cm-opti - AI implementation and process automation
Intro call

Services

AI solutions for business: AI agents, document capture, computer vision, forecasting and systems integration

How they work, what we need from you and what results they delivered in our team’s projects.

AI process automation

Solutions that take over repetitive work with emails, documents and images. We build them with large language models, AI agents and computer vision. Where the model is uncertain, a person makes the call.

AI assistant for company knowledge

Answers questions based on your documents, procedures and regulations, and cites the source.

How it works and what we need

How it works

The assistant uses RAG (Retrieval-Augmented Generation). It first finds the relevant passage in your documents and then builds its answer on it, so every answer can be checked against the source. It can also suggest the next step, such as what else to check.

What we need from you

The documents the assistant should use, such as contracts, manuals, procedures or regulations, and a list of the questions your team asks today.

From our projects

At a construction company operating in Germany, an answer to a question about the HOAI and VOB regulations now takes 10 minutes instead of about 1.5 hours, including the justification.

Analysis of inquiries and quotes

Reads the customer’s requirements, compares them with previous quotes and production data, and prepares a draft quote for approval.

How it works and what we need

How it works

The inquiry arrives by email or through a form. The assistant extracts the parameters, such as material, dimensions, quantity and deadline, and finds similar past orders. It calculates the price using your rules and points out what is missing for a complete quote. A salesperson or process engineer reviews the result and decides whether to send it.

What we need from you

Sample inquiries together with the quotes prepared for them, and the rules you use to price today, such as a costing spreadsheet.

AI agent for multi-step tasks

Takes a case through several systems: gathers information, decides what to do next and prepares the result for approval.

How it works and what we need

How it works

The agent connects an LLM (Large Language Model) with tools such as an email inbox, ERP, CRM or document repository. It decides the next steps itself, but within limits we set together: which data it can access, what it can change and what needs approval from your team. Every decision is logged, so it is easy to trace how it was made.

What we need from you

A description of a task that currently passes through several people or programs, examples of how it runs and access to the systems the agent will use.

Visual quality inspection (computer vision)

Analyzes photos and videos of products or goods and flags defects, damage or mismatches with system inventory.

How it works and what we need

How it works

We train the model on photos or videos from your process, including defective items. It analyzes video frame by frame. It assesses every item against the same criteria and compares the result with a reference or with the data in your system. Your team only reviews the cases the model has flagged. We check its accuracy in the first sprint, using images from your line, before the solution covers your entire production.

What we need from you

Photos or videos of the area you want to inspect, including examples of defects and damage, and access to the data the images should match.

From our team’s experience

In a retail chain, the model compared shelf photos with system inventory. Discrepancies became visible before a shelf ran empty or an order was calculated from the wrong stock level.

AI document capture (OCR)

Turns scans, PDFs and photos of invoices, orders or contracts into data in your system, with no manual retyping.

How it works and what we need

How it works

OCR (Optical Character Recognition) recognizes the text on the scan, and a second step identifies each field: order number, date, amount, address. The data goes into the system ready for approval. Only documents that do not follow a known layout are handled manually.

What we need from you

Examples of each document type you want to process, and a list of the fields that should go into the system.

From our projects

At a transport company, shipping orders go into the TMS (Transportation Management System) without retyping, and staff only approve them.

AI assistant for the email inbox

Reads incoming messages and attachments, recognizes what each case is about and routes the data to the right system or department.

How it works and what we need

How it works

A message arrives in a shared inbox, for example for orders, complaints or claims correspondence. An NLP (Natural Language Processing) model classifies it by topic, and OCR reads the attachments. The data goes straight to the right person or system.

What we need from you

Access to the inbox or a copy of it, and sample messages sorted into the categories your team uses.

From our projects

At an insurance company, the model identifies the risk category in 90% of cases. About 25% of documents no longer go into the main queue, and the correspondence backlog fell by 30%.

Machine learning and forecasting

Machine learning models trained on information from your systems. They forecast demand and failures, recommend products and show the effects of changes before you make them. They rely on well-structured data and reports, and we retrain the models as new data comes in.

Demand forecasting

Forecasts sales of every item and builds a replenishment plan that accounts for transport costs and warehouse capacity.

