Service · App development

Building an IBM Watson app.

A custom app on the IBM Watson and watsonx platform, for enterprises with an existing IBM stack, banks, insurers, healthcare and government bodies that want to use AI without losing control of data and compliance. We build on watsonx.ai, watsonx.data, watsonx.governance, Watson Assistant, Discovery, NLU and Speech-to-Text where it fits, and we are honest when another route works better. No vendor pitch: Watson is one of the options alongside Anthropic, OpenAI, Mistral or a local model.

watsonx.aiwatsonx.governanceWatson AssistantAI Act & GDPREU residency

When IBM Watson is the right choice, and when it isn't.

IBM Watson has been quietly reinvented in recent years. The original Watson from the Jeopardy era has largely been retired, and what remains under the Watson banner is a collection of enterprise AI building blocks that works exceptionally well for one specific type of organisation: institutions whose existing systems run on IBM (DB2, MQ, mainframe COBOL, z/OS, Power Systems, Db2 Warehouse), where the procurement relationship with IBM is a given, and where audit, governance and data residency matter more than the latest benchmark score.

The new centre of gravity is called watsonx, and it consists of three connected products: watsonx.ai (foundation models and deployment of your own models), watsonx.data (an open lakehouse built on Iceberg and Presto) and watsonx.governance (model oversight and AI Act documentation). Alongside these, the older but still widely used components — Watson Assistant, Watson Discovery, Watson NLU and Watson Speech-to-Text — run in the same IBM Cloud environment. We build custom apps that call these components from your own front end and your own domain logic. Our broader approach to this kind of work is described on the page about AI development, with the more general LLM layer covered under custom LLM integrations.

What sets us apart: we are vendor-independent. We are not an IBM partner with a sales quota, and we have no commercial incentive to choose Watson when another route fits better. For a retail start-up with a fresh cloud budget and no IBM history, watsonx is rarely the right choice; for a Dutch bank with a mainframe in Diemen, it is often exactly the right answer. We can talk through how that trade-off works out in your case in a vendor advice session before we discuss building anything.

Three typical forms.

The Watson projects we encounter in practice usually fall into three patterns. Which one suits you depends on where the value lies: in conversational interfaces for customers or employees, in document AI and search over your own corpus, or in a deeper custom app where watsonx.ai acts as the brain for your own domain.

Conversational · Watson Assistant · customer contact

Customer support and employee assistant

An app or chat layer that handles customer questions or internal employee questions based on a controlled corpus. We use Watson Assistant for dialogue management where it fits, or set up your own orchestrator on watsonx.ai with your own prompts, retrieval and guardrails, depending on how much control you want over the conversation. For banks and insurers, this often includes a handover mechanism to human agents and logging suitable for compliance investigations. We connect to an existing contact centre (Genesys, NICE, Avaya), CRM (Salesforce, Dynamics, Siebel) or telephony stack.

Watson Assistantwatsonx.aiRAG on your own corpusAudit logging
Document AI · Discovery · NLU

Document AI and enterprise search

An app that reads, classifies, summarises and makes searchable large streams of documents — contracts, policies, case files, regulations, knowledge bases, process documentation. Watson Discovery becomes the retriever layer on top of your own corpus; watsonx.ai generates the answers, summaries or structure. For regulated industries, you can keep the corpus within IBM Cloud Frankfurt, with granular access rules per role or file type. We also build integrations with existing document management systems (FileNet, OpenText, SharePoint, NetDocuments) and mainframe sources.

Watson DiscoveryWatson NLURAGFileNet · OpenText
Custom app · watsonx.ai · domain-specific

Custom AI app on watsonx

Your own app with its own domain, its own workflow and watsonx as the engine, for example a fraud investigation tool for an insurer, an underwriting assistant for a bank or a triage app for a hospital. We build the front end natively or as a web app, handle the integration with the mainframe and data warehouse, and use watsonx.ai for the reasoning. watsonx.governance keeps track of what the model does and how that relates to the AI Act. Our broader approach to this kind of project is described on enterprise AI implementation.

watsonx.ai foundation modelswatsonx.governanceDB2 · MQ · mainframeOn-prem option

What you get at the end.

A production-ready custom app that leans on the Watson and watsonx building blocks where that makes sense, runs in your IBM Cloud environment or on-prem watsonx installation, and fits into your existing architecture. No lock-in to us as an agency: you own the code, prompts, models and infrastructure definitions.

  • Native or cross-platform appA mobile app (iOS, Android, Flutter, React Native) or web app, built on the Watson and watsonx layer, with a UI that matches your brand and the work your end user does.
  • Your own orchestrator on top of watsonx.aiYour own prompt, retrieval and guardrail layer around watsonx.ai or Watson Assistant — no hard-coded vendor dialogues, just manageable flows you can adjust yourself.
  • RAG pipeline on your own corpusDocument ingestion, embedding and retrieval — via Watson Discovery, watsonx.data or an open-source vector store — so answers are grounded in your own knowledge rather than generic web data.
  • Integrations with your IBM and non-IBM stackConnections to DB2, IBM MQ, z/OS, Power Systems, mainframe COBOL, FileNet, OpenText and, where needed, non-IBM systems (Salesforce, Dynamics, SAP, ServiceNow). A Watson app rarely lives on its own.
  • EU residency and hybrid deployment optionsIBM Cloud Frankfurt for most clients, watsonx on Cloud Pak for Data on-premises or on a private cloud where requirements are stricter — for example in government or at banks with strict IT policies.
  • watsonx.governance integrationModel oversight, model version records and logging in a way that meets AI Act requirements — no separate governance spreadsheet alongside the app, but integrated into the production pipeline.
  • Evaluation and regression suiteA set of tests on realistic examples from your domain, which we run against the model with every release. A Watson app that quietly gets worse on edge cases after an update must not go unnoticed.
  • Compliance evidence: DPIA, AI Act, NIS2, DORAThe documentation and logging you need for your data protection officer, and for compliance, audit and DNB (Dutch central bank) processes — not a loose PDF at the end, but part of the architecture.
  • Fully transferable codebaseThe source code, infrastructure as code (Terraform or OpenShift manifests), prompts and evaluation set are yours. If you later want to continue with another agency or take it over in-house, you can, without us withholding any components.
  • Maintenance contract (optional)Monitoring, model evaluation, security patches and further development — only if you want us to handle that. Many clients only have us do the build and then maintain it in-house.

Who we build Watson apps for.

Eight patterns we see again and again in our conversations about IBM Watson and watsonx. If you recognise your organisation or use case in one of them, we're happy to talk further — and we'll also be honest when another platform or a hybrid approach suits you better than pure Watson.

Banks

Underwriting and KYC support

An app that helps bank staff with credit underwriting, KYC investigations or fraud analysis, with watsonx.ai as the reasoning layer and existing core banking data from DB2 and the mainframe as raw material. For clients under DNB and AFM supervision, audit logging and explainability carry great weight — watsonx.governance helps build those requirements in.

Insurers

Claims triage and policy AI

An app that classifies incoming claims, detects fraud indicators, explains policy terms and proposes settling standard cases, while the human claims handler stays in control. Often built with Watson Discovery on the policy archive and watsonx.ai for the reasoning, connected to Guidewire or a proprietary core insurance system.

Healthcare

Clinical decision support and triage

Apps that turn clinical data into decision support for doctors or triage staff — for example preliminary assessment of outpatient referrals or administrative assistance with discharge documentation. In healthcare, NEN 7510 applies from day one; we deliver suitable integrations with HiX, Epic, ChipSoft and CGM, and keep audio and patient data within the EU.

Government

Case file AI for implementing organisations

Implementing organisations (the Tax Authority, UWV, SVB, municipalities, ministries, regulators) with large case files and dense regulation want to deploy AI without giving up control over data and explainability. Watson on IBM Cloud Frankfurt or on-premises Cloud Pak then fits within existing procurement, security and architecture frameworks.

