Most mid-size companies run on a mix of systems built at different times: an ERP that has been in place for over a decade, a custom order or billing tool, and a set of spreadsheets that fill the gaps between them. These systems still handle daily operations reliably, so replacing them is rarely a simple decision.
At the same time, pressure to adopt AI keeps growing. Leadership teams expect faster customer service, more accurate forecasting and less manual work, while IT departments have to deliver those results without disrupting the systems the business depends on. In most cases, a full rebuild is not the first step.
Existing CRMs, help desks and ecommerce platforms can be connected to models from OpenAI, Google or Mistral while core systems stay in place. The AI transformation firms in this article follow a phased approach, modernizing one system at a time, and each of them is small enough that the specialists who scope a project also take part in delivering it.
To keep things concrete, this article follows one example throughout: a UK builders’ merchant running an AS/400 system for orders and stock, a web shop, and a handful of spreadsheets that hold everything together.
Why the Big Rewrite Keeps Failing
Anyone who has sat through a full core-system replacement knows the pattern. The budget grows. The go-live date moves twice. New features freeze for a year because nobody wants to build on a system that’s about to be switched off.
The bigger risk is the rules nobody wrote down. The AS/400 at the builders’ merchant almost certainly holds trade discounts agreed on a handshake in 2003, a special VAT treatment for one customer group, and a fix for a supplier that went under years ago. A rewrite team working from a requirements document won’t know any of that exists, and customers find out first.
Engineers often point to Martin Fowler’s strangler fig pattern instead. The name comes from a tropical fig that grows around a host tree and eventually replaces it. New services are built beside the old system, traffic moves across in small pieces, and the old core is retired once nothing depends on it anymore.
AI fits into this without much fuss. A chatbot for trade customers can read stock levels straight from the AS/400. A demand forecast can run on a nightly export. Both go live without touching the core.
Some tools help with the migration too. IBM’s watsonx Code Assistant for Z translates COBOL into Java, Amazon Q Developer handles Java version upgrades and GitHub Copilot is good at writing tests for code that has none. They make engineers faster. They don’t tell anyone which module to move first, and at the builders’ merchant that decision will cause more arguments than any line of code.
Small Teams, Fewer Handoffs
Big consultancies are good at programs that run across dozens of countries and hundreds of staff. A three-person IT team with one ERP and one web shop gets little out of that machinery and pays for it anyway.
At a smaller firm, the architect who scoped the job in March is often still reviewing code in June. Questions get answered by someone who has actually seen the system. The limit is headcount: a team of 40 can’t run ten parallel workstreams, and some will bring in a partner for niche skills like mainframe work. For a phased project that changes one or two components a quarter, that limit hardly matters.
The Shortlist
Each firm below has public evidence of legacy system modernization or AI integration work and AI running in production. The order is not a ranking.
| Firm | Main focus | Best suited for |
|---|---|---|
| Alpha Craft AI | AI consulting, agents, workflow automation | Small firms starting their first AI project |
| DigitalSuits | AI added to current stores, CRMs, and workflows | Ecommerce and mid-size businesses |
| Leobit | .NET modernization and AI features | Software firms with aging .NET products |
| Addepto | AI and data engineering | Manufacturing, logistics, retail data |
| Vstorm | AI agents and LLM software | Mid-market teams automating workflows |
| DLabs.AI | Custom ML added to existing apps | Product teams adding AI to live apps |
| Neurons Lab | AI for financial services on AWS | Banks, insurers, wealth managers |
Seven Firms Doing This Work
The profiles differ in length because these firms publish very different amounts about their clients.
1. Alpha Craft AI
Alpha Craft AI works out of Glasgow and keeps its offer short: custom AI development, AI consulting and strategy, AI agent development, and automation connected to the workflows a company already has. Its blog is a fair guide to the kind of client it has in mind. Recent posts compare AI tools for small business customer communication, look at Power BI agents for reporting, and explain the network setup automation relies on.
That mix suits a company that hasn’t shipped any AI yet and isn’t sure where to begin. A consulting stage can cut a long wish list down to one process, and an agent is then built to work inside tools staff already use. For UK firms around the size of the builders’ merchant, a nearby partner that starts with strategy can save months of guessing.
Website: https://alphacraftai.com/
2. DigitalSuits
Registered in Estonia, DigitalSuits started out as a Shopify Plus and headless commerce agency, and AI has grown into a large share of its work. Projects include support agents and chatbots, semantic search, OCR-based document processing and speech-to-text built on OpenAI’s Whisper, plus AI connected to Salesforce and HubSpot. The team plugs models from OpenAI, Anthropic, Google, AWS or Azure into a client’s existing product instead of building a new one. Its AI work also covers recommendation engines, virtual try-on for fashion and beauty stores, and translation pipelines, and every system it ships is monitored after launch for accuracy and model drift.
