Know what the promise will cost.
The shared-service centre is becoming an agent queue, and the chargeback will arrive before the saving does
The internal centre that once consolidated payroll and HR is the immediate precedent for an AI operations layer. It shows why a faster case is not yet a cheaper function.
What it feels like
The strange irritation of watching a simple HR request become faster while feeling less and less like it belongs to anyone.
A resolved case and a realised saving are different events.
Your small request has been made easier. You’re irritated, and you can’t say why.
The portal remembers the form.
An answer arrives with a confirmation number.
Nothing is wrong, exactly. The correction is done.
And yet your request feels lifted out of your hands and returned without a person who can be wrong in the old, comprehensible way.
You didn’t ask for your work to become a queue. You asked for the queue to stop swallowing your afternoon.
You’ve got payroll to sort out, a manager waiting, a child to collect, and no appetite for explaining a changed address to three different inboxes.
Taking the faster route isn’t surrender. It’s the ordinary decision of somebody whose day is already full.
No machine has plans for you.
There is a bill, though.
It’s about a bill.
Your internal service centre is learning to become an agent queue.
The old promise was one front door for HR, finance, payroll and procurement. The new promise is that the front door can read your case, fill the fields, prepare the change, and sometimes announce it’s complete.
That promise has an older shape than anybody is admitting.
2011 — You email HR because the portal can’t find your old address.
2017 — A ticket number arrives before a person does.
2023 — The form fills in the fields you forgot.
2026 — An agent says your payroll record is corrected. You wonder who’s allowed to be wrong.
Naming the unit
The shared-service centre isn’t a metaphor for this.
It’s the immediate precedent.
It took work that had names, local habits and your private email trails, and made it catalogue work. A requestable unit with a scope, evidence, an outcome, a clock and an owner.
“Employee data support” was too broad for anyone to forecast or measure. “Process a legal-name change in payroll and benefits after approved evidence” could be counted, routed, audited and charged for.
That act of naming is genuinely useful.
The UK government’s 2024 HR reference model lists more than 2,000 common glossary terms and more than 750 individual process maps.
Those aren’t targets for you. They’re a record of the granularity your modest request needs once it crosses policy, data, workflow and hand-off boundaries.
Your simple form only feels simple because somebody named the fields underneath it.
Then the centre does something that sounds clerical and is really a theory of authority.
Triage checks your completeness, your entitlement, your route. Then it assigns work by skill, language, jurisdiction, urgency and system access.
A case needing a statutory check should be recognized before a standard change propagates through three systems.
Waiting for your document can pause the service clock. Teams arguing about who owns the problem can’t be allowed to make their delay disappear the same way.
That distinction is what the old apparatus has to teach the agent pitch.
A standard case can be made fast. An exception cannot be made imaginary.
When your payroll record conflicts with an immigration document or a local legal rule, a trained agent in a centre doesn’t invent a workaround to protect a metric.
Your case goes to a named specialist. A global process owner decides whether the standard is missing. A local policy owner decides whether your variant is real.
Your request has become visible. Not frictionless.
Transition is a second service
The move itself has always been a second service, with its own queue.
The Hera programme moved three legacy business systems into one in six months while serving more than 20,000 civil servants.
Cutover proves nothing by itself. Stabilization needs an exit test for volume, accuracy, aged work, and what happened to you.
The HR benchmark makes the same quiet point. It defines first-contact resolution as a rate, and average speed to answer in seconds.
A quick answer isn’t a good outcome if it manufactures your next contact.
In the 2019 Deloitte survey, 37% of sampled centres supported one function while 32% supported three to five.
The appeal of a single customer view grows as work crosses functions. So does the need to say who owns your exception.
The UK’s Synergy baseline puts shared-service volume around 15 million transactions a year across HR, finance, payroll and procurement.
At that scale, classification and validation matter. So does the fact that one quiet defect travels fast.
The same move, deeper down
The agent is that move again, further down the stack.
It takes a case that was catalogue work and turns it into a bounded run. One immutable identifier tied to you, a record, a policy version, a time.
That isn’t bureaucratic excess. It’s the difference between “the agent fixed it” and a system that can tell you what actually happened.
An agent may interpret documents and choose tools where a fixed workflow can’t. But a deterministic workflow is usually the better choice where the rule is fixed.
The old centre already knew the underlying point. Not every request needs a generalist. And not every deviation deserves a guess.
The language changed from service catalogue to tool catalogue. The limit didn’t.
Your bank-change agent can read named sources and create a reversible draft. It can’t activate the change because it assembled a persuasive sentence.
The old version of that rule was plainer. Don’t give the queue a decision that belongs somewhere else.
