Own the System, Rent the Model

Every few months a new AI model arrives that is faster, cheaper or smarter than the last. Leaders ask which one to pick. The better question is what surrounds it, because that is where the results come from.

Our new white paper, Own the System, Rent the Model, explains in plain terms why the model is the easy part, and what it takes to turn any large language model into dependable work.

A model keeps nothing

A large language model reads the text it is given, its context, and predicts what comes next, including requests to call tools. It keeps nothing between requests. It cannot open your ERP, does not know how your organization approves a purchase and has no idea what it was asked to do yesterday. Everything that makes it useful has to be supplied, every time, by the software around it.

That is why two organizations using the same model get very different results. The difference is in what goes into the context, what the agent is allowed to do and what is kept afterwards.

Nineteen building blocks, on your side

The paper sets out nineteen building blocks in five groups that assemble the context for every step:

• Knowledge: skills, memory and storage, so the model follows your way of working and uses your facts.
• Action: connectors, tools, plugins, devices and artifacts, so it can do real work and leave reviewable output.
• Work: requests, projects, schedulers and notifications, so it knows what it is working on, and when.
• Control: roles, approvals, credentials and the audit trail, so the right person decides and every action is recorded.
• Agents and models: named agents, the models and plans that power them, and evaluations that prove they still work.

Kept in an agent hub, these become company assets: reusable, reviewable, auditable and independent of any one AI vendor.

Any model, under a plan

When the building blocks live on your side, the model becomes the one replaceable part. Claude, ChatGPT, Gemini, Grok or the next model connects under a provider plan, passes an evaluation on your own tasks and gets to work. Nothing you have built is lost, and no single vendor's price, terms or outage can stop your operations.

How to decide what to automate

The paper is practical about where AI belongs. Automate processes, not tools: break each process into steps and give every step to the cheapest worker that can do it reliably. Rules handle predictable steps, agents handle varied inputs such as documents and email, and people keep the consequential decisions. A worked example shows a supplier-invoice process where only four of eight steps involve an agent, and none commits money without a person.

Each process then climbs one rung at a time, from assist to draft, to act with approval, to act and report, to running on its own, and only when evaluations and error rates say it is ready.

The right size is smaller than most expect

Every extra tool, agent and schedule costs context, review and attention. The paper gives starting ranges for a mid-sized organization: a sharp skill for each repeatable process, usually fewer than thirty tools per agent, and a handful of agents organized by role rather than one agent for everything.

Inside the paper

• How an LLM works, and the three facts that shape everything else
• The nineteen building blocks and the eight-step work lifecycle, with escalation from any step
• Vendor agents and the hub, and when to build, buy or use a managed service
• Governance and security mapped to OWASP and NIST, plus public-sector requirements for records, disclosure and accessibility
• Data readiness, an illustrative cost model, the operating model for people, and a pilot scorecard
• A 15-question readiness self-assessment, a reference architecture and a glossary

Who should read it

CEOs, CIOs, CFOs and COOs, and the IT, operations and department leaders deciding how to put AI to work, especially in the public sector and mid-market organizations that need AI to do real work safely.

Download the free white paper to see the full framework, the sizing guidance and the readiness checklist.

*Worked examples, cost figures and sizing ranges in the paper are illustrative guidance, not client results or benchmarks. Product and model names are trademarks of their respective owners; their use does not imply partnership or endorsement.*