i2Business — Consulting & Training : Intelligence to your business
Home/Case Studies/Industries

Industries

AI applied to the realities of your industry

A connector or an agent only adds value if it speaks the language of your operations. We start from your processes, your existing systems and your regulatory constraints to deliver use cases that hold up in production.

Your processesthe starting point of every project
Your systemsERP, CRM and existing tools taken into account
Pilot firstwith indicators defined upfront, before scaling up
Your datahandled in line with your confidentiality constraints and the law

Choose your industry

Don't see your industry?

Manufacturing & Textiles

Tunisian production sites run heterogeneous lines, face pressure on export lead times and keep largely manual quality documentation. AI first pays off on data that is already produced but never used.

Challenges we see

  • Visual quality control that depends on the operator's eye and is hard to trace.
  • Unplanned line stoppages and spare parts ordered in a rush.
  • Export demand forecasts built in Excel and revised too late.
  • Quality reports and customer audits written by hand for every order.

Use cases deployed

Vision-based quality controlDetection of fabric and finishing defects on line images, with an alert threshold adjustable per item.Goal: less scrap
Predictive maintenanceAnalysis of machine readings to anticipate failures and schedule interventions outside production hours.Goal: fewer stoppages
Demand forecastingForecasting model by item and by export customer, reconciled with the order book.Goal: less tied-up inventory
Quality documentation agentGeneration of audit reports and non-conformity forms from production records.Goal: time saved every week

Systems we connect

OdooLocal ERPsMES / production planningExcel & Google SheetsLine sensorsSharePoint

Related skills & resources

MCPSwiver MCP Connector (ERP)Connect Claude to your Swiver ERP: customers, products, suppliers, invoices, quotes and payments, viewed or created in natural language. 31 tools, 5-step setup with your own API key.AgentsMonitoring & reasoned recommendationsMonitoring and decision-support agent. Use when an open business question requires gathering recent information from the web before deciding — choosing a tool or provider, responding to a competitor move…AgentsProspect hunterIdentifies importers, distributors, wholesalers and buying groups for a product in priority countries. For the lead-finding phase of export prospecting, in any industry.PluginsFollow-up assistantSorts the organization's incoming emails (Gmail), spots key people, identifies the main recipient of each request, classifies it and creates or updates the tracking record in ClickUp. No email is sent: drafts only.
All resources for Manufacturing & Textiles
Discuss this industryRelated training courses

Industrializing AI

From trial to production: how companies industrialize AI

A working pilot is not enough: what matters is what it takes to move into production and stay there. We present real deployments published by the model vendors, described without naming the companies, and draw out what they have in common.

Please read first: these cases are not i2Business engagements and do not guarantee the same results for you. The effects described are those reported by the vendor and its client, without independent audit.

What these cases have in common

Whatever the industry, successful production deployments rely on the same levers.

A narrow scope, then expansion

Start with a narrow, well-defined case, then expand in stages: one category, one region, one request type at a time.

Rapid prototypes

Integration with real systems

The AI reads from and writes to the existing ERP, claims system, messaging, or sensors, instead of living on the side.

ERP and CRM integration

Evaluation from the start

A set of reference examples, continuous monitoring, and comparative tests show whether the AI is doing its job well.

Success criteria

Humans in the loop

Confidence thresholds decide when the AI acts on its own; edge cases and the decisions that matter are still reviewed by a person.

AI agents

Traceability and observability

Every action is logged, and the AI's inputs, outputs, and reasoning are tracked to diagnose and improve.

Support and monitoring

Reusable building blocks

The model is chosen for the task, and connectors (MCP, API) serve multiple uses rather than being rebuilt for each project.

Strategy and architecture

The cases, industry by industry

Public case · not an i2Business engagementOpenAIBanking & Insurance

An industrialized voice assistant for first notice of loss

Major North American insurer · Technology: OpenAI Realtime API

The starting challenge

  • During catastrophic events, claims pour in within a few days and exceed a call center's capacity.
  • Customers expect immediate help right after an accident.

What runs in production

  • An autonomous voice assistant guides customers through their first notice of loss for auto claims, answers questions about their policy, and submits the claim.
  • It is connected to the insurer's claims management systems, orchestration layer, and internal tools.
  1. Customer call
  2. AI voice assistant
  3. Claims system

What made it possible to move from trial to production

  • A narrow scope to start: first notice of loss for a single claim type.
  • A phased rollout: a few regions first, then the entire territory within a short time.
  • Humans stay in the loop: specialists focus on complex cases.
  • Continuous operation, day and night, including during peaks.

What changed, according to the vendor and its client

  • The vast majority of customers who use the assistant complete their claim with it.
  • No more waiting on the phone, even during catastrophic events.

Our take

A narrow scope, a phased rollout, and humans reserved for complex cases: the recipe for a production deployment that holds up during peaks. It translates well to a Tunisian customer service operation.

Caveats

  • No public data on costs, processing times, or satisfaction.
  • The share of customers who choose the assistant is not disclosed.
Case study published by OpenAI Related i2Business services:AI AgentsAI Governance
Discuss a similar case
Public case · not an i2Business engagementAnthropicBanking & Insurance

A layered AI platform to industrialize claims processing

Systems provider for insurers, brokers, and administrators · Technology: Claude API

The starting challenge

  • Documents arrive by email, fax, paper, or phone and must be re-keyed or captured through OCR.
  • According to the provider, this processing took up a significant share of claims adjusters' time (estimate), and most decisions relied on human judgment.

What runs in production

  • An orchestration platform built on the Claude API: document processing (classification, extraction), underwriting support, workflows, and task-specific agents.
  • A compliance layer logs every AI action.
  1. Incoming documents
  2. Layered AI platform
  3. Insurance systems

What made it possible to move from trial to production

  • Model routing by difficulty: lightweight models for simple tasks, a frontier model for long files.
  • Logging of every action, to enable audits.
  • Task-specific agents rather than a single general-purpose assistant.
  • Human judgment retained for the decisions that matter, with the model flagging uncertainty.

What changed, according to the vendor and its client

  • Near-instant document classification.
  • Claims processing delays sharply reduced, from several days to a few minutes according to the provider.
  • A much smaller share of claim calculations still requiring human judgment.
  • Complex calculations prototyped in a single workday.

Our take

A layered platform, traceability of every action, and model choice based on difficulty (and therefore cost): these are the ingredients of industrialized AI, not just a prototype.

Caveats

  • All results are reported by the provider, without independent evaluation; several are estimates.
  • The page is published by the model vendor: the tone is promotional.
  • The goal of largely autonomous decisions is an ambition, not a result.
Case study published by Anthropic Related i2Business services:AI AgentsSupport & Monitoring
Discuss a similar case
Public case · not an i2Business engagementAnthropicRetail & Distribution

A quality-documentation assistant, expanded one category at a time

Major Asian home shopping operator · Technology: Claude, via an enterprise messaging platform

The starting challenge

  • Product launches were delayed by back-and-forth between the quality team and partners.
  • Test reports, documentation, and compliance markings had to be checked through exchanges across multiple channels.

What runs in production

  • An AI assistant analyzes documents and answers routine partner requests, around the clock.
  • It is deployed through an enterprise messaging platform that provides the infrastructure, security (strong authentication, message filtering), and distribution.
  1. Partner question
  2. Document assistant
  3. Consistent answer

What made it possible to move from trial to production

  • A start with a single partner category, then expansion to other categories.
  • A channel partners already use: no new tool to adopt.
  • Security infrastructure provided by the messaging platform.
  • Gradual enhancements, such as automatic certification checks currently being integrated.

What changed, according to the vendor and its client

  • Fewer repetitive requests sent to quality staff.
  • Consistent answers, available at any time.

Our take

An assistant connected to compliance documents, deployed where users already are and expanded in stages: the principle of a search engine over internal documents (RAG) put into production.

Caveats

  • The extent of the drop in requests has no published baseline or measurement method.
  • The improvement in launch times and partner satisfaction comes from a survey with no detailed results published.
Case study published by Anthropic Related i2Business services:RAG & Knowledge BasesRapid Prototypes
Discuss a similar case
Public case · not an i2Business engagementOpenAITourism & Hospitality

Reusable MCP building blocks to make a hotel chain visible in ChatGPT

International multi-brand hotel group, with an integration partner · Technology: ChatGPT (plugin) and MCP

The starting challenge

  • Travelers increasingly plan their stays with AI and shortlist hotels before even visiting the hotel's website.
  • The group wanted to be visible earlier in this planning journey.

What runs in production

  • A ChatGPT plugin lets users see hotels on a map, compare prices, and review amenities, then complete the booking on the group's website.
  • An MCP server and APIs, built by an integration partner, connect the hotel group's offering to ChatGPT.
  1. Traveler in ChatGPT
  2. Hotel group plugin (MCP)
  3. Booking on the website

What made it possible to move from trial to production

  • Very fast development thanks to agent-assisted development.
  • A reusable building block: the same MCP server and APIs also power sponsored listings and future features.
  • Booking stays on the hotel group's website: the existing sales journey is preserved.
  • Attribution measurement put in place to assess the real impact.

What changed, according to the vendor and its client

  • A higher conversion rate than organic search, according to early results.
  • A significant share of campaign bookings attributed to ad views.

Our take

MCP connectors are not just for internal operations: well designed, they can be reused and make a brand visible in AI assistants. Worth considering for a Tunisian hotel operator or tour operator.

Caveats

  • Limited hindsight: these are early results.
  • The comparison is with organic search, not with other paid channels.
  • Results depend on the attribution model used.
  • A distribution and marketing case more than an internal operations one.
Case study published by OpenAI Related i2Business services:ERP & CRM Integration (MCP)AI Agents
Discuss a similar case
Public case · not an i2Business engagementOpenAIFood & Beverage

Ordering agents put into production with continuous evaluation and confidence thresholds

Food distribution platform connecting restaurants, suppliers, and distributors · Technology: OpenAI API (including Realtime)

The starting challenge

  • Orders arrived by email, text message, voicemail, images, and handwritten notes, then were re-keyed by hand into the ERP: slow and error-prone.
  • The hardest part was not reading but the implicit context: customer-specific item references, units, and habits that lived in the heads of the order-entry teams.

What runs in production

  • An ordering agent converts emails, text messages, images, and documents into structured, ERP-ready orders.
  • A voice agent takes orders by phone, including outside business hours.
  1. Order (text, image, voice)
  2. Ordering agent
  3. Distributor's ERP

What made it possible to move from trial to production

  • An optional “autopilot” mode: automation only kicks in when the confidence level is high enough.
  • Human review of edge cases, and learning from corrections.
  • An evaluation framework from the start: reference datasets, continuous monitoring, A/B tests in production.
  • AI-specific observability: inputs, outputs, and reasoning traces.
  • Adoption without changing habits: customers keep ordering by phone, text message, or email.

What changed, according to the vendor and its client

  • A sharp reduction in manual order entry.
  • Sales teams that handle more volume without additional hires.
  • Order taking available day and night.
  • A low error rate, with adjustable automation thresholds.

Our take

The lessons the company cites are those of industrialized AI: start evaluation on day one, even with a small set of reference examples; monitor AI inputs and outputs; set realistic expectations, because AI is probabilistic. For a food wholesaler or manufacturer, order taking is an excellent first use case.

Caveats

  • The reported gains are upper bounds, not averages.
  • Self-reported results; this is a technology company, not a traditional distributor.
Case study published by OpenAI Related i2Business services:Rapid PrototypesAI AgentsERP & CRM Integration
Discuss a similar case
Public case · not an i2Business engagementAnthropicManufacturing & Textiles

From reactive to predictive maintenance, on legacy equipment

European industrial software vendor, with a spirits producer as a client · Technology: Claude API, MCP, Skills, Agent SDK

The starting challenge

  • Data is fragmented: technicians react to breakdowns instead of preventing them. At the industrial client cited, a significant share of repairs was reactive.
  • Equipment is old and environments are demanding: the AI has to work in the field.

What runs in production

  • An assistant analyzes equipment images, correlates sensor data, and detects problems before a breakdown.
  • During repairs, it provides real-time diagnostics right alongside the engineers.
  1. Images and sensors
  2. Diagnostic assistant
  3. Maintenance and parts

What made it possible to move from trial to production

  • MCP to connect existing systems, including legacy infrastructure that is hard to integrate.
  • Skills to manage context drawn from decades of industrial data.
  • An agent that coordinates diagnosis, parts ordering, and final verification.
  • Security and compliance requirements built in from the design stage, essential in critical industries.

What changed, according to the vendor and its client

  • Breakdowns resolved faster.
  • Less excess inventory at utility companies.
  • A shift from reactive repair to predictive maintenance, with savings expected at a pilot site.

Our take

Predictive maintenance starts from data a plant already produces without using it: sensors, breakdown histories, documentation. Connecting AI to what already exists, rather than rebuilding everything, is the shortest path from trial to production.

Caveats

  • The announced savings are a forward-looking estimate, not a measured result.
  • The gains cited are the vendor's, with no published method.
  • This is the case of a product vendor, not a plant building its own solution.
Case study published by Anthropic Related i2Business services:ERP & CRM IntegrationRAG & Knowledge Bases
Discuss a similar case

A pilot that never makes it to production? Tell us about your situation: we will identify what is blocking you and what can be carried over from these cases.

Talk to a consultant

Method

How we approach an industry

The same approach whatever the business: understand before tooling, prove before scaling.

1Business scopingField interviews and mapping of the most time-consuming processes.At kickoff
2Use case selectionPrioritization on three criteria: measurable gain, technical feasibility, team adoption.After scoping
3Real-world pilotOne use case deployed on a limited scope, with indicators defined upfront.Duration set during scoping
4Scale-up & trainingScaling up, data governance and upskilling of internal teams.Ongoing

References

Trusted by clients across every industry

GIZ
UAS
Monoprix
EY
Sidi El Heni
Cellcom
Hamila Duty Free

Let's talk about your business

A one-hour conversation is enough to identify two or three use cases you can activate in your organization, with an estimate of effort and gain.

Request an assessment