How to add an AI chatbot to Macola
Short answer
An AI chatbot for Macola answers questions by retrieving live order, stock and invoice data from curated views and your documents, then responding in plain language within the user's permissions. It can be an internal assistant, a customer portal widget, or both, and can run fully on-prem.
Most Macola support questions are repeats: where is my order, do we have this part, what was invoiced, what are the terms. They consume order desk and planner time because the ERP screens are built for transactions, not for answering a caller in thirty seconds.
A chatbot changes that only if it is grounded. The failure mode to avoid is a friendly assistant that invents ship dates. The design below keeps the model on a short leash: it can only state what a tool returned or a document says, and it shows its source.
There are two audiences with different risk. Staff can see more and tolerate rougher edges. Customers need strict scoping to their own account. Build the internal bot first.
What teams use it for
Internal order desk assistant
A CSR types an order or customer name and gets status, ship dates and backorders without opening four Macola screens.
Customer self-service order status
A portal widget lets a logged-in customer ask about their own orders and invoices, restricted by customer ID.
Shop floor part and location finder
A supervisor asks where a part is stocked and how many are free, from a tablet at the line.
Policy and procedure Q&A
Staff ask about return policy, credit terms or inspection procedure and get an answer from approved documents with the source shown.
New hire Macola helper
New users ask how to do a task in Macola and get steps drawn from your own training documents.
Macola integration surface
- Curated read views on the Macola database (SQL Server or ODBC on Progress based installs) behind a few fixed lookup functions.
- Customer identity: for portal use, your web authentication must map to a Macola customer number. That mapping is your responsibility, not Macola's.
- Documents: training guides, quality manuals and terms and conditions, indexed for retrieval with citations.
- Vendor web services or SDK, where present, for optional actions like creating a quote request, handled with confirmation.
- Notifications: escalations can go to email or a ticketing system so the bot hands off to a human cleanly.
How it works
- 1
Collect the top 50 questions
Ask the order desk and planners for the questions they field most. Let those drive which tools and documents you build.
- 2
Build lookup tools over curated views
Order status, invoice status, item availability and customer terms. Each returns structured data and an as-of time.
- 3
Add document retrieval
Chunk approved manuals and policies, index them locally, and return the passage with its source.
- 4
Constrain the model
Instruct it to answer only from tool results or retrieved passages, to say when it does not know, and to offer a human handoff.
- 5
Enforce identity and scope
Pass the authenticated user to every tool. Customer users get a hard filter on their customer number applied in code, not in the prompt.
- 6
Launch internally, then expand
Measure answer accuracy against a sampled review before opening the portal widget.
Example flows
Customer asks about a delayed order
Trigger: Customer types 'when will PO 7712 ship?' in the portal.
- Portal passes the authenticated customer ID.
- Bot calls order status filtered by that customer.
- Tool returns lines with ship dates and backordered items.
- Bot summarizes and offers to open a ticket if the date is unacceptable.
Result: The customer gets an accurate answer without calling, and the CSR sees the ticket only if needed.
Supervisor asks for a part
Trigger: Supervisor asks 'do we have any 2 inch ball valves free?'
- Bot searches item descriptions through the item lookup tool.
- Tool returns matching items with free quantity by location.
- Bot lists the two best matches and asks which one.
Result: Part located in seconds, with locations and quantities from Macola.
Keeping data on-prem
A Macola chatbot can run entirely inside your network. Serve an open-weight chat model such as Llama, Qwen or Mistral with vLLM or Ollama on a local GPU, run retrieval with a local embedding model and vector index, and keep the tools on the LAN beside the database. For customer facing use, place only the chat front end in a DMZ and have it call an internal service over a tightly controlled path; the model and ERP data never sit in the DMZ. A modest GPU handles a small team's internal usage. Customer portals with many concurrent users need capacity planning, and short queues are acceptable for a status bot.
Security checklist
- Apply customer scoping in code on every tool call, never in the prompt.
- Return only fields the audience may see; hide cost, margin and other customers' data from portal users.
- Rate limit and authenticate the public endpoint to prevent scraping through the bot.
- Log questions and answers with tool calls for review, with retention rules agreed with legal.
- Add a refusal path for questions outside scope instead of letting the model improvise.
Ways to get it done
Off-the-shelf chatbot with a Macola data feed
Effort: Days to weeks, but live accuracy is limited by the export cadence.
Small teams needing basic FAQ and exported status data.
Build in-house on an open source stack
Effort: About 6 to 10 weeks for an internal assistant.
A dev team that can maintain retrieval, tools and evaluation.
Partner built, such as Netray
Effort: A scoped pilot of several weeks, then a customer-facing phase.
You want on-prem models, role scoping and an evaluation set done properly.
Pitfalls
- !Letting the bot answer ship dates from memory of earlier turns instead of a fresh lookup.
- !Relying on prompt instructions to separate customers' data.
- !No handoff path, so frustrated customers loop with the bot.
- !Indexing outdated procedures and having the bot confidently quote them.
- !Skipping evaluation, so quality drifts after a Macola update changes a view.
Frequently asked questions
Can the chatbot show real-time Macola data?
Yes, if it queries curated views at question time. Results are as current as the database, including posting delays. Show an as-of time so users understand what they are seeing.
Is it safe to expose to customers?
It can be, with hard customer scoping in code, field filtering, rate limits and a DMZ design. Begin with an internal bot and open the portal only after you have reviewed accuracy and logs.
Do we need a cloud LLM?
No. Open-weight models served locally are adequate for lookups and document Q&A. A cloud model can help with harder reasoning, but only if your policy allows the data to leave.
What stops it from making things up?
Grounding rules, a retrieval and tool-only answer policy, citations, and an evaluation set that you re-run after changes. It will still sometimes be wrong, so keep human handoff available.
Does Macola have a built-in chatbot?
Check with your Macola vendor for current features, as I cannot confirm a built-in one. Most sites get better results from a bot designed around their own views, documents and customer portal.
Want AI chatbot working on your Macola?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment