Smotly integrates generative AI and business agents into your e-commerce platforms, PIMs, ERPs and business portals. Our experts orchestrate models, secure your data and industrialize what worked in POC. From scoping through run-phase improvements. Zero vendor lock-in.
The problem
According to Gartner and IDC studies published in 2024-2025, more than 80% of generative AI proof-of-concepts never reach production. The cause is almost never the model. It lies in five blind spots that pure-play AI agencies systematically underestimate.
A POC running in a Notebook or local Streamlit says nothing about the system's ability to integrate with your ERP, PIM, CRM, SSO directory and data lake. Moving to production is exactly that work.
An AI assistant on internal documentation assumes that the documentation is clean, classified, deduplicated and indexable. In 9 cases out of 10, it is not. RAG does not save a dirty knowledge base; it spreads its defects at scale.
GDPR, audit trail, secrets management, prompt traceability and log retention are constraints that break POCs when they were designed without them. IT security arrives for production review, and everything has to be redone.
A POC calling GPT-4 for 50 internal users costs a few dozen euros. The same usage at the scale of an e-commerce site, a 200-agent support team or a 100,000-reference catalog can quickly exceed 5,000 euros per month. Without a cost model, the project is challenged precisely when it starts working.
An AI agent imposed without training, feedback channels or guardrails is rejected in six weeks. The cause is cultural, not technical, and it is the hardest one to repair after the fact. This is also why, at Smotly, change management is not an optional module: it is a deliverable in every AI production rollout (see Methodology section).
Putting AI into production requires exactly what any other information-system project requires: scoping, clean data, mature governance, a cost model and change management. That is precisely what Smotly has been doing for 15 years, now with an AI layer.
Our posture
The distinction has major consequences. A pure AI agency delivers a model that performs well on evaluation prompts. A SaaS integrator installs an add-on in a suite you already pay for. Smotly does something else : we design and operate the platforms (e-commerce, PIM, ERP, portals) on which AI creates value, and we integrate AI agents as another building block of the information system, not as a parallel project. Our POCs start from a quantified business use case, our agents run on your real data from week two, and production does not require a new project because it is in scope from the start.
Pure-play AI agency
R&D, models, prompts
SaaS integrator
Vendor suite (Salesforce, Microsoft...)
Smotly
Information-system design + AI layer
Who it's for
AI requests are coming from every business department. You need a partner who understands your information system, can choose between a France-hosted model, a private Azure OpenAI model or a self-hosted open-weight model, and owns governance end to end.
Your PIM contains 50,000 to 500,000 references. Product sheets are incomplete, translations are late and product SEO is dormant. You need a partner who can industrialize without breaking Akeneo, Strapi or Medusa governance.
You process thousands of invoices, delivery notes, contracts and customer emails every month. You know OCR alone is not enough and dedicated SaaS tools do not cover your edge cases. You want to automate without reinventing everything.
Your members, distributors or employees search for information in a dense, sometimes regulatory document base. A well-built AI agent saves hours per day, provided it does not hallucinate and remains traceable.
Two families of use cases
Our AI services fall into two families with distinct logic. The first augments the platforms we design (e-commerce, portals, corporate sites) with AI features visible to your users. The second automates your internal operations (catalog, back office, support, documentation) with agents that work in the background or alongside your teams.
Visible to your end users
In the background or with your teams
↓ SmotFlow methodology · scoping → POC → run → improvement
Our AI services
Smotly works across the full lifecycle of an AI project: audit, scoping, POC, build, production rollout and application maintenance. Six services cover almost all the requests we receive. Every service is fixed-price or milestone-based, never billed as open-ended time and materials.
2 to 4 weeks
We audit your information system across 5 axes: AI use cases to prioritize, data quality and accessibility, regulatory constraints, relevant inference models and change management.
4 to 6 weeks
Scoping, build and measurement of a single AI use case on your data, in your environment. No isolated Notebook: the POC is delivered in a pre-production stack, with governance, observability, production estimates and an adoption plan by the end of the POC.
10,000 to 500,000 sheets · 5 to 25 languages
Large-scale generation, enrichment, normalization and translation of product sheets, integrated into your PIM (Akeneo, Strapi, Plytix, Pimcore) or e-commerce back office (Medusa, Shopify). Human validation can be configured to match your governance.
Stack: pgvector / Qdrant + MCP + Langfuse
AI agents connected to your internal documentation, business databases or SaaS tools through MCP. The agent answers in natural language, cites sources, refuses out-of-scope questions and logs every exchange.
ERP, CRM, accounting, contracts
End-to-end automation of back-office workflows: OCR extraction + LLM structuring of incoming documents, routing and categorization, ERP/CRM writing, anomaly alerts and summaries for human validation. Orchestrated in n8n.
B2B · DTC · marketplaces
Natural-language search, AI concierge and contextual recommendations ("I'm looking for a product for...", "I need to replace reference X"). Connected to your Medusa, Strapi or Shopify catalog and your B2B pricing logic.
Concrete use cases
For each major area of the information system, here are the AI use cases we actually put into production in 2026, with orders of magnitude observed on our projects or on peer projects.
E-commerce
Multilingual product-sheet generation
5 to 25 languages, preserved brand voice and worked-through SEO. Typical gain: 6 months of catalog delay recovered in 4 to 6 weeks.
Automatic attribute enrichment
Materials, dimensions, compatibilities and keywords generated from existing descriptions and visuals.
Conversational search
"I'm looking for a 50-liter compressor compatible with my tools", "replace this obsolete reference".
AI recommendations & pre-sales concierge
Contextual B2B cross-sell, visitor qualification and handoff to sales.
Review and product Q&A moderation
Filtering off-topic content and suggesting replies for customer support.
We do not sell what we do not practice. Smotly uses internally orchestrated AI agents across its own project phases: scoping, graphic design, build, acceptance testing, documentation and application maintenance. The "AI-augmented" differentiator is not a marketing claim; it is our daily operating model.
Sector readings
The IT perimeters above are expressed differently depending on your sector. Here are the use cases we actually see emerging in 2026, sector by sector.
Multilingual PIM industrialization, technical-sheet generation from supplier datasheets, lead qualification agent, conversational search across catalogs of 50,000 to 500,000 references and complex quote automation.
Dominant pain point: catalog backlog that hurts SEO and customer experience.
OCR and extraction for delivery notes and purchase orders, technical support agent over product documentation, failure prediction from intervention history and pre-filled field reports.
Dominant pain point: scattered technical documentation and loss of senior knowledge.
Member assistant over regulations and FAQs, topic summaries for committees, routing agent toward the right contact or training, and automation of incoming requests.
Dominant pain point: repeated questions saturating permanent teams.
Compliance agent over the regulatory corpus, KYC and document-classification automation, client-file summaries for advisors and product assistant over terms and conditions.
Dominant pain point: compliance, which requires self-hostable models and a rock-solid audit trail.
Citizen guidance agent for administrative procedures, semantic search across deliberations and decrees, incoming-mail processing automation and file summaries for case officers.
Dominant pain point: data sovereignty, requiring French or EU infrastructure and model transparency.
Key takeaway: whatever your sector, the success pattern is the same: clean data, governance from scoping, a model aligned with sensitivity and integrated change management. What changes is the tone, vocabulary, regulatory constraints and priority use-case mix.
Related reading
Our AI stack
We do not sell a model or a vendor. For each project, we choose the right tool for the right use, with strong requirements for portability, sovereignty and governance.
Inference models
OpenAI (Azure OpenAI for EU compliance), Anthropic Claude, Mistral (French sovereignty), Llama / Qwen / DeepSeek for self-hosting. Choices are arbitrated case by case.
RAG & vector databases
Postgres/pgvector for operational simplicity, Qdrant or Weaviate for high volumes, hybrid Elasticsearch for semantic + lexical search.
Agent orchestration
LangGraph, n8n and MCP-native in-house frameworks. No imposed proprietary framework.
Tooling protocol: MCP
Model Context Protocol connects agents to your SaaS tools (CRM, ERP, ticketing) without rewriting custom connectors.
AI observability
Langfuse, Helicone and OpenLLMetry to trace prompts, costs, latency and quality.
Governance
Versioned system prompts, auditable logs, configurable retention, PII anonymization on the fly and per-use-case kill switches.
Information-system integration
SAP, Dynamics, Sage, Odoo, Salesforce, HubSpot, Akeneo, Algolia, Strapi, Medusa and Shopify connectors.
Hosting
AWS, OVH, Scaleway or your private cloud. French or EU sovereignty can be covered end to end.
Methodology
Every AI project at Smotly follows the SmotFlow methodology: six connected phases (scoping, design, build, acceptance testing, run, improvement), with internal AI agents assisting our experts at every step and senior humans validating every deliverable.
On AI projects, this means scoping that estimates production rollout before the POC, a build that addresses governance and observability from the first iteration, acceptance testing that measures model quality and inference cost, and a run phase that monitors model drift over time. No AI project is delivered as an isolated POC thrown over the production wall.
