Picture a Monday morning at the new NHS South West Cluster ICB. A GP practice manager in rural Wiltshire needs to know the status of a patient's continuing healthcare funding application. She sends an email. Within three minutes, she has received an acknowledgement, a case reference number, the name of the clinician managing the file, and the expected response timeline — without a single human being having touched the keyboard.
Across the cluster, a family in Dorset has submitted a formal complaint about a hospital discharge. In Swindon, a primary care network is chasing a response to a commissioning query. In Bath, a local council officer wants to arrange a meeting with the Place Director. Three different things, three different email inboxes — and in each case, the right information has arrived in the right person's hands within minutes, not days.
This is not science fiction. It is, broadly, what the application of what are called "managed AI agents" could look like in an NHS administrative setting. And it is precisely what we set out to think through the efficiency opportunities in the redesign of the new NHS South West Cluster ICB — a merged organisation covering Bath and North East Somerset, Swindon, Wiltshire, Somerset, Dorset, and Bournemouth, Christchurch and Poole, with a combined commissioning budget of around £8 billion.
This piece describes a strategic analysis carried out using AI tools to examine the proposed structure of the new NHS South West Cluster ICB. It explores where AI-driven automation could improve efficiency and patient-facing services. It is not a proposal for imminent implementation — it is a contribution to thinking about how future NHS organisations might be designed differently.
The challenge: redesigning a very large organisation
Merging three Integrated Care Boards — NHS BSW, NHS Somerset, and NHS Dorset — into a single cluster is one of the most complex organisational transformations happening in the NHS right now. The new structure will employ hundreds of people across six geographic "places", five clinical and operational directorates, and a shared cluster-wide leadership team.
The proposed structure, when you look at it laid out in full, is extensive. There are over 344 distinct job roles, ranging from Band 3 administrative assistants through to Very Senior Managers and clinical directors. Many of these roles exist in geographic multiples — the same function replicated across each of the six places, because historically, each ICB ran its own version of the same administrative team.
The question that was put to us was a genuinely difficult one: given the financial constraints that NHS ICBs are operating under, given the technology that now exists, and given the direction of travel from NHS England toward leaner, more strategic commissioning organisations — is there a smarter way to design this?
What the AI actually did — and why it mattered
Before we get to the conclusions, it is worth being clear about something: the AI tools used in this analysis did not make decisions. They did not tell us who should or shouldn't have a job. What they did was read, compare, and summarise — at a speed and scale that would have taken a human team several weeks to replicate.
In practice, this meant processing all 344 job descriptions, extracting the key functions from each role, grouping similar functions together, identifying where the same tasks appeared across multiple roles, and producing a structured analysis of which parts of each role were primarily administrative and which required professional judgment or clinical expertise.
Think of it less like a robot making decisions, and more like having a very thorough research assistant who can read 344 documents over a long weekend, never gets bored, never misses a line, and produces a neatly organised summary without you having to ask for it twice.
The analysis revealed something that many NHS managers already sense but rarely get to see clearly laid out: a very significant proportion of the working week across administrative and coordinator roles is spent on tasks that follow predictable rules. Acknowledging emails. Checking documents. Logging information. Chasing actions. Preparing meeting papers. Updating spreadsheets. Sending reminders.
None of these tasks require professional expertise. None of them benefit from years of NHS experience. They require accuracy, consistency, and availability — which, as it turns out, are things that software is rather good at.
- 45% Admin & coordination (meetings, papers, action logs)
- 20% Correspondence & communication management
- 12% Reporting, data returns & performance monitoring
- 15% Meeting attendance & diary management
- 8% Professional judgment & specialist decisions
Based on analysis of key accountabilities listed across Band 4–6 administrative and officer job descriptions in the proposed cluster structure. The 77% figure represents the proportion of typical working week spent on tasks that follow defined rules and processes — the category that AI agents are designed to handle.
What is a "managed agent" — and why does it matter?
