On 1 October, IBM announced general availability of self-hosted deployment for IBM Bob, its agentic software development platform. Organisations can now run Bob in on-premises, private-cloud, sovereign-cloud and air-gapped environments, rather than only as a service. On the third-quarter results call on 21 October, IBM mentioned it as a way for clients in regulated and sovereign environments to keep code and models on their own infrastructure.
For a regulated organisation, that removes a familiar objection to AI coding tools. For a licensing team, it raises a different question. Once a tool moves inside your own estate, you are no longer buying one thing. You are buying several, and each has its own terms.
What IBM has announced
IBM describes Bob as a platform that helps teams move beyond code generation to apply AI across software delivery and modernisation. The self-hosted option is aimed at organisations that need greater control over data residency, security policies, governance and development workflows. With a supported self-hosted model configuration, code, development context and build artifacts can stay inside the customer-managed environment.
There are choices on where the model runs. Customers can run supported models on premises, including in air-gapped environments, using supported models they have licensed. Alternatively they can connect Bob to supported external model services in a hybrid configuration. IBM’s press release states that enterprises can run supported models they have licensed. SiliconANGLE reports that IBM says the self-hosted option keeps core Bob capabilities, including its BobShell environment, parallel tool calling, skills and operating modes.
The licensing stack has several layers
Running Bob yourself means paying for the product and also for what it needs around it. The first layer is the Bob entitlement itself. None of the announcements we reviewed publishes a price or a metric for it, so that is something to ask IBM directly.
The second layer is the model. IBM’s wording is that organisations run supported models they have licensed, which means the model licence is a separate agreement that you hold. Open-weight models in particular come with their own licence terms, and the permitted uses are not always the same as for open source software. The third layer is infrastructure. Self-hosted inference needs hardware with enough capacity, and someone to run it. The fourth is operations, because with a self-hosted model the customer installs and manages the model.
Questions to put to IBM
Ask what metric Bob is licensed on and whether it differs between the hosted and self-hosted options. Ask which models are supported, who licenses each one, and what happens to support if you use a model that is not on the list. Ask how updates, security patches and model changes reach an air-gapped environment, and who is responsible for them. Ask what the support boundary is when a problem could sit in Bob, the model or your own infrastructure.
Ask also what is counted. In an agentic tool, usage can mean seats, tasks, tokens or agent actions, and the choice changes the economics. If the entitlement is by seat, a small team may be cheap to run. If it is by usage, a large modernisation project could look very different.
Where it fits
IBM also announced an extension of its AI-powered IBM Z application modernisation to self-hosted deployment, with new analysis capabilities. For organisations with large mainframe estates and strict data rules, that combination is the headline. It is worth testing against a real project rather than a demonstration, because the effort in modernisation sits in the details of each application.
Who this is for
The clearest fit is an organisation with sensitive source code or regulated data that has so far kept AI coding tools out of its environment. Banks, insurers, public sector bodies and defence suppliers are the obvious examples, and so are companies with large estates of older applications where the code cannot leave the building. For them, the self-hosted option removes a reason to say no.
It is less obviously a fit for a small team that simply wants help writing code. For them, the hosted option avoids the work of running models and infrastructure, and the data concerns may not apply. Self-hosting is a choice to take on more responsibility in exchange for more control, and it only makes sense if the control is worth having.
Governance questions to settle first
Moving an AI tool inside your estate does not remove the need for governance. It changes who is responsible. Decide who can use Bob and on which repositories. Decide what is logged, where logs are kept, and who reviews them. Decide how generated code is reviewed before it is merged, and how you record that an agent made a change. IBM’s UK announcement says the option is designed for organisations that need greater control over data residency, security policies and AI governance, and those are exactly the areas to document before rollout.
Check the legal side as well. Ask what rights you have in code produced with the tool, and what the licence of each model says about its outputs. Those answers matter for intellectual property, and they are easier to get before a project starts than after.
Start with a pilot
A sensible first step is a limited pilot on a real but contained application. Measure what it costs to run, including infrastructure and staff time, and what it produces. Compare that with a hosted alternative if one is available to you. Use the results to decide whether to expand, and to negotiate from facts rather than from a demonstration.
Costs that do not appear on the licence
Self-hosting has costs that never show up as a line on an IBM order. Skilled people are needed to run models and the surrounding platform. Hardware must be bought or rented and kept current. Security and compliance teams will want to assess the deployment. Each model update means testing before it goes live. A fair comparison with the hosted option has to include all of these, not just the subscription.
It also helps to agree an exit position early. If you invest in hardware and skills for a self-hosted tool, what would it cost to move away from it later, and how portable are the prompts, skills and workflows your teams build? Those questions are cheap to ask now and expensive to answer in three years.
Conclusion
IBM Bob’s self-hosted option opens AI-assisted development to environments that could not use hosted tools, and it gives customers a choice of where models run. It also moves licensing, infrastructure and model management onto the customer’s side of the line.
Before committing, get the metric, the model licence terms, the infrastructure needs and the support boundaries in writing. The sovereignty benefit is real, and so is the extra work of owning more of the stack.