Scoping
Costed use case + production rollout plan
Design
Data architecture + security + observability
Build
POC → pre-production on the target stack
Acceptance
Model quality + inference cost + UAT
Run
Drift monitoring + application maintenance
Improvement
New use cases, new models
Change management
A technically working AI that nobody uses is a failure in its own right. More than 80% of AI POCs that die before production do so less for technical reasons than because users were not brought onboard. That is why every Smotly AI project includes change management from scoping, aligned with the target audience, business sponsor and organizational culture.
From the POC, we identify advanced users who will help spread the right practices.
Persona-based sessions for the target user profiles before production rollout.
Your sponsor tracks who uses what, how often and with what satisfaction.
A dedicated channel surfaces out-of-scope questions, hallucinations and gaps to feed iterations.
Usage retention happens after production rollout, not during the first two weeks.
Measured adoption KPIs
| KPI | Typical target | Why |
|---|---|---|
| Weekly usage rate | > 50% of the target at 90 days | Measures whether AI has entered the work routine |
| Useful-answer rate | > 80% | Measures perceived quality, not just model quality |
| 90-day retention rate | > 70% | Measures whether usage lasts over time |
| Out-of-scope question volume | < 15% | Measures whether the agent scope is correctly calibrated |
| Average time-to-proficiency | halved vs manual onboarding | Measures whether adoption accelerates |
Note: Smotly is an IT integrator, not a SaaS adoption-platform vendor. On very large-scale rollouts (> 5,000 users), we work with specialized AI change-management partners. We own the scope we know how to deliver and integrate the other building blocks when the project requires it.
Governance & security
AI in production without governance is a major operational and regulatory risk. Smotly integrates governance constraints from scoping, never at production review.
Incoming data mapping
Nature, sensitivity, legal basis and retention period.
Model choice aligned with sensitivity
Self-hosted open-weight models for sensitive data, EU-managed models otherwise.
PII anonymization on the fly
Before sending data to a managed model when relevant.
Complete audit trail
Who asked which question, which prompt, which sources, which answer, which latency and which cost.
Configurable guardrails
Allowed topic perimeter, out-of-scope refusal and prompt-injection detection.
Sources cited by default
For RAG agents, with links to the source document.
Kill switch per use case
Disable an agent in under 5 minutes without touching the code.
AI Act compliance
System classification, technical documentation and risk register.
Reversibility plan
Change model, provider or host without rewriting agents, through MCP and internal abstractions.
Comparison
The enterprise AI market is structured around three types of players. None is wrong; they answer different needs. Here is how each one responds:
| Criterion | Pure AI agency | SaaS integrator | Smotly |
|---|---|---|---|
| Core skill | Models, prompts, fine-tuning | Feature activation in a suite | Custom information-system platform + AI |
| Ideal use case | AI R&D, advanced NLP | Already-paid suite, AI included | Business platform to augment |
| Real data, real information system | No | Partial | Yes |
| Production rollout included | No | Partial | Yes |
| Governance from scoping | Partial | Partial | Yes |
| Maintenance & run-phase improvements | No | Partial | Yes |
| Zero vendor lock-in | Partial | No | Yes |
In short: call a pure AI agency for an R&D topic or a specialized model. Activate AI in your SaaS if your use cases fit inside it. Call Smotly when your AI has to live inside a custom business platform, on your data, with your governance, and without locking you into a vendor.
Client cases
Extracts from recent AI projects: sectors are anonymized for cases that cannot be named publicly. For a detailed case in your sector, ask us for a tailored presentation.
PIM industrialization
B2B distributor · 80,000 references
Context: 9 months of catalog delay on multilingual product sheets. Undersized content team and outsourced translations that were too slow.
Solution: Generation + enrichment agent connected to Akeneo, configurable human validation, 8 languages.
RAG business agent
Professional federation · 12,000 members
Context: Dense regulatory knowledge base, saturated hotline and members unable to find answers independently.
Solution: RAG agent on Qdrant, MCP toward the extranet, systematically cited sources, audit trail and adoption plan with 18 internal champions.
Adoption results
Back-office automation
Industrial mid-market company · 12,000 invoices/month · Sage ERP
Context: Manual supplier-invoice entry, accounting team under pressure and error rate around 4%.
Solution: OCR + LLM extraction + n8n + Sage connector, with human validation only on anomalies.
FAQ
An enterprise AI agency designs, deploys and maintains generative AI and automation solutions for organizations by integrating them into the existing information system. Smotly is an IT integration agency for e-commerce, PIM, ERP and portals that adds an AI layer to the platforms it builds, rather than a pure AI agency delivering isolated models.
AI audit, POC scoping, catalog industrialization, business agents, back-office automation: we call you back within 24 business hours with a senior contact who understands your information system.
