AI agent / WhatsApp / B2B operations

OXALIS

A DIGITAL PRODUCT BUILT FOR FLOWERS MARKET

Oxalis, the AI agent that turns WhatsApp conversations into controlled B2B orders.

Oxalis understands text, voice messages and images, uses catalogue and availability data to guide customers, prepares an order draft and requests confirmation before the team takes over. When the available information is not sufficient, the conversation reaches a person.

Oxalis flow from WhatsApp conversation to interpretation, draft, confirmation and team takeover
AI moves the conversation forward while confirmation and exceptions remain controlled.

Text, voice and imagesone WhatsApp conversation

Operational datacatalogue, packaging and availability

Explicit confirmationa draft before the order

Human in the loopcontrolled handoff and review

From conversation to operation

Not a chatbot that improvises. A digital colleague with clear boundaries.

In B2B distribution, a WhatsApp conversation is rarely a simple question and answer. A customer may send an incomplete flower name, a photograph, a colour, a quantity or a quick voice message. Behind a useful answer sit a catalogue, packaging rules, availability, commercial context and a team that must be able to continue the discussion.

Oxalis was built for this context at Flowers Market Holland. It does not replace the webshop or the team. It reduces the distance between the customer's natural language and the structured information required by the operation.

The agent belongs to the ecosystem described in the Flowers Market digitalisation case study. This page examines Oxalis separately: the product problem, its public architecture and the rules that keep it useful and controllable.

The product problem

Customers want a simple conversation. The business needs structured data.

01

Customer language

Popular names, abbreviations, typos, colours, photographs and voice notes must be understood before a product is selected.

02

Catalogue reality

Products have families, variants and packaging. An approximate match must never be presented as certainty.

03

Volatile availability

The customer receives useful qualitative guidance, not exact internal inventory or an unsupported promise.

04

Consequential action

A draft can be built in conversation, but sending an order request requires the customer's explicit confirmation.

05

Commercial exceptions

Final price, substitutions, ambiguity and questions without a safe source must reach the team.

06

Human continuity

An operator receives enough context to continue naturally without asking the customer to repeat everything.

How it works

A controlled path from message to order draft.

  1. 01

    Receive the customer's chosen format

    Text, voice or images enter the same conversation. Oxalis requests clarification when the product, colour, quantity or packaging cannot be inferred safely.

  2. 02

    Interpret the intent

    The agent distinguishes product discovery, availability questions, order preparation, controlled repeat orders and requests for a person.

  3. 03

    Consult the right source

    Catalogue data, operational availability and approved knowledge have different authority and update rhythms. They are not mixed arbitrarily.

  4. 04

    Build a verifiable draft

    Discussed products and quantities are gathered in a summary. Informational questions must not silently alter the draft.

  5. 05

    Request explicit confirmation

    The customer sees what will be sent and confirms the intention. Exploration is not treated as a firm request.

  6. 06

    Hand over to the team

    Flowers Market validates the request operationally. If a safe source is missing, the conversation can reach an operator earlier.

Sources of truth

An operational agent must know not only what to say, but how it knows.

Catalogue and packaging

Product identity, variants and rules required to build a coherent quantity.

Operational availability

Current data translated into qualitative guidance without exposing exact internal stock.

Conversation memory

Context for clarifications and the controlled selection of a previous order.

Approved knowledge

Stable, customer-safe answers. A new human response does not automatically become public truth.

Multimodal conversation

Text is only one way customers explain what they need.

Voice messages

Voice can be more natural in warehouses, flower shops and on the move. It is interpreted within the same context, with clarification when names or quantities remain uncertain.

Images

A photograph can indicate a product type, colour or visual reference. It helps orientation but is not treated as infallible identification.

Packaging and quantities

In B2B, “ten” may mean stems, pieces, bunches or packages. The agent connects quantity with the selling unit and clarifies contradictions.

Availability without inventory exposure

The answer is useful enough for a decision—unavailable, low, medium or sufficient—without publishing exact inventory.

Human in the loop

A good human handoff is a product feature, not an AI failure.

Escalation happens when the customer requests it, sources do not cover the question, information conflicts or a commercial judgement is required.

01

Recognise the limit

The agent does not fill gaps with guesses merely to remain fluent.

02

Route the conversation

The appropriate team can take over with context.

03

Keep human control

While an operator is active, the agent does not compete with parallel answers.

04

Return explicitly

After resolution, the conversation can return to Oxalis in a controlled way.

Controlled learning

Good human answers can improve the product without becoming rules automatically.

01

Identify the gap

Separate missing knowledge from data, integration, ambiguity or freshness problems.

02

Propose, do not publish

A reusable answer may become a redacted, deduplicated candidate. Transient prices and stock cannot.

03

Human review

An administrator validates meaning, audience and validity. Conflicts cannot auto-publish.

04

Versions and audit

Approved knowledge evolves through controlled versions that can be traced and withdrawn.

Deliberate boundaries

What Oxalis was designed not to do.

  • No exact internal stock exposure. Customers receive qualitative guidance.
  • No indicative price presented as final. Commercial confirmation belongs to the team.
  • No order without confirmation. Draft and action remain separate states.
  • No forced answer. A handoff is better than invented confidence.
  • No automatic publication of operator replies. Learning requires filtering and approval.
  • No hidden AI identity. Customers should know whether Oxalis or a person is replying.

Transferable lessons

What other B2B companies can learn from Oxalis.

Start with real conversations

Map recurring questions, vocabulary, exceptions and the systems consulted by the team.

Define sources before prompts

Catalogue, ERP, CRM and approved procedures have different authority and freshness.

Separate information from action

Answering, changing a draft and sending an order have different risk levels.

Design handoff from day one

Ownership, context and the agent's return should not be improvised after launch.

Measure usefulness

Track grounded coverage, successful clarifications, handoffs, errors and confirmed drafts.

Launch gradually

Use scenarios, limited pilots, monitoring and a quick return to the human flow.

Read the generic guide to a WhatsApp AI agent connected to ERP, the complete Flowers Market case study and our software development services for SMEs.

Frequently asked questions

What Oxalis is—and is not.

Is Oxalis a WhatsApp chatbot?

It is a conversational agent connected to Flowers Market processes and operational sources, with confirmations and human handoff.

Can it understand voice and images?

Yes. Text, voice and images are interpreted in one conversation, with clarification when confidence is insufficient.

Does it show exact stock?

No. Availability is communicated qualitatively and final operational confirmation remains with Flowers Market.

Does it send an order automatically?

Not from an exploratory conversation. It prepares a summary and requests explicit confirmation.

What happens when it does not know?

It asks for clarification or hands the conversation to an operator rather than inventing an answer.

Does it learn automatically?

New reusable answers remain redacted review candidates until a person approves them.

Does it replace the webshop or sales team?

No. It is a complementary channel that advances conversations and routes exceptions.

Build and validation

A product created where technology meets operations.

AYSA and Web-Developmentproduct direction, design, development and integration

Flowers Marketprocesses, operational validation and use

Marius Dosinescufounder and direction coordinator

See the customer-facing Oxalis page, the AYSA case study and Web-Development.ro.

AI connected to real work

Do you have repetitive conversations that could become a clearer operational flow?

We start with customer questions, sources of truth, exceptions and the decisions that must remain human. Only then do we design the agent and its integrations.