology keeps them apart as three attributes, each with its own definition, and the cancellation policy names the one it tests. Showcase estate.

The Tuesday review: 412 columns, two people, 2 h 10 min

Mapping review · Tuesday 15 September · 5 of 412 columns · version 1 published at 10:40
Registered columnOntology attributeConfidenceStatusWhy Juno proposed it
EKPO.MENGESAP · 12,000.000→Ordered quantityPurchase order line98%confirmed · A. Novakstandard SAP field · numeric with a unit of measure
PO_LINES_ALL.QUANTITYOracle · 500→Ordered quantityPurchase order line95%confirmed · A. Novakstandard Oracle field · same value shape as EKPO.MENGE
EKPO.ZZ_CXL_WDWSAP · 21→Cancellation window (days)Purchase order line62%confirmed · D. Okaforcustom field · name suggests “cancel window” · whole numbers 0 to 90 · reviewed 09:20
EKKO.BSARTSAP · NB→Purchase order typePurchase order71%adjusted · D. Okaforproposed as document category · the buyer set it to order type
CONFIRMED_SHIP_DTPortal · 2026-09-30→Supplier confirmed dateSupplier confirmation·added · A. Novakplaced below 60% · mapped by hand at 09:42

Five of 412 columns. The last column holds Juno’s reason; the status names the person who decided. Showcase estate.

By Tuesday morning Juno had proposed a meaning for 371 of the 412 columns, each with a confidence and a reason. A. Novak opened the queue at 08:30 and confirmed 338 in one action. D. Okafor joined at 09:15, adjusted 33 and settled the custom field that decides the case, the cancellation window, placed at 62%. The team added 41 by hand; version 1 was published at 10:40. Mapping review

Every confirmation and correction is kept as evidence for the next table; Juno also links the same supplier numbers across SAP and Oracle. The published mapping is a version, so the Wednesday question and the Friday policy work from the same meaning. Linking across systems

Stage three · Ask

The same meaning answers a question in 4.2 seconds

The stage that changes the week first: the answer arrives while the window is open.

ChatLima · L. Tan · supply planner · Wednesday 09:12showcase estate
LT
Which open purchase order lines are still inside their cancellation window this week, and what are they worth?
L

31 open purchase order lines worth $348,000 sit inside their cancellation window at plants 1210 and 1220, and the window closes on Friday 18 September. Two suppliers hold 24 of them.

SupplierLinesValueWindow closes
Delta Fasteners Ltd15$186,200Friday 18 September
Brightline Metals SA9$121,300Friday 18 September
4 other suppliers7$40,500Friday 18 September
purchase order linecancellation windowprojected excessSAP S/4HANAOracle Fusionplanning suite
Show the query
SELECT s.supplier_name, COUNT(*) AS lines, SUM(l.open_value_usd) AS value
FROM   ontology.purchase_order_line l
JOIN   ontology.supplier s ON s.supplier_id = l.supplier_id
WHERE  l.status = 'OPEN'
  AND  l.cancel_window_end >= CURRENT_DATE
  AND  l.projected_excess_weeks > 8
GROUP  BY s.supplier_name ORDER BY value DESC
4.2 s3 sources read in placeyour access: plants 1210, 1220
Which of these lines are marked NCNR?Show the same list by plant

Five steps sit between the typed sentence and the answer. ChatLima recognises the intent and the business objects, frames the goal and finds the tables mapped to those objects. It writes the query across one source or several and runs it with the asking person’s access. The answer carries its working: the terms, the systems, the time and the query. From a question to the right query

The object in the query is the same purchase order line the agent acts on, with the same cancellation window and projected excess. A person asking and an agent acting read one meaning, so the Wednesday answer, that week’s 31 lines, and the Monday drafts, the 38 lines the push-out left, agree on what a line is. ChatLima has been in production since 2023.

Stage four · Teach

Your own words, taught once, for questions and agents alike

Vocabulary is one half of teaching; feedback on answers and on drafts is the other.

TermWhat it means in this companyKindAdded byWhen
NCNRNon-cancellable, non-returnable: a line the supplier holds to the agreed dateAcronymD. Okafor · buyerThu 09:00
LT fenceLead time fence: the date inside which the plan stays firmAbbreviationL. Tan · supply plannerThu 09:20
“strat supplier”The strategic class on the supplier recordAliasP. Sharma · master-data stewardThu 09:40

Three of the 96 entries in the first vocabulary file, which P. Sharma owns and reviews. Showcase estate.

