Workflow and actors
The data team prepares order, catalogue and customer history; models produce forecasts, classifications, churn risk and commercial scores. Procurement, master-data and sales teams consume each output in their existing workflow.
Service · Custom AI · Distribution & B2B industry
Custom AI for B2B distribution that learns from your orders.
Demand models per SKU and customer, B2B churn prediction, catalog classification, datasheet extractors and opportunity scoring, trained on your ERP and your datasheets, not on a retail benchmark.
The workflows, examples and figures on this page are illustrative composites and modelled targets, not measured client results. In a real project, we define the baseline, thresholds and human review with your data before rollout.
What we solve
Your catalog doesn’t look like Amazon.
In B2B distribution, demand isn’t driven by consumer seasonality: it depends on orders per customer, project cycles, industrial calendar and negotiated pricing. Standard ERP forecasting treats a customer who orders weekly the same as one ordering three times a year. Catalog classification and deduplication get done by hand every time a new supplier comes in.
We build models trained on your real data: daily forecast per SKU and customer, B2B account-churn scoring, catalog classifier for deduplication and categorisation, field extractor from datasheets and spec sheets, and commercial-opportunity scoring. Integrated with SAP Business One, Holded, Odoo and Microsoft Dynamics 365 without touching your master-data model.
Implementation contract
How it runs in production
Systems and data
SAP Business One, Holded, Odoo or Dynamics supplies orders, inventory, catalogue, families, customers and returns; the CRM supplies activity and closing data for training and evaluating commercial signals.
Exceptions and risks
New SKUs or customers, duplicate master records, families with limited history, sudden demand changes, drift and low-confidence predictions are flagged for specific handling rather than forced through automation.
Human review
Procurement approves replenishment changes, master-data owners validate categories and equivalences, and the sales lead decides any action involving an at-risk account or prioritised opportunity.
Implementation pattern
Versioned pipelines by family, batch training, scheduled or API inference, and quality and drift monitoring. Every prediction retains its version, features and confidence level.
Relevant integration
The ERP remains the system of record; predictions return to extended fields in SAP Business One, Holded, Odoo or Dynamics and to the CRM without changing the native data model.
What we build for this sector
Use cases that ship to production.
Forecast
Demand model per SKU and customer
Daily prediction per SKU and per SKU–customer pair using your order history, industrial calendar, sector behaviour and negotiated pricing, feeds replenishment and lets you anticipate specific drops, not just aggregate ones.
Modelled target: MAE −35–50% vs ERP forecast, by family
Predictive
B2B customer-churn predictor
Model trained on lost accounts: estimates 90-day churn probability combining order frequency, average size, product mix, logistic incidents and shifts in buying behaviour, to open a real retention window.
Modelled target: Recall > 0.7 on at-risk accounts
Classification
Catalog classifier and deduplicator
Categorises products into your taxonomy, identifies duplicates across suppliers and proposes merges by equivalent key, critical to onboarding new catalogs without your master-data team spending weeks mapping.
Modelled target: Precision > 0.95 on deduplication, by family
Extraction
Field extractor from datasheets and spec sheets
Reads PDF datasheets and spec sheets (multilingual, with tables and mixed units), extracts structured technical attributes and normalises units, ready for ERP and PIM with no retyping.
Modelled target: 30–80 fields per sheet, by category
Commercial
Commercial-opportunity scoring
Predicts close probability and expected deal size using account history, ordered SKU mix, buyer behaviour and activity signals, so the sales team prioritises accounts with a real probability of closing.
Modelled target: Top 25% opportunities = 65% of closings
Illustrative composite scenario · modelled figures and targets, not client results
A wholesaler with 12,000 SKUs and 1,800 active accounts.
Industrial distributor on SAP Business One, four years of order history and a catalog that grows 8% a year through new suppliers. Illustrative composite scenario; the figures below are a modelled baseline and targets, not measured client results.
Modelled baseline
Baseline assumption: Forecast by family and month with an ERP model, average per-SKU error of 28%. Replenishment by gut on long-tail SKUs, with frequent stock-outs and excess on others. Onboarding a new catalog (3,000 SKUs) takes the master-data team 4–6 weeks. Account losses are spotted when the rep calls and the customer is already buying elsewhere.
Modelled target
Modelled operating target: Per-SKU–customer forecast with average error of 14% and an early signal for critical SKUs across the top-200 accounts. Catalog onboarding is done in 5 days with human review only on low-confidence cases, not on all 3,000. Churn flags 72% of at-risk accounts 60–90 days ahead, giving the rep room to act. Opportunities arrive in the CRM prioritised by score.
Modelled target: −42% stock-outs on A SKUs · −18% inventory on C SKUs
We connect to your stack
Integrations that matter in this sector
CRM
Salesforce
Enterprise CRM with fine-grained permissions, AI workflows that respect the data model.
CRM
Microsoft Dynamics 365
Enterprise CRM/ERP suite in the Microsoft ecosystem, native fit with 365 and Power Platform.
ERP
SAP Business One
Reference ERP for mid-market distribution and manufacturing, document extraction and ops orchestration.
ERP
Holded
Spanish cloud ERP widely adopted by SMBs, invoicing, expenses and reconciliation automation.
ERP
Odoo
Modular open-source ERP, AI agents and workflows on top of sales, inventory and project modules.
COMMS
Microsoft 365 / Outlook
Email, calendar and SharePoint as channel and context, triage, drafting and RAG over your inbox and files.
Frequently asked
What clients ask us
- 01
Our master data isn’t clean. Will it still work?
No, and we say so in the first meeting. Quality and consistency in product master, customer codes and initial categorisation are the foundation, without that, no model holds up. The good news is that part of the project is exactly that: the catalog classifier and deduplicator clean and normalise as we train. We start with families that have healthy data, prove results, and extend from there.
- 02
How does it integrate with SAP Business One, Holded or Odoo without changing the data model?
We connect via standard API, intermediate connector or scheduled exports, depending on the ERP and version. Predictions (forecast, churn, scoring) are written into custom fields or extended tables, without modifying the native model. If you need traceability for audit or ISO, we keep an audit log of every prediction and the model version that produced it.
- 03
What happens when a new SKU or a new customer comes in with no history?
For new SKUs we start with cold-start based on similar SKUs (same supplier, same family, comparable technical attributes) and refine as orders accumulate. For new customers we use sector, size and initial-mix profiling. The initial prediction carries a lower confidence flag and updates automatically, the team always knows how much to trust the number.
- 04
How long until it’s in production and when do we see results?
The first model (typically family-level forecast or catalog classification) usually reaches production in 8–12 weeks, depending on initial data quality and pilot scope. From month three we compare forecast error, churn and classification time with the real baseline; any improvement depends on the data and operation and is not guaranteed in advance.
Same sector, other services
Trust
Safe, traceable AI,
enterprise-ready.
We design for privacy from the start, human control, traceability, usage limits, permissioning and documentation. For sensitive processes, we help assess risk and applicable obligations under GDPR and the EU AI Act.
- 01We never train models on your data without explicit authorization.
- 02Human review built-in for processes where risk demands it.
- 03Traceability: prompts, sources, permissions, errors and metrics, all documented.
- 04Privacy, security and control integrated from day one.
- 05Solutions engineered to be maintained, audited and improved over time.
GDPREU AI ActAEPDISO 27001 readyEU data residency
Personal diagnosis
We work with
few clients.
Every engagement is led personally by one of the partners. If there's a fit, you get a personal first read of your case within one business day, not a canned demo.
How we work
- 01Tell us which process eats your time
- 02Personal reply within one business day
- 0320-minute call, no demo, no pitch