Reported operating use · Consumer goods and order-to-cash operations
Unilever NOVA: AI-supported order processing
Microsoft reports NOVA in 11 pilot markets and thousands of orders processed in production. The source attributes reported speed and touchless figures to Unilever and supplies no independent measurement method.
Read and check customer orders, then route incomplete or exceptional cases to people
Unilever
Microsoft’s 26 October 2026 account reports NOVA in 11 pilot markets and says thousands of orders have been processed in production. It describes an operating workflow, while the expansion beyond those markets remains planned.
A published third-party example, not a Nuvepro customer case or evidence of a partnership.
The workflow at a glance
Unilever’s NOVA order-processing account
Illustrative summary of the published account
- Customer order input
Orders arrive in varied formats
Microsoft says around two million annual orders still arrive through email, PDFs, spreadsheets and photos of paper forms.
- NextFour NOVA agents
Read, interpret and extract the order
Four agents read the email, interpret its context and extract order information.
- NextFifth NOVA agent + people
Check orders and review exceptions
A fifth agent checks the work. NOVA is reported to read, check and enter orders; missing information and exceptions go to a person. The source does not specify the post-review route or exact entry authority.
What the source reports
Microsoft reports that Unilever receives around two million annual orders in formats such as email, PDFs, spreadsheets and photos of paper forms. NOVA, built on Microsoft Foundry, uses five agents: four read the email, interpret context and extract the order; a fifth checks the work. Missing information and exceptions go to a person for review. Microsoft says the system reads, checks and enters orders, and that the same technology is being applied to claims. Unilever’s Head of Technology for Customer Operations reports an 11-market pilot, thousands of orders processed in production, 95% of orders touchless and an order first-stage time reduction from 15–20 minutes to around 90 seconds. The article separately reports refund claims processed twice as fast and onboarding reduced by 90%; it does not provide methods or an independent audit for these measures. Unilever plans expansion to more than 50 markets.
Reported technology: Microsoft Foundry multi-agent system NOVA. Scope and outcomes are source-reported.
What the enterprise environment would need
The following is a proposed project design for an enterprise doing similar work. Published details are marked; the remaining systems, data and scenarios would be agreed for your project.
Proposed practice environment
Rehearse customer-order processing with fictional records
Illustrative project design to scope with your enterprise.
- 1 · Set up the environment
Versioned order documents
Load fictional emails, purchase orders, attachments, product records and mock terms into a disposable workspace.
- 2 · Preparation and agent contributions
Extract fields and prepare a mock order
An agent reads the documents, links each extracted field to its source and prepares a proposed record without contacting customers or writing to a live ERP.
- 3 · Independent checks
Reconcile fields and duplicate state
An independent checker verifies customer, ship-to, SKU, quantity, price, currency, terms, document versions and duplicate submissions against a separately prepared reference.
- 4 · People decide acceptance
Order operations accepts the checked case
A named operations owner resolves conflicts and accepts or rejects the exact versioned practice output. Failed checks withhold mock entry and return for repair.
A reviewable evidence package
A source-linked mock order record, field-level check results, exception notes and owner disposition.
If a check fails → withhold acceptance.
The same order is submitted twice with a changed quantity and an obsolete price attachment. Hold entry, preserve both versions, repair the case, rerun independent checks and return it to the owner.
Failure or owner rejection returns to review
- Hold
- → 2 · Revise work
- → 3 · Rerun independent checks
- → 4 · Owner review
Operating boundary. Fictional orders and mock systems only. No customer messages, live orders, payments, ERP writes or production claims. Acceptance criteria and rules must be agreed with the enterprise owner before implementation.
- Systems and tools
- Published: NOVA on Microsoft Foundry. Proposed: a disposable order-processing environment with mock email/document intake, an order ledger and a mock entry endpoint; no live customer or ERP connection.
- Data and state
- Proposed: version each fictional purchase order, attachment, customer, ship-to address, SKU, quantity, price, currency and applicable terms. Keep source documents and extracted fields linked.
- People and handoffs
- Published: missing information and exceptions go to a person for review. Proposed: order operations staff resolve source conflicts and accept only independently checked practice outputs.
- Rules and authority
- The source does not disclose decision rules, approval thresholds or write permissions. Proposed: mock-only order entry; no customer communication, real order entry, payment or ERP write.
- Normal and exception cases
- Proposed: clear order, multiple attachments, duplicate submission, missing ship-to, SKU mismatch, conflicting quantity, stale price, currency mismatch and an unreadable paper-form image.
- Reset and replay
- Proposed: restore the same source-document and policy versions before each run; retain extracted values, checker results, exceptions and owner dispositions for comparison.
A proposed simulation
An illustrative practice design derived from this work, not an executed assessment.
Proposed: process a bounded fictional customer order from source documents to a checked mock order record, with exceptions held for people.
Review gate: A separate checker reconciles every extracted field to the source documents and active mock terms, detects duplicate or conflicting inputs, and the order-operations owner accepts the exact versioned practice record.
Challenge the result: A resubmitted order contains a changed quantity and an old price sheet. The checker must flag both conflicts, prevent mock entry, preserve the source versions and return the case for owner resolution and independent recheck.
What needs checking
- Source boundary: Microsoft reports the five-agent order workflow, human review for missing information and exceptions, 11 pilot markets and thousands of orders processed in production.
- Source boundary: 95% touchless, 8x faster, 15–20 minutes to around 90 seconds, twice-as-fast refund claims and 90% onboarding reduction are figures attributed to Unilever in a Microsoft account; independent measurement methods are not provided.
- Proposed: independently reconcile extracted customer, ship-to, SKU, quantity, price, currency and terms against the exact source documents and versioned rules; detect duplicate submissions.
- Proposed: withhold mock order entry when required fields conflict or fail checks; an order-operations owner resolves the case and accepts the exact checked version. Designed, not executed.
Review environment fidelity, human capability and agent performance separately. The enterprise’s accountable owners decide acceptance of the shared workflow and any operating use.
How an enterprise could build similar work
A proposed build order for an enterprise doing similar work, written by Nuvepro from the published account and our interpretation of the environment and checks above. It is not how Unilever built or operates this work.
- Define the fictional order scopeSelect mock customer orders and required fields such as ship-to, SKU, quantity, price, currency and terms.
- Version source documents and rulesPin purchase orders, attachments, product records and applicable mock terms. Keep the expected field map separate from the agent.
- Prepare an order recordHave the agent extract fields with references to source documents and stage a mock record.
- Check the record independentlyReconcile fields, detect conflicting attachments and duplicate submissions, and hold failed cases.
- Repair and replay exceptionsPreserve the failed evidence, resolve missing or conflicting data, and rerun checks against the corrected version.
- Order operations acceptsAn accountable owner reviews the exact checked practice result. No live order entry or customer action is included.
Your systems, data, rules and accountable owners decide the real project. Agree acceptance criteria, escalation owners and operating responsibility with the people accountable for the work before anything runs on live records. Treat the first run as a rehearsal on test records: compare the draft output with an independently prepared answer, and record what the people reviewing it change.
Questions about this example
- What does Unilever’s NOVA system do?
- Microsoft says four agents read order emails, interpret context and extract order details; a fifth checks the work. Missing information and exceptions go to a person for review.
- Is NOVA in production?
- The 26 October 2026 Microsoft account says the system was in 11 pilot markets and had processed thousands of orders in production. Unilever planned to expand to more than 50 markets; the article does not say that expansion is complete.
- How strong is the performance evidence?
- Microsoft attributes figures such as 95% touchless orders and an 8x speed increase to Unilever. The article does not publish independent audit results, samples or measurement methods.
- What would a similar business environment need?
- A proposed, unexecuted exercise uses fictional order documents, mock terms, field-level source links, independent reconciliation, duplicate detection and an order-operations owner. It makes no live ERP changes.
Shape this around your enterprise
Bring a similar workflow and the people responsible for it. Together, define the records and systems to use, the result your team will build or configure, and how its owners will check it. Nuvepro scopes the simulation, environment and preparation. Agree any existing-agent evaluation, implementation assistance and production responsibilities separately.
Sources and limitations
Curation review: . This is the date this source example was reviewed, separate from source publication dates.
- Microsoft publishes this vendor account and quotes Unilever staff; no independent operational evaluation is provided.
- The account does not disclose model details, full agent prompts, customer-specific business rules, exception taxonomy, identity controls or audit procedures.
- The reported performance figures are company-reported and methods, samples and independent verification are not supplied. Do not treat separate claims/refund or onboarding measures as the order-specific 95% touchless figure.
- The proposed practice environment is fictional and unexecuted; it is not a reproduction of Unilever’s NOVA system or proof of equivalent results.
- Unilever’s multi-agent creating customer value at global scaleSource publication: 2026-10-26