Sixty per cent of a UK AI integration's lifecycle cost arrives after go-live, in token consumption, licensing and support retainers, and Primewise's 2026 pricing guide brackets the build itself between £15,000 for a single workflow and more than £750,000 for an enterprise programme. That split is usually read as a budgeting warning. It is better read as an architectural one: the expensive half of an LLM system is the half that runs every day, and what it runs inside is a business process designed before the model existed.
CIO reports that CDW has put assistants in front of more than 14,000 coworkers and built more than 12,000 agents, and that usage climbed without telling the company whether the business was performing better. That is the correct test, and it exposes the usual failure. A task that finishes in seconds and then joins a queue that clears over days has not improved the process containing it. The intelligence rose; the throughput did not. The constraint did not disappear. It relocated to the next handoff, approval or re-keying step.
The bottleneck worth architecting against is therefore not the model call. It is the path the output takes after the model returns. Integration work now means redrawing the process an LLM sits inside, then building the architecture that keeps the redrawn process observable, bounded in cost, reversible when a step misfires, and honest about where a human still carries accountability. Treating those as separate jobs produces a fast task inside a slow system. Treating them as one job is the discipline this page is about.
The handoff is the constraint, so architect the process before the model call
A process does not speed up because a task inside it did; the constraint relocates to the next queue, approval or exception path. The first instinct after that realisation is to point a model at every step in the existing flow. CIO names the error plainly: "Don't automate the workflow you have." The steps in a mature process exist for reasons that have often expired, built around the limits of the previous system or around a departmental boundary that no longer justifies its cost. Automating each one produces a faster version of a process that should have been redesigned.
The business outcome is the useful unit of analysis. Define the result the organisation needs, then work backwards through the steps that still earn their place. Some steps disappear. Some become a rule. Some become a model call. Some remain with a person because judgment, accountability or a regulatory signature cannot be delegated to a probabilistic component. This is not a technology choice made once at the start. It is a sequence of design decisions that should be recorded, owned and revisited when the model or the process changes.
Anthropic's engineering team draws a distinction that helps here. Its guide to building effective AI agents separates workflows, where LLMs and tools are orchestrated through predefined code paths, from agents, where the model directs its own process and tool use. That distinction is architectural. It determines how much of the process is fixed in code, how much is decided at runtime, and how much evidence the system can produce when someone asks why a step happened.
A deterministic workflow is the right choice when the path is known, the exceptions are enumerable and the value comes from consistency. It is easier to test, cheaper to run and simpler to audit. Its weakness is that it cannot improvise when the input does not match the expected shape.
An agent earns its place when the number of steps cannot be predicted, the environment can return ground truth at each turn, and the cost of a wrong turn is contained by a sandbox, a budget ceiling or a human checkpoint. Its strength is that it can move work forward through an open-ended task. Its weakness is that autonomy compounds errors and spend, so it needs tighter instrumentation than a fixed workflow.
The choice is not prestige. It is fit. RaftLabs' cost breakdown for AI workflow automation makes the same point from the delivery side: data quality and integration count move budgets more than model choice. That matches what the handoff analysis predicts. The expensive work is making the process legible, connecting the systems it touches and defining the exceptions. A cheaper model does not repair a process that nobody has mapped.
This is why the cost model should be per completed business outcome. A workflow that completes a quote, resolves a claim or closes a reconciliation has a measurable cost and a measurable cycle time. A licence count does not. The moment finance can see cost per completed outcome, it can compare a rule, a workflow and an agent on the same basis. The moment operations can see rework rate and queue time after each handoff, it can tell whether the model improved the process or merely moved the waiting.
Build a workflow contract that makes the LLM a bounded component
The workflow contract is the artefact that keeps an LLM integration from becoming an unowned API call with a business process wrapped around it. It should exist before the build, be versioned with the code and be short enough for a business owner, an engineer and a risk reviewer to read in the same meeting. The contract does not describe the prompt in detail. It describes the boundary: what outcome the workflow serves, who owns it, which systems it may touch, which actions it may take, where approval is required, what happens when the model fails and which measure proves the process changed.
A functional reference architecture for LLM-integrated systems published on arXiv is useful because it treats the model as one layer among several. It separates presentation, application logic, LLM integration and data management, with monitoring and guardrail sidecars spanning those layers. That structure is a practical reminder: observability and policy enforcement are not features added after launch. They are cross-cutting components that must be designed into the workflow from the start.
workflow: quote_to_order
outcome: signed_order_within_policy
owner: commercial_operations
trigger: quote_request_received
systems: crm, pricing_service, document_store
allowed_actions: draft_quote, request_margin_approval
approval_gate: margin_below_policy
fallback: human_queue
measure: cycle_time_and_rework_rate
The fields are deliberately boring. The outcome is a business result. The owner is a role with authority to remove steps. The trigger and systems list bound the blast radius. The allowed actions separate drafting from committing. The approval gate names the condition under which a person must sign. The fallback says what happens when the model is unavailable, uncertain or wrong. The measure ties the workflow to an outcome the business already tracks.
Version that contract with the code. A prompt change, a model change or a tool change is a workflow change, and it should require the same review as a change to an approval rule or a pricing calculation. This is where many LLM integrations lose control. The model is swapped quietly, the prompt is edited in a console, the tool schema changes underneath an agent, and nobody can reconstruct which version produced a customer-facing decision. A contract does not prevent change. It makes change attributable.
CIO's report makes the same point at enterprise level. The CIO needs to provide the foundation quickly enough and usefully enough that business teams build on it. That foundation combines approved tooling, trusted data with business context, connections into line-of-business systems, governance, and visibility into cost and value. Without it, teams find their own tools and build their own connections, and IT inherits fragmented data, duplicated spending, inconsistent controls and little visibility into what AI is doing across the organisation.
The human side has to be designed with the same precision. Broad access builds fluency, because people learn what a model can and cannot do by using it on their own tasks. Leadership focus supplies authority, because only a business owner can remove a handoff that crosses functions. Neither scales if every team solves tools, data, security and cost for itself. The foundation has to be useful enough to become the default and constrained enough to remain governable.
Measure what changed. Licences, usage counts and hours saved are activity metrics. They can rise while the business stays flat. The question that survives scrutiny is whether a workflow now completes with less rework, a shorter cycle time, a lower cost per outcome or a better decision at the approval gate. If the answer is no, the model may be impressive and the integration may still be decorative. The architecture should make that visible before the next budget cycle.
Sources
- Your AI is smarter. Is your business still working the same way?
- AI Integration Cost UK 2026: Real Pricing From £15k to £750k
- Building effective AI agents
- A Functional Software Reference Architecture for LLM-Integrated Systems
- AI Workflow Automation Cost in 2026: Full Breakdown
Source: Your AI is smarter. Is your business still working the same way?, CIO