And this is where the cheerful sentence — the agent resolved it — starts concealing work.
A pending change isn’t a committed change. A notification saying approval was requested isn’t a gate. A human being nearby isn’t an approval control.
You can feel why it matters even in the harmless HR case.
The calm completion message gives your interaction an owner-shaped surface. The real question isn’t whether the prose sounds certain.
It’s whether the downstream system returned a confirmation number, a record version and a timestamp.
One of those is a story about a transaction. The other is the transaction.
That’s the first correction. You’re not resisting efficiency.
You’re noticing that a resolved case and a realised saving are different events.
Where the money hides
The shared-service precedent is unusually clear here, because the money never hid its mechanics very well.
The Synergy business case starts from over £550 million a year spent on HR, finance, payroll and procurement services.
It doesn’t start from an assumed saving rate. It treats service cost and transaction base as your denominator.
In its second full business case, discounted benefits are £1.238 billion against discounted costs of £656 million. Net present value £582 million.
Those are an estimate for one programme, not a portable return. Their usefulness is the symmetry. Costs and benefits appear in the same discounted frame.
What vanishes when somebody says an agent will make a function cheaper is usually the baseline.
Your baseline isn’t only the salaries of people presumed to transfer or shrink.
It’s your time and your manager’s time, contractors, systems and licences, office and equipment, rework, error correction, consultants, data remediation, and the small pieces of work hidden inside a business-unit role.
It needs volume and complexity too. A low-headcount team can be expensive because other teams spend their days repairing its exceptions. Perhaps yours.
Then comes the cost of becoming orderly enough to automate.
Programme management. Process documentation. Data cleansing. Consulting you. Training you. Dual running. Migration. Integrations. Testing. Cutover support.
And after go-live, your centre staff have descendants. Supervisors, experts, knowledge management, platform subscriptions, cloud charges, cyber controls, quality assurance, retained local roles, supplier management, data stewardship, exception monitoring, rule maintenance, technology renewal.
The agent didn’t remove the service. It rearranged its cost and its authority.
The chargeback arrives first
Allocation is a decision about who gets to feel that rearrangement.
Direct costs belong to a known customer. Shared costs — leadership, common platform, controls, improvement — need an allocation pool of your money.
Demand-driven costs can follow a completed transaction, an active employee, a call, a purchase order, a device, an invoice, or raw consumption.
In the Deloitte survey, 44% reported headcount allocation, 41% volume or service charging, 22% revenue allocation. Respondents could name more than one.
No formula arrives innocent.
A subscription pays for common availability. A variable transaction price exposes incremental demand. Showback reports what a unit would pay before any money moves. Chargeback actually transfers budget.
Those choices change the story when demand falls.
A pure transaction price makes the approved centre look expensive at precisely the moment your business unit decides whether to use it.
Pass every productivity gain straight back to customers and you starve improvement. Keep every gain and your customers start resisting the catalogue.
In the same survey, 48% reported reinvesting productivity savings, including 20% in technology and 14% in process improvement.
Reinvestment is a policy choice. It isn’t proof a saving occurred.
The old record even supplies the qualification management presentations prefer to skip.
More than 45% of Deloitte respondents reported a headcount reduction above 10% within a year of a significant shared-service implementation.
Self-reported historical evidence, not a target. And it says nothing about whether service cost stayed down.
Headcount movement, capacity release and cash saving are three different things, and only one of them shows up in your budget.
Ask the local systems. The retained managers. The exception queue.
The Synergy case applied 46% optimism bias to total investment cost, and found its net present value turned negative once benefits fell by 85%.
Those figures don’t set anybody’s agent programme tolerance. They show what an honest business case admits.
Adoption can be slower. Volume lower. Dual running longer. Technology delayed. Attrition higher. A country exception real. And your claimed benefit merely capacity rather than cash.
The machine can make your form gentler. It can retrieve the record, prepare the proposal, and take the standard case out of your afternoon.
That’s a real improvement.
It can’t make transition vanish. It can’t turn your local exception into a standard one by calling it confidence. And it can’t make an internal bill a saving by changing which ledger it lands in.
The shared-service centre became a queue because the work had to become legible before it could be consolidated.
The agent queue inherits that legibility, and every question it never answered.
What’s the unit? Who owns the exception? What is allowed to commit a change? Who funds the idle hours? And where does the released capacity actually go?
Faster, flatter, less visible. That’s the feeling.
Not that nobody is responsible. That responsibility has been divided into pieces small enough to arrive with a confirmation number.
Put this to work
- A service catalogue makes a request measurable by defining its scope, inputs, outcome, owner, and exception route.