The phrase "AI agent" tends to provoke one of two reactions: either eyes glaze over, or people picture something from a science fiction film. Neither response is particularly useful.
A managed agent, in the context we are describing, is something much more mundane — and much more practical. Think of it as a very well-trained member of staff who operates entirely through email, follows a detailed rulebook, never takes a day off, processes work instantly, and immediately escalates anything unusual to a real human being.
Each agent is defined by a detailed set of written instructions — in essence, a very thorough job description — that tells it exactly what it is allowed to do, what it must pass to a human, what it must never do, and who to contact in different situations. It operates within those boundaries absolutely. It cannot deviate. It cannot be persuaded to do something outside its rules. And every single thing it does is logged.
Imagine hiring a temp who arrives on their first day with a complete manual for every scenario they might encounter. For ninety-five per cent of enquiries, they follow the manual and handle it flawlessly. For the other five per cent — anything unusual, sensitive, or requiring professional judgment — they immediately refer to the appropriate senior person. That is essentially what a managed agent does, except it works twenty-four hours a day, handles dozens of enquiries simultaneously, and costs a fraction of what a human employee costs.
In our analysis of the cluster structure, we identified 24 distinct managed agent roles that could collectively handle the administrative functions currently performed by the administrative and lower-band coordinator workforce. These are not general-purpose AI systems doing whatever they feel like — each one has a specific job, a specific set of rules, and a specific human supervisor it reports to.
A selection of proposed agents and what they do
Triage & Routing Agent
Receives all inbound email, classifies the type of enquiry, and routes it to the right team or agent within minutes — 24 hours a day, 365 days a year.
Complaints Agent
Handles the administrative lifecycle of formal complaints — logging, acknowledgement, chasing, and tracking — while a human Complaints Manager focuses on resolution.
Place Business Support Agents (×6)
One agent per place area — Wiltshire, Somerset, Dorset, Swindon, BaNES, BCP — handling all diary management, meeting coordination, and stakeholder correspondence for the Place Director's team.
Information Governance Agent
Manages the administrative side of FOI requests, Subject Access Requests, and DSPT submissions — tracking statutory deadlines and preparing draft responses for human sign-off.
Finance Support Agent
Processes routine invoices and payments within approved thresholds, manages budget monitoring cycles, and flags anything requiring a qualified Finance Business Partner to review.
PALS & Engagement Agent
Handles patient enquiries and concerns, distinguishing between informal PALS matters and formal complaints, coordinating engagement events, and maintaining stakeholder contact lists.
Three scenarios: what changes in practice
Abstract descriptions of AI agents can sound impressive on paper while meaning very little in practice. So let us look at three specific, real-world scenarios — the kind that happen dozens of times a week across a large NHS ICB — and see what changes when an agent is handling the administrative side.
Scenario 1: A formal complaint from a member of the public
What happens now
Staff involved in admin steps: 3–4 people
What happens next
Staff involved in admin steps: 0 — fully automated
Scenario 2: A GP practice requesting a CHC assessment update
A common source of frustration
GP practice satisfaction: A known pain point
Speed and clarity from the start
Benefit: Clinical staff freed to focus on assessments
Scenario 3: An FOI request from a journalist or member of the public
A statutory obligation with tight deadlines
Statutory compliance, automated
What about the people?
This is, understandably, the question that matters most. If agents are handling the administrative work, what happens to the staff who currently do it?
It is important to be direct about this: an honest analysis of the efficiency opportunities does involve a reduction in the number of lower-band administrative posts. That is a consequence that cannot be softened by clever language. However, several things are also true at the same time.
First, the NHS is not currently short of things for skilled people to do. The organisations merging into the new cluster face significant pressures on primary care access, health inequalities, integrated care development, and population health improvement. Many of the professional staff who would be freed from administrative burden by agents are exactly the people who should be spending more of their time on those challenges — not chasing action logs and preparing meeting papers.