Logistics

Supply chain assistant on an SAP/IBM stack

Logistics companies with a hybrid SAP/IBM architecture (DB2, MQ, mainframe batch, alongside SAP S/4HANA) are a typical profile where Watson fits well — not because it's sexy, but because the data already lives there and the IT organisation knows how to work with it. We build apps for purchasing, planning or inventory AI that run on this type of stack.

Energy & utilities

Field service app and asset AI

Energy and grid companies that want to build predictive maintenance, asset AI or field service apps on an existing IBM Maximo and DB2 environment. Watson Speech-to-Text for hands-free reporting in the field, watsonx.ai for reasoning about faults or asset condition. Often in line with the NIS2 requirements that are tightening for the energy and water sectors.

Regulators & compliance

Regulatory AI for supervision and audit

Regulators and in-house compliance teams who search dense regulation (Wwft, MiFID II, DORA, NIS2, AI Act) for passages, precedents and risk signals. Watson Discovery on top of your own corpus, watsonx.ai for summaries and explanations, watsonx.governance for traceability.

Existing IBM customers

A first AI step within an IBM relationship

Organisations with a large IBM installed base (DB2 Warehouse, Db2 on z, Cognos, OpenPages) that want to take their first serious AI step without bringing in a second cloud platform. For these customers, Watson is often the practical choice; we help with the first app, the evaluation and the question of whether a second platform will eventually be needed after all.

The Watson and watsonx platform unpacked.

A short overview of each building block, as we walk through them in a first session. Not everything on this list belongs in your app, so we choose together which components add value.

Foundation models · custom model deployment

watsonx.ai

The reasoning layer: a set of foundation models (IBM Granite, Llama, Mixtral and other open-source families) that you can use via a hosted API in IBM Cloud Frankfurt or run yourself in a Cloud Pak in your own data centre. For customers whose audio or source data must absolutely not leave the organisation, the on-prem route is decisive.

Lakehouse · Iceberg · Presto

watsonx.data

An open lakehouse architecture on top of Apache Iceberg and Presto, with a choice of query engines and integration with existing warehouses. For customers whose data is spread across Db2, Cloud Object Storage and existing analytics environments, this provides a single layer for the AI app to build on, without having to migrate everything first.

Model oversight · AI Act documentation

watsonx.governance

Model registration, monitoring for bias and drift, and the documentation you need for an AI Act classification as high-risk. For banks and healthcare organisations that have to account to DNB, DORA or NEN 7510, this is the building block that makes the difference between "we're doing something with AI" and "we can demonstrate what we're doing".

Conversational · dialogue management

Watson Assistant

The older but still widely used conversational platform with visual dialogue management, intents and entities. It works well for high-volume customer contact where you want tight control over flows, for example a claims bot at an insurer. For open-domain conversations, we more often choose a custom orchestrator on watsonx.ai.

Enterprise search · document AI

Watson Discovery

The search and document AI layer: ingestion of PDFs, Word, HTML and mainframe files, with semantic retrieval and entity extraction. It works well as a retriever beneath a watsonx.ai app for RAG, and has connectors that work with existing document stores.

NLU · classification · sentiment

Watson NLU & Speech-to-Text

Two older but useful building blocks: NLU for categorisation, sentiment and entity extraction from text, and Speech-to-Text for transcription within IBM Cloud. For specific speech recognition needs, we also look at other engines. See the speech recognition app development page for the broader comparison.

Not sure about a large project yet?

Test your idea first: a working prototype in 1 day

With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.

View OneDayBuild →

How a Watson project works.

1

Introduction & vendor comparison

A conversation in which we get to grips with your organisation, your existing IBM relationship and the use case. We are upfront about it: not every AI app deserves a Watson layer, and not every IBM customer should move everything onto Watson. Sometimes a focused advisory session is enough — see our page on digital consulting for the broader advisory route. If Watson is the right choice, we document why; if it isn't, we say so. A vendor assessment belongs before the first sprint, not after it.