How it works and what we need

How it works

We train a machine learning model on sales history, seasonality, trends and external factors that affect demand. From this, it calculates how much stock to send to each warehouse and when. The planner sees several scenarios and compares them before deciding.

What we need from you

Item-level sales history and data on deliveries, warehouses and transport costs.

From our team’s experience

At a distribution company operating in several markets, the forecast covered individual items instead of product-group averages, and each market got its own seasonality. Less stock sat in the warehouse, and the replenishment plan matched actual demand.

Predictive maintenance: machine failure prediction

Predictive maintenance: machine data shows which machine needs servicing before the line stops.

How it works and what we need

How it works

We train the model on data the machines already record, such as temperature, vibration or power draw, and on the history of past failures. When current readings start to resemble the pattern before one of those failures, the maintenance team gets an alert, by email or in the CMMS (Computerized Maintenance Management System), and schedules the repair at a convenient time.

What we need from you

Access to data from PLCs (Programmable Logic Controllers) or a SCADA system (Supervisory Control and Data Acquisition), plus the history of faults and servicing. In the diagnosis, we check how much history is enough. If there is no such history yet, we start with downtime tracking.

Recommendation systems

Use purchase history and customer behavior to suggest further products or items frequently bought together.

How it works and what we need

How it works

Market basket analysis shows which products customers buy together and how often. When transactions can be linked to individual customers, a machine learning model compares similar purchase histories and suggests what the next buyer may be interested in. The recommendations feed sales, the website or inventory planning.

What we need from you

Receipt history from brick-and-mortar stores or online orders. The longer the period, the more reliable the patterns.

From our team’s experience

For a retail chain, we analyzed hundreds of thousands of receipts. The same model suggested cross-selling opportunities, helped plan inventory and powered the website’s recommendations.

Process digital twin

A virtual, data-based copy of a process where you test the effects of a change before it reaches the shop floor.

How it works and what we need

How it works

We model machines, flows, cycle times and constraints, using data from your systems. On this copy you test a change, such as a different order sequence, and see its impact on throughput and costs with no risk to production.

What we need from you

Data on how the process runs, such as times, volumes and downtime, and the knowledge of the people who manage it.

From our projects

We built a prototype of such a model for a large industrial company.

Downtime tracking and OEE analysis

Operators log the cause of every stoppage in a simple app, and you see what drags down OEE (Overall Equipment Effectiveness) the most.

How it works and what we need

How it works

The app runs on a tablet or phone at the machine. The operator picks the cause from a short list agreed with the production manager. The downtime duration is recorded automatically or pulled from the controller. The report shows availability, performance and quality losses for every machine and shift. The history you collect also lays the groundwork for predicting failures later.

What we need from you

A list of machines and downtime causes, if one exists, and conversations with the operators and the production manager.

Reports and dashboards (BI)

KPIs from many systems in one place, updated automatically.

How it works and what we need

How it works

We start by asking which decisions the BI (Business Intelligence) report should support and by choosing the KPIs (Key Performance Indicators). We define how each metric is calculated and automate collecting and cleaning the data in an ETL process (Extract, Transform, Load). The dashboard shows where your company stands on a single screen.

What we need from you

Access to the systems the numbers come from, such as your ERP, CRM or spreadsheets, and your current reports, if there are any.

Systems integration

Your company’s software exchanges data on its own, with no manual retyping. We design the architecture so further systems can be connected without rebuilding everything.

ERP, CRM and online store integration

You enter an order, customer data or a stock level in one program, and the others have it right away.

How it works and what we need

How it works

We look at which programs you use and how data flows between them today, including data that is retyped by hand. Then we connect them through an API (Application Programming Interface), a defined way for applications to exchange information. If a system has no API, we use its database or a file export, and if that is not enough, we write a connector that moves the data. An order status changed in the ERP updates in the CRM and the online store.

What we need from you

A list of the systems you use and access to their documentation or a contact at the software vendor.

Transportation management system (TMS) integration

Data from orders, emails and scans flows automatically into the TMS and from there to your other company systems.

How it works and what we need

How it works

Document capture prepares the data for the TMS, and the connections to your other systems pass it on. Instead of retyping, staff review and approve. A next step could be automatic data exchange with customs and veterinary systems.