Second, the redesigned structure still requires a substantial and skilled human workforce. Clinical roles, strategic roles, finance and governance professionals, IT specialists, and data scientists are all retained in full. What changes is that every one of those remaining professionals has their time freed to focus entirely on the work that actually requires human expertise.
Roles that remain fully human
New roles created by the agent model
The numbers: what the analysis found
When we applied this analysis across the full proposed structure of the NHS South West Cluster ICB, the results were significant. Across the six directorates and six place teams, the agent model could support the removal or consolidation of approximately 82 posts — a reduction of around 23% of the overall establishment.
The savings are concentrated at the administrative end of the pay scale. The clinical, strategic, analytical, and technical workforce is largely unaffected. What changes fundamentally is the shape of the organisation — moving from the traditional wide-based pyramid to a flatter structure where every human is focused on professional or strategic work.
The full financial picture
The salary saving is only one side of the equation. Running a network of 24 AI agents does carry costs — but they are considerably smaller than one might expect.
| Item | Annual cost |
|---|---|
| Costs of the agent model | |
| AI API usage (Claude, with caching — ~500 interactions/day across 24 agents) | ~£20,000 |
| Infrastructure: N8N automation platform, NHS.net integration, cloud hosting | ~£150,000 |
| Security, audit logging, DSPT compliance monitoring | ~£50,000 |
| Maintenance, agent configuration updates, IT support | ~£60,000 |
| AI Operations Lead (Band 8A fully loaded: salary + pension + NI) | £78,268 |
| Directorate Interface Leads × 4 (Band 6 fully loaded) | £225,502 |
| Total annual cost of agent network | ~£584,000 |
| Annual benefit | |
| Salary + on-costs saving from 82 eliminated posts | £4,655,940 |
| Net annual saving | ~£4.1 million |
Remarkably, the AI technology itself — the actual computing power that runs the agents — accounts for only around £20,000 of the annual cost. That is approximately what the NHS currently spends on a single Band 3 administrative post. The dominant costs in the agent model are the human governance layer and the infrastructure to run it safely on NHS systems — not the AI itself.
What this means for patients and partners
The efficiency case is compelling. But there is an equally important argument that has nothing to do with money — and everything to do with experience.
Anyone who has ever tried to navigate an NHS ICB to get an answer to a query about continuing healthcare funding, or to follow up on a complaint, or to find out why a referral has not progressed, will recognise a familiar frustration: long waits, unclear processes, and the sense that your query has disappeared into an administrative void.
Much of that experience is not the fault of individuals — it is the fault of systems under pressure. Overloaded administrative teams, too many emails, too few hours in the working day. When an agent is handling the administrative layer, the system no longer relies on a human finding time to send an acknowledgement or log a case. It happens automatically, instantly, and consistently.
For GP practices and primary care networks, this matters enormously. They currently spend significant time trying to track down information from the ICB — information that is sitting in a system somewhere, but is not being surfaced in a timely way. An agent model changes that dynamic fundamentally.
For local authority partners and VCSE organisations, it means that the Wiltshire Council officer who wants to arrange a meeting with the Place Director does not wait two weeks for a response. It means that the Healthwatch volunteer who wants to understand the ICB's engagement calendar gets a clear answer the same day.
| Stakeholder | Today | With the agent model |
|---|---|---|
| Patient making a complaint | ✗ 1–3 day wait for acknowledgement; unclear timeline | ✓ Acknowledged within the hour; case reference issued; timeline confirmed |
| GP practice chasing a CHC query | ✗ 3–7 day wait; often no update unless chased again | ✓ Status update within 2 hours; escalated to clinician if needed |
| Council partner requesting a meeting | ✗ Email to shared inbox; may sit for days | ✓ Place Agent responds within the day; diary options offered |
| VCSE organisation with a service query | ✗ Unclear who to contact; often rerouted multiple times | ✓ Triage Agent classifies and routes immediately; right person alerted |
| Journalist submitting an FOI request | ✗ Manual logging; deadline tracked in a spreadsheet | ✓ Logged immediately; statutory clock tracked automatically; no missed deadlines |
| Primary Care Network asking a commissioning question | ✗ Sits in queue; may be rerouted between teams | ✓ Classified and routed to the right Commissioning Business Partner within minutes |
What keeps it safe? The guardrails that matter
It would be irresponsible to describe the potential of AI agents in an NHS setting without being equally clear about the safeguards required. The NHS is not a commercial call centre. It handles some of the most sensitive information about some of the most vulnerable people in the country. Getting this wrong would be seriously harmful.