- A complete baseline includes local manager time, rework, licences, data remediation, and hidden work as well as transferred salaries.
- An agent needs a bounded action, scoped tools, and a binding approval gate instead of a broad instruction.
- Subscription, transaction, showback, and chargeback models distribute the benefit of falling demand differently.
- Transition cost, retained stewardship, exception monitoring, and technology renewal remain part of an AI business case after deployment.
- A model message is not proof of completion; the downstream system result and trace are the transaction evidence.
Questions
- Why does a faster HR request not automatically save money?
- A faster HR request reduces elapsed time, but it becomes a cash saving only when spending changes. Released capacity becomes cash when an organization avoids hiring, reduces overtime or contingent labour, eliminates a position, reduces supplier spend, or closes infrastructure. If capacity clears a backlog or improves service, it remains a throughput or service benefit and should not also be counted as a headcount saving.
- What does a service catalogue actually change?
- A service catalogue turns a broad label such as employee data support into a defined requestable service with a customer, scope, inputs, expected outcome, service level, exclusions, privacy notice, and escalation route. That definition creates a measurable unit of work and lets the requester select a service without knowing the centre's internal teams. It does not remove legitimate local or statutory exceptions.
- What belongs in the cost baseline before work moves to a shared centre or an agent?
- The baseline includes local employee and manager time, contractors, systems and licences, office and equipment, rework, error correction, local consultants, data remediation, and work hidden inside business-unit roles. It also records volume and complexity. Transferred salaries alone do not show whether a function costs less, because old local costs can remain while a new platform and transition team appear.
- What is the difference between showback and chargeback?
- Showback reports a business unit's consumption and indicative cost without transferring budget. Chargeback is an internal bill that moves service cost to the business unit using the service. Showback can establish trust before formal billing. Chargeback changes incentives immediately, so its allocation basis needs to be understandable, stable enough to budget, difficult to game, and related to the demand driver.
- Why can a transaction price make the approved service less attractive?
- A pure transaction price makes demand visible, but it can discourage business units from using the approved service or encourage strategic classification of work. A hybrid model commonly uses a subscription for shared availability and a variable price for incremental demand. This separates the fixed cost of a safe common capability from the demand-driven cost that a customer can influence.
- What makes an AI agent a bounded service instead of a broad automation promise?
- A bounded agent begins with a defined purpose such as preparing a specific change for review, not a vague instruction to manage a whole function. The runtime binds the request to a case, record, policy version, and time; exposes only named tools; and validates state transitions. The agent can read evidence and create a draft without automatically receiving authority to commit the change.
- Why is an approval message not the same as an approval gate?
- An approval gate binds an authorized person's decision to a stable proposed action, its evidence, its policy version, and its expiry. The gate checks those conditions before the tool runs and fails closed when they cannot be confirmed. A message that says approval was requested while the agent continues does not block the action and therefore does not create approval control.
- What proves that an agent actually completed a change?
- The authoritative proof is the downstream system's committed result, such as a confirmation number, resulting record version, and timestamp. An agent's final message is not proof because it can describe a result that did not commit. A useful trace links the request, evidence, policy decision, credential scope, approval, tool arguments, external result, and recovery event for later review.
- Why does an agent need different permissions for reading, drafting, and committing?
- Reading retrieves information, drafting creates a reversible proposal, and committing changes a business, legal, or physical state. Each effect needs different constraints. A policy engine independently checks the tool name, arguments, resource, principal, state, budget, and approval before execution. Narrow authority limits the effect of manipulated content, excessive permissions, and excessive autonomy.
- What should happen when an agent cannot verify a request?
- When required evidence or authority cannot be verified, the agent enters a failed or awaiting-evidence state, escalates, or parks the case according to policy. It does not treat confidence as approval. If an activation call times out, the runtime checks the downstream system using the idempotency key before recording completion, retrying safely, or sending the case to an operator.
- Why do exceptions remain part of the business case?
- Exceptions require named decision-makers, retained specialists, monitoring, rule maintenance, and sometimes localized process variants. A recurring exception can reveal poor master data, a missing catalogue scope, or a legitimate local requirement; it is not automatically evidence that standardization failed. These costs continue after deployment and belong beside platform subscriptions, controls, quality assurance, and data stewardship in the run-rate model.
- What should an organization test before calling an AI service cheaper?
- The financial case tests slower adoption, lower volume, a longer dual run, technology delay, higher attrition, country exceptions, and benefits that release capacity rather than cash. It separates transition expenditure, recurring run rate, allocated cost, avoidable cost, and reinvestment. A saving exists only when the whole service cost falls or spending changes, not when a charge moves from a local budget to a central ledger.