2

Architecture and data residency

We map your existing IBM stack (Db2, MQ, mainframe COBOL, FileNet, Db2 Warehouse) and decide together where the AI work will sit: IBM Cloud Frankfurt for most customers, or watsonx on Cloud Pak for Data on-premises or on a private cloud for customers with stricter requirements — healthcare, defence, government, some banks. We capture data residency, retention, sub-processors and encryption strategy in a single architecture document that your security architect and data protection officer can review.

3

Compliance classification: GDPR, NIS2, DORA, AI Act

For banks, insurers, healthcare and government, a Watson app is by definition a compliance question. We carry out the AI Act classification (low risk, limited transparency or high risk), start the DPIA, and, for financial customers, translate the DORA and NIS2 requirements into the design. For healthcare customers, the architecture aligns with NEN 7510 as described on our page on NEN 7510-compliant software. watsonx.governance handles a large part of the logging requirements automatically, provided it is included in the design.

4

Building in sprints with an evaluation set

Every sprint delivers a working build on a staging environment, with a growing evaluation set containing examples from your domain. Every time we swap the model, adjust the prompt or add a new RAG source, we run the evaluation set and see immediately whether quality improves or worsens per cohort. For app development we use the usual tooling: Flutter, React Native or native iOS/Android for the front end, FastAPI or Spring for the back end, connected to watsonx endpoints and an IBM Cloud account.

5

Rollout, governance handover and maintenance

A phased rollout via a pilot group, with training for end users and a handover document for the model owner within your organisation. We link watsonx.governance dashboards to the right role, often a compliance officer or risk manager, so that oversight after go-live doesn't become administrative leftovers. Maintenance can run through us as an agency or through your internal team; we care about a clean handover and no artificial dependencies.

Compliance is not an afterthought.

For organisations that choose Watson, compliance is almost always one of the reasons. We take care of the legal, security and governance layer from the first sprint, not as a paperwork exercise at the end, and not as a separate consultancy track alongside the build.

AI Act and watsonx.governance

For high-risk applications, the AI Act requires documentation, monitoring, transparency and human oversight. watsonx.governance is designed for exactly that: model registration, evaluations, drift detection and risk classification in a single toolchain. We build that toolchain into the app, not alongside it.

GDPR & DPIA

For every use case involving personal data (customer contact, employee data, case files, claims) we start the DPIA and put data processing agreements in place with IBM and any sub-processors. Audio and sensitive text stay within the EU; we show exactly which processors see what.

NIS2 and DORA for the financial sector

Banks, insurers, critical infrastructure and large service providers fall under DORA and NIS2. We align the Watson architecture, supplier management and incident response with those frameworks, so the app isn't called back by risk and compliance at a later stage. In practice, that means contractual requirements for IBM as a third party, an exit strategy, and operational resilience testing.

EU residency: IBM Cloud Frankfurt

By default we choose IBM Cloud Frankfurt for EU data residency. For clients with stricter requirements — some banks, healthcare providers and government bodies — we run watsonx on Cloud Pak for Data on-premises or on your own private cloud. We document explicitly which route was chosen and why.

NEN 7510 for healthcare providers

For healthcare clients, we adapt the Watson architecture to the provider's NEN 7510 controls: access management, logging, encryption and supplier management. We align with the existing ISMS and with the working methods of the data protection officer and the information security function.

Audit trail & human oversight

Every interaction with the Watson or watsonx system leaves a traceable record: which model, which prompt, which retrievals, which output, which human check. For high-risk applications, we build in explicit human-in-the-loop steps before the output becomes operational.

FD
Fabian van Dijk Business developer · Appfront · fabian.vandijk@appfront.nl

Frequently asked questions.

What clients usually want to know before we start on a Watson or watsonx project.