The agent model proposed here has several layers of protection built in from the outset.
Explicit limits on what each agent can do. Every agent has a clearly defined list of actions it is permitted to take — and a list of things it must never do. An agent that handles FOI requests cannot make clinical decisions. An agent that manages meeting diaries cannot commit the ICB to any financial obligation. These limits are written into the agent's instructions and cannot be overridden.
Mandatory escalation for anything unusual. Every agent has a set of defined escalation triggers — situations where it immediately passes control to a named human. A safeguarding concern is always escalated to the QSE Duty Lead within minutes, regardless of what time it arrives. A media enquiry is always passed to the Communications Director. An MP's email is always flagged to the Head of Corporate Affairs. The agents do not make judgment calls in sensitive territory. They escalate.
Complete audit trails. Every email an agent sends, every decision it makes, every escalation it triggers — all of it is logged. This means that if something goes wrong, there is a clear record of exactly what happened and when. This level of audit trail is, in practice, more thorough than most human administrative processes.
DPIA and DSPT compliance for every agent. Before any agent could go live on NHS systems, a full Data Protection Impact Assessment would be required for those handling patient-identifiable information. Each agent would also need to be registered as an information asset under the NHS Data Security and Protection Toolkit. These are not optional extras — they are prerequisites for go-live.
Humans still make all the important decisions. Clinical eligibility decisions, strategic commissioning choices, media statements, financial commitments, safeguarding responses — every one of these remains with a named, accountable human professional. The agents handle the pipeline. The humans make the calls.
A consideration for the future — not a plan for tomorrow
This analysis was carried out to inform thinking about how NHS organisations might be designed differently as AI capabilities mature. It is not a proposal that the new NHS South West Cluster ICB should implement a managed agent network in the next six months. That would be premature, and likely counterproductive.
The implementation pathway we outlined is a ten-phase programme spanning around two-and-a-half years, starting with the most straightforward administrative functions and expanding only as each phase is proven to work safely. The early phases — which would involve only triage, complaints administration, and information governance — would themselves require months of sandbox testing, IG sign-off, DSPT registration, and careful human oversight before any agent communicated with an external party.
The bigger picture
The most important thing this analysis reveals is not the cost saving — though £4 million a year is not a number to dismiss lightly in a stretched NHS environment. It is the scale of the administrative burden that NHS professionals currently carry, and how much of their working week is spent on tasks that contribute nothing to clinical outcomes, strategic thinking, or organisational improvement.
A senior commissioning professional who spends 40% of their week chasing actions, preparing meeting papers, and managing email queues is not working at the level of expertise that their training and experience qualifies them for. Neither is a Band 7 programme lead whose mornings are taken up with coordinating meeting logistics. When agents absorb that administrative load, these professionals do not become redundant — they become genuinely strategic. They do the work that NHS organisations were built to do.
The technology to make this happen already exists. The main barriers are not technical — they are about governance, trust, and the entirely reasonable caution that any organisation should apply before changing how it operates at scale. Those barriers can be overcome with the right planning, the right safeguards, and the right pace.
What this analysis demonstrates, above all, is that the question is no longer "could AI do this?" The question has become "how do we make sure it does it safely?" That is a much more productive conversation to be having — and one that the NHS, with its extraordinary institutional knowledge, its professional standards, and its deep commitment to the people it serves, is entirely capable of leading.