Two Wednesday answers came back wider than expected: the questions used words Juno had yet to learn. On the Thursday three people added three terms in forty minutes: the buyer NCNR at 09:00, the planner the lead time fence at 09:20, the master-data steward the strategic supplier alias at 09:40. As owner of the vocabulary file, P. Sharma reviewed all three. Training

Training ran in the background in six minutes. From then on a typed acronym is understood on the first try. A later version of the policy can say “keep NCNR lines” and the rules engine reads the same definition, and the approval queue uses the buyer’s own words.

Mark an answer right or wrongA wrong answer names the term ChatLima misunderstood.
Correct the termThe correction changes the vocabulary; the next answer from anyone uses it.
Save a verified pairA question and its checked answer become the answer everyone sees.
Approve, edit, rejectEach one on an agent draft is feedback too, recorded and used in the next cycle, as chapter 4 showed. Feedback
Stages five and six · Act and Learn

The Friday and the Monday you have already watched

The two stages you watched most closely run on the four before them.

ACT AND LEARN: THE FRIDAY, THE MONDAY AND EVERY MONDAY AFTER01State the policyL. Tan, Friday 09:00, inplain language02Planning tablefive rows · simulated onlast quarter · version 303Rules engine runsMonday 06:00 · 38 linesmatch version 304Named approvalD. Okafor · 29 drafts · A205Write-backconfirmedthe ERP’s acknowledgementon each lineLearn: every approve, edit and reject is recorded and shapes the next version
Act runs left to right, from the policy sentence to the confirmed write-back; Learn is the dashed line back. Showcase estate.

Act is the Friday and the Monday of chapter 5. L. Tan’s policy sentence became a five-row planning table, released by J. Ridley as version 3 at A2 with D. Okafor as the named approver. Each morning at 06:00 the Lima Rules Engine runs those rows the same way. On the first Monday 38 lines matched, 29 drafts reached the queue, and each approved cancellation was written back with the ERP’s acknowledgement. Planner Agent · Write-back

Learn is the loop chapter 4 drew. Each decision is stored with the rule version, the facts at that moment, the recommendation, what D. Okafor did and the outcome at day 14. After six weeks the record read 212 drafted, 203 as written, 9 edited, and it carried the level review that moved lines under $25,000 to A3. From the buyers’ behaviour Juno proposed the exclusion j.park approved as version 4. Learning loop · Audit trail

Standards

The ontology lines up with standards your teams already use

Four public standards supply the process names and identifiers your teams already use.

StandardWhat it coversWhere it shows in the ontology
SCOR Digital StandardA process reference model for plan, order, source, transform, fulfil and return, with its performance metrics.Process names and measure definitions such as the perfect order and cash-to-cash.
APQC Process Classification FrameworkA common way of naming business processes, used across industries.The phases and process steps the agent blueprint is organised by.
GS1 identification keysGlobal identifiers for trade items, parties and locations.Identity attributes that let one supplier or one location be recognised across systems.
ISA-95 · IEC 62264The vocabulary of the plant: equipment, material, personnel, operations schedule and performance.Plant-side objects and the link from a production order to the line that runs it.

Lima keeps its own business vocabulary and documents where it lines up with each of these, so an auditor can trace a measure back to a public definition.

Terms on this page
Supply chain ontologyLima’s built-in map of supply chain things, their attributes, relationships and measures.
MappingThe confirmed, versioned link between one column in your system and one attribute in the ontology.
Knowledge indexJuno’s record of every registered table, column and value, held in business language.
Planning tableThe reviewable rows a policy compiles to, which the rules engine runs the same way every time.
Questions
Do all six stages have to be finished before the first answer?

Two registered systems with their columns mapped are enough for the first questions. Teaching, agents and the learning loop follow on the same data.

Can we use our own data model in place of the Lima ontology?

Yes. The built-in ontology takes your own objects, attributes and relationships as extensions, and a model of your own can be loaded as the target of mapping. The review screen is the same for both.

The question this leaves

Where does the language model sit in all this, and what does it get to see?

Answered next

Architecture and trust

Design time and run time, the boundary your row data stays inside, and the record an auditor reads.

In the platform

Where Juno sits among the four capabilities

Each one builds on the one before it.

One shared foundationCatalogLineageAccess policyWrite path

See Juno answer and act on your own data

Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.