Are you an IBM partner?
No. We build apps on the Watson and watsonx platform because it is the right choice for a specific type of client, but we are not an IBM partner with a sales quota. That is a deliberate choice: we want to be able to give vendor-independent advice. For clients with an existing IBM relationship, that is actually an advantage — we look at what works in your architecture and compliance context, not at what brings a vendor target closer. See also our page on vendor choice in AI projects.
When do you choose watsonx, and when Anthropic, OpenAI or Mistral?
In broad terms: watsonx wins when the existing IBM stack runs deep, when EU residency and explainability weigh more heavily than the very latest benchmark, and when a strong compliance framework (AI Act, DORA, NEN 7510) calls for a central governance platform. Anthropic, OpenAI or Mistral often wins when the organisation is cloud-native, has no IBM legacy, or when a specific model capability is decisive. In practice we often use a hybrid: watsonx for the regulated layer, and another model for parts of the pipeline where it fits better. Our reasoning is set out in more detail on custom LLM integrations.
Can we run watsonx on-premises?
Yes. watsonx is available in IBM Cloud (we usually use Frankfurt for EU residency) and as part of Cloud Pak for Data, which runs on-premises or on a private cloud — for example on OpenShift in your own data centre. For clients whose data absolutely must not leave the company — defence, some healthcare providers, some government bodies — the on-premises route is decisive. We advise per project which route fits, taking the broader data architecture into account.
How do you deal with the AI Act?
For every Watson app, we carry out an AI Act classification up front: low risk (general text, search or summarisation applications), limited transparency obligation (chat with customers) or high risk (medical decision support, credit assessment, HR selection, legal decision support, biometrics). For high-risk applications, we take care of the mandatory documentation, data quality requirements, monitoring and the human oversight mechanism. watsonx.governance is the practical engine here — it takes much of the logging and version administration off your hands, provided we set it up properly during the build.
How do you deal with DORA and NIS2 for financial clients?
DORA applies to banks, insurers and large financial service providers; NIS2 also affects a wider range of critical infrastructure players. We translate these frameworks into concrete requirements for the Watson architecture: contractual requirements for IBM as a third party, operational resilience testing, classification of the Watson environment as a critical ICT service, and alignment with the third-party register that is mandatory under DORA.
Can you integrate with our mainframe, DB2 and MQ?
Yes. For customers with a DB2 or mainframe layer, we work with IBM MQ as the messaging bus, with DB2 connectors, or via a data warehouse layer in watsonx.data. For COBOL or CICS work, we use integration services and API fronting. We don't touch the mainframe unnecessarily, but we make sure the Watson app communicates reliably with your existing systems.
Do you also work with Cloud Pak for Data?
Yes. Cloud Pak for Data is the broader on-prem or multi-cloud packaging that watsonx falls under. For customers running OpenShift in-house, it is often the natural landing place. We then run watsonx.ai, watsonx.data and watsonx.governance within your existing platform, with OpenShift deployment via GitOps. For customers who are just getting started, we more often choose IBM Cloud Frankfurt first and migrate later if scale demands it.
What if we already have a Watson Assistant implementation that isn't working well?
We hear this regularly: a Watson Assistant implementation that started well but became rigid or poorly maintained over time. We then carry out a short review: how does it perform on an evaluation set, and is a refactor within Watson Assistant sensible, or is moving to your own orchestrator on watsonx.ai the better route? We often refactor to a hybrid set-up with more LLM and fewer rigid intents.
Who owns the code, models and prompts?
You do. We deliver the full source code, the infrastructure as code (Terraform, Helm charts, OpenShift manifests), the prompts, the RAG configurations and the evaluation set. The watsonx resources run in your IBM Cloud account or on your Cloud Pak installation. A later handover to another agency or an internal team is a clean handover.
What determines the scope of a project like this?
The biggest scope drivers are the complexity of the domain, the choice between IBM Cloud hosting and on-prem Cloud Pak, the number of source systems to connect, and the compliance burden. A high-risk AI Act app requires considerably more documentation and logging than a low-risk search app. We work in sprints with clear sprint goals, so you can steer the scope sprint by sprint. An initial conversation quickly gives a realistic picture of the range.

Talk to us about your Watson app.

A no-obligation introductory call of half an hour. Tell us who will use the app, in which architecture and in what compliance context. We'll think along, give direction, and be honest about when watsonx is the right choice and when another platform or a hybrid approach fits better. Sometimes a short advisory phase is the right first step before we talk about building. You can also email us directly at fabian.vandijk@appfront.nl.

Edit content