For a business like the builders’ merchant, the agency’s habit of starting at the storefront makes sense. AI product search can go onto the existing web shop first. Catalog automation follows. A headless rebuild comes later if the numbers justify it, and the back office stays untouched until there’s a reason to change it.
Website: https://digitalsuits.co/
3. Leobit
Leobit was founded in 2014 and has its main engineering center in Lviv, with smaller offices in Austin, Tallinn, Kraków and London. Its 170 or so engineers mostly work in .NET, web and AI.
Legacy work comes up again and again in its Clutch reviews: a rebuilt CRM for a company that makes nonprofit software, maintenance on an ageing loan system, a modernization assessment one client singled out for its detail. Leobit has also run a webinar on whether a legacy system should be modernized or rebuilt from scratch. Software vendors sitting on an old .NET product, with AI features now on the roadmap, are the most likely to get value here.
Website: https://leobit.com/
4. Addepto
Addepto was set up in Warsaw in 2018 by two data scientists, Artur Haponik and Edwin Lisowski, and employs between 50 and 100 people. Its first projects were small machine learning models that improved how clients’ existing ERP systems ran. No new platforms.
The work has since grown into demand forecasting for a parcel delivery company, product traceability and generative AI. For older setups, the firm’s data engineering is arguably the bigger draw. Stock data spread across an AS/400, a web shop and three spreadsheets has to be cleaned and joined before any forecast built on it can be trusted, and Addepto puts much of its effort there.
Website: https://addepto.com/
5. Vstorm
Vstorm is based in Wrocław, was founded in 2017 and has somewhere between 11 and 50 staff. It calls itself a boutique AI agent engineering consultancy. The team builds LLM software, RAG systems and agents on LangChain, LlamaIndex and Pydantic AI, is an official Pydantic implementation partner, and appeared in Deloitte’s Technology Fast 50 for Central Europe.
Its project with Xchanger, a platform connecting payment institutions, ran for about two years. Vstorm reworked the platform’s architecture first and added generative AI for customer experience and internal operations afterwards. Mid-market companies that want agents handling real daily tasks will recognize the approach.
Website: https://vstorm.co/
6. DLabs.AI
DLabs.AI was founded in 2015 by Przemysław Majewski, Maciej Karpicz and Ewa Rajewska, originally to build AI software for people managing diabetes. The Polish company has since taken on machine learning and automation projects in many other sectors.
A good part of its portfolio involves adding AI to products that are already live. An education platform, for instance, hired DLabs.AI to add machine learning features to its web and mobile apps while they stayed online. Product teams with an established user base and no wish to rewrite will find that track record relevant.
Website: https://dlabs.ai/
7. Neurons Lab
Neurons Lab was founded in 2019 and has offices in London and Singapore, with a team of roughly 50 to 200. It works only in financial services, covering banks, insurers and wealth managers, and holds AWS Advanced Partner status with a generative AI competency.
Banks can’t take their core systems offline for a weekend, and every release passes through compliance, so phased change is the only realistic option. Neurons Lab’s projects (KYC and AML automation, document agents, copilots for relationship managers) run beside the core banking platform rather than replacing it. Regulated firms with older cores are its entire market.
Website: https://neurons-lab.com/
Questions Worth Asking a Partner
Start by asking which workflow they would change first. A serious partner names one, for instance trade customers’ stock queries or supplier invoices, and explains how the result will be measured. “We’ll review the whole estate” is not an answer yet.
Then ask how long the old and new systems will run side by side. A few months of overlap is normal. It also costs money, and it’s missing from a surprising number of budgets, along with data cleanup and staff training. Cleanup takes longer than planned; training gets squeezed into the last week.
With a small firm, ask what happens if the lead engineer leaves halfway through. Who replaces them, where is the documentation kept, who owns the code? Those answers belong in the contract.
The First Few Months
At the builders’ merchant, a realistic plan would start with three or four weeks of discovery. The partner maps the AS/400 data, finds out which spreadsheets the business actually depends on, and agrees on one painful workflow, say trade customers phoning branches to check stock.
The first AI feature, perhaps an assistant that answers those stock questions from a nightly export, connects without changing the core. It runs alongside the old phone-and-check routine for a month or two, and the two get compared. If branch staff get their afternoons back, the next phase is an easy sell. If not, the business has lost a few weeks rather than a year.
Summing It Up
Most mid-size companies will run legacy systems and AI side by side for years, and the builders’ merchant is no exception. A phased plan lets useful features reach customers now while old components are retired when it’s safe. Smaller partners fit that rhythm because the people who design the plan stay to deliver it.
The simplest next step is to write down the workflow that frustrates staff the most and send it to two or three of the AI transformation firms above. How each one proposes to break it into stages will say more than any capabilities deck.
