RECO
    Back to blog
    AI Stack

    Which AI Tools Should SMEs Actually Use in the Second Half of 2026?

    Markus C. Weiss, BScMarkus C. Weiss, BScCTO, RECOPublished Last edited 6 min read

    When a management team asks me which AI tool to buy, I answer with a question of my own: where does your data go, and how do you get it back out? For typical office work, the models from the big providers are, in my view, now close to each other. For an SME, the differences lie in the architecture around them.

    So my thesis for a service-oriented SME is this: five or six AI tools are enough. Each additional tool mainly adds another place where customer data sits, without anyone keeping track. The failed AI project addresses the gap between existing files and actual company knowledge. A tool for knowledge preservation should keep customer context retrievable and exportable, with knowledge validation clarifying which captured information has actually been checked.

    The opposite view is common, and it has a point. Let every department test freely and the best tool will win. Testing is right, as long as it happens with non-critical data. The trouble starts when, a few months later, quotes, HR notes and customer correspondence are spread across many tools, some of which nobody ever approved. That is shadow IT, only grown faster.

    The data check before any tool

    From a technical point of view, I check every tool against these five points first, before looking at features.

    • Where are inputs processed and stored, in which country and with which sub-processors?
    • Are your inputs used for training by default, and can you exclude that contractually?
    • Is there a data processing agreement under Art. 28 GDPR?
    • Can the tool connect to your user directory via single sign-on, so access ends when someone leaves?
    • In what format do you get your data back if you cancel?

    My rule: as long as any of these questions is open, the tool stays in a test with non-critical data.

    In my view, AI pays off in an SME today mainly in six categories.

    • General assistance and research
    • Writing, translation and editing
    • Knowledge documentation and conversation intelligence
    • Sales and customer success
    • Finance, controlling and reporting
    • Development and automation

    General-purpose assistant: choose by stack, pay by usage

    ChatGPT Business, Claude in its Team or Enterprise edition, or Microsoft Copilot in Microsoft 365 are the obvious candidates. I advise deciding based on your existing software stack. If you already work deep inside Microsoft 365, Copilot has the advantage that documents and permissions already sit where the assistant works. If you do a lot of research and analysis, ChatGPT or Claude give you more flexibility.

    In my assessment, licensing policy drives the cost at least as much as the choice of vendor. Do the calculation with your own quoted price. Unused licence cost per year is the price per person per month, times twelve, times the difference between licences paid for and people who actually use the tool regularly. Decide beforehand what regularly means, for example at least once a week. The second figure comes from the usage reports in the admin console, not from a survey. Check before you buy whether the vendor provides such reports. My recommendation: after the first few weeks, reallocate licences based on those reports and hand them out again on request.

    Writing and translation

    For translations with fixed terminology, such as contracts or product documentation, I rely on a specialised tool like DeepL with a stored glossary. A general-purpose assistant translates well but sticks to your terms less reliably.

    Knowledge documentation and conversation intelligence

    This is where I see the biggest gap in the SME stack. For assistance and translation there are mature standard tools. What is missing is a layer that records what gets decided in meetings, handovers and employee reviews, and makes that knowledge searchable later.

    That is what we build RECO for, and since it is our own product, read this paragraph with appropriate scepticism. Technically, three things matter to us. Processing runs on EU servers with a dedicated server in Germany, and there is an on-premises option for companies with stricter requirements. Speech recognition runs on Speechmatics, whose startup programme RECO is part of. Spoken German with dialects and technical terms is a hurdle of its own in the DACH region; our post on speech-to-text models and dialect explains why. And every summary stays a proposal that a person reviews and approves.

    Sales, customer success and controlling

    The same rule applies in all three areas: use AI where the data already lives. The AI features in HubSpot or Salesforce help with lead qualification, email drafts and sanity-checking forecasts. In customer success, the same goes for your existing helpdesk. In controlling, AI-supported analytics in Power BI, Tableau or the ERP are only as good as the data underneath. If you start from scattered Excel files, you just move the problem into the AI.

    Development and automation

    For teams that build their own software, coding assistants like GitHub Copilot or Cursor are now widespread. For workflow automation, n8n, Make or Zapier remain the pragmatic options. One technical note: n8n can run on your own infrastructure. That matters as soon as automations move personal data between systems.

    A realistic stack for a service SME

    1. A general-purpose assistant as the base, usually Microsoft Copilot or ChatGPT Business.
    2. DeepL for multilingual communication.
    3. RECO for knowledge documentation, employee reviews and handovers.
    4. AI features in the existing CRM for sales and customer success.
    5. AI-supported analytics in the existing BI or ERP tool.
    6. Optionally, a coding assistant for development teams.

    The list is short on purpose. More tools mean more context switching, more contracts and more places where data sits.

    What has changed since the spring

    Since 2 August 2026, the transparency obligations in Art. 50 of the EU AI Act apply. If you run a chatbot on your website, for instance, check that visitors can clearly tell they are chatting with an AI. The Digital Omnibus (Regulation (EU) 2026/1744) also rewrote Art. 4. Companies must take measures to support the AI literacy of their staff, but they do not have to guarantee a specific level for each individual. For tool selection, that means planning a short introduction for every tool you roll out and recording who received it. Our post on the EU AI Act for SMEs has the details. This is a technical assessment, not legal advice.

    Five questions before you sign

    • Which concrete decision do we make better with this tool?
    • Which hour does it save us per week, and how do we measure that?
    • Who in the company owns the tool, professionally and contractually?
    • How does the tool handle data protection, the works council and auditability?
    • What happens to our data if we switch it off again?

    I consider the last question the most important, because it shows whether you are introducing a tool or a dependency.

    Sources

    Related articles

    In a 30 minute software demo we show you how RECO secures and transfers knowledge in your company. Request a demo now and we'll get back to you with a proposed time.