What we need from you

Sample orders and shipping documents, plus access to the TMS and the systems it should connect to.

From our projects

At a transport and freight forwarding company, standard documents go through automatically, and only exceptions are handled manually. Processing a single document takes seconds instead of minutes.

Deployment on your servers or in the cloud

The solution runs on your infrastructure or with a cloud provider, in compliance with the GDPR.

How it works and what we need

How it works

We agree on where it runs at the start, together with the security requirements. Where regulations require it, we build an isolated environment the data never leaves, and every processing step can be demonstrated in an audit. Code and data stay on your side.

What we need from you

Information about your servers and the requirements of your IT or security team.

From our projects

At an insurance company, the model worked on anonymized data in an isolated environment built for that project.

Industries and technologies

Industries
Insurance, construction, distribution, transport and logistics, retail and e-commerce, manufacturing.
Technologies and methods
Natural language processing (NLP), RAG (Retrieval-Augmented Generation), AI agents, computer vision, machine learning, sales forecasting, replenishment and inventory planning, recommendation systems, market basket analysis, text clustering, systems integration through APIs, databases or custom-built connectors.

How we work

We work with Scrum, in sprints of 2-3 weeks each. After every sprint, another part of the solution is running at your company. We work mostly remotely and come on site when a project stage requires it. Where and how we work

  1. Step 1

    Intro call and diagnosis

    We start with a free, 60-minute online intro call, after which you get our conclusions in writing. If the goal is clear, that is enough to prepare a proposal. For a broader scope, we propose a paid diagnosis: we talk to the people who run the process, map it, check the data in your systems and calculate how much the company is losing. We agree on the duration and price of the diagnosis before we start, and if you order the implementation, we deduct it from the first sprint. The result is a task list ranked by value (product backlog) and a proposal, and the documents stay with you.

  2. Step 2

    Sprint

    At the start, we pick the tasks from the list that deliver the most value. We turn them into a working part of the solution and put it into operation at your company. A person on your team who knows the process checks the results as we go, for example whether the assistant answers correctly, and we fine-tune the solution based on their feedback. You pay for each sprint separately.

  3. Step 3

    Review and decision

    At the end of the sprint, we show what we built, and your team tests it on its own data. Feedback goes onto the task list, and together we set the scope of the next sprint: fixes, another part of the process or a connection to another system. You decide whether we start it. We repeat steps 2 and 3 until the solution covers the full planned scope.

  4. Step 4

    Handover and support

    When we finish, whether after the last planned sprint or an earlier one, we train the users and hand over the documentation and code. If you want, we stay on: we keep the solution running, fix bugs, add features and retrain the models on current data.

Questions about solutions and implementation

You will find more answers on the home page.

What is the difference between computer vision and OCR?

OCR reads text from images and documents. Computer vision analyzes the entire content of an image: it detects and recognizes objects, their position and their condition. OCR can be one part of such a solution.

What is an AI agent, and when is it used?

An AI agent is a system that completes a multi-step task on its own: it plans, uses tools and makes decisions within set limits. We use it where rule-based automation is not enough, for example when an order has to be read from an email, checked in several systems and answered with a reply ready for approval. If a simpler solution will do, we choose it, because it is cheaper to run.

Which solution should we start with?

With the process that weighs most on your team and has a measurable cost, such as retyping orders by hand, visual inspection of products or compiling reports from several systems. In the diagnosis, we calculate what this process costs today and how much the solution could save. The first sprint covers a selected part of it, so you see results after 2-3 weeks.

Can an AI solution be connected to our ERP?

Yes. We connect it to your ERP, CRM, TMS or online store through an API (Application Programming Interface), the database or a file export. If the system offers none of these, we write a connector that moves the data.

How much data does an AI solution need?

It depends on the solution. An assistant for company knowledge uses the documents you already have. A demand forecast needs sales history covering at least one full season. A vision model learns from images of your line, including defective items. In the diagnosis, we check how much data is enough in your case.

Free intro call

Describe your process in a few sentences. On a 60-minute online call, we will check which of these solutions makes sense for you. If none does, we will tell you so, and you get our conclusions in writing.

Go to the form