The outsourcing saving only existed after the old costs left

Being congratulated for a saving you can’t feel is its own small insult.

The spreadsheet says the function costs less. The supplier invoice is lower than the old payroll line.

Meanwhile you’re chasing exceptions, the old system is still open because somebody needs it, and the people who knew the odd cases are reviewing work that was supposed to have left.

You’re not bad at numbers because that feels unreal.

You’re not refusing progress because you notice the old work continuing round the edges of the new service.

You’ve got targets, a team to protect, a budget meeting, and a day full of requests that can’t wait for the business case to become elegant. When the number arrives presented as settled, accepting it feels like the only adult move available.

Nothing turns on you in this story.

It’s arithmetic.

A capacity number can put on a cash-saving costume long before the old costs have gone anywhere.

2001 — Finance puts your payroll total beside a supplier quote. The gap looks clean because the systems, the managers and the severance sit in other columns.

2008 — Your team stays. It manages the contract, judges the exceptions, and explains the work the contract never learned to name.

2024 — An AI demo answers in seconds. The saved time is visible enough to make the rest of the ledger look fussy.

2026 — The licences, the reviewers, the supplier consumption and the old system are all still on the invoice.

The saving had a before

Outsourcing learned this trick in public, which is why the record is so useful to you.

Nobody could responsibly compare an internal service desk with a provider’s monthly price until somebody described the service.

Applications, devices, users, locations, hours, languages, incident types, security boundaries, internal-only cases, ticket demand, response and restoration time, reopen rate, satisfaction, backlog age, and the work of keeping the desk alive.

“IT support” was never a unit. It was a fog with a payroll total attached.

The Government Accountability Office puts two phases before contract development.

Sourcing strategy, then the operating model.

That order is the whole lesson. Issue a tender before you understand your service and you’ve let the supplier’s price make your design decision.

UK sourcing guidance carries eleven key policies and asks for a pilot the first time a service goes out. Neither is caution as theatre. They’re how you find out whether the costed service is the service anyone needs.

Then came mobilisation, which a neat comparison treats as a corridor rather than a room.

Process mapping. Data clean-up. Recruitment. Transfer administration. Training.

Access design. Technology configuration. Parallel running.

Site changes. Asset inventory. Knowledge capture.

All of it before the new desk could answer a single call.

And you paid too. Diverted specialists, programme management, legal and procurement support, security review, severance or redeployment, and the retained team you kept.

A provider can bury some of that in a monthly charge, quote it as a project, or recover it through a longer commitment.

It doesn’t stop being paid for.

The reviewed projects put planned transition at thirty to ninety days.

Those are examples, not your timetable.

Their value is duller than that. Transition was an explicit, funded phase. It had a start, a dependency chain, and an acceptance condition.

Your saving didn’t begin when somebody signed a contract.

The old ledger has a name worth keeping. A transition ledger. Owner, cash timing, accounting treatment, dependency and acceptance condition beside every cost.

And it leaves room for what only appears when a provider touches your real records. The undocumented exceptions. The unusable data. The legacy integrations.

The people who don’t transfer.

A bid assumption isn’t a fact because it appears in a bid.

The bill after the bill

Once live, an outsourced service never had one price.

Fixed monthly fee. Per transaction. Per full-time equivalent, site, asset, user, device, call. Usually some mixture.

The provider still paid for staff, subcontractors, software, facilities, training, quality assurance, working capital, insurance, overhead, profit. You still paid for governance, internal experts, demand management, enterprise technology, audit, security, sometimes the licences the provider used.

One invoice was never the whole function.

And your pricing unit decided whose problem success became.

A per-ticket rate can reward self-service and automation. It can also reward a fast closure that sends the same person back tomorrow.

A per-FTE model can preserve expertise. It can also preserve headcount after your demand falls.

A fixed fee makes planning easier, if its volume band and change mechanism are real.

The question was never whether the rate looked low. It’s whether lower demand cuts your cost, raises the unit rate, or quietly becomes the provider’s margin.

Now put the AI demo beside that, and notice you don’t need a dramatic analogy.

A model call has a price. A production service has model inference, hosted seats, storage, cloud compute, security tooling and supplier support.

It also has product and process ownership, data engineering, integration, evaluation, security review, legal and procurement work, training, help desk, monitoring, incident response.

Then it has the change itself. Redesigning work. Running in parallel. Auditing output. Absorbing errors.

Salaried time doesn’t become free by avoiding an invoice.

Stanford’s AI Index records a GPT-3.5-equivalent query falling from US$20 per million tokens in November 2022 to US$0.07 per million tokens in October 2024. More than a 280-fold change.

That’s a real change in a real input price.

It isn’t proof that the service around the query is nearly free. It’s the reason your model-price assumption needs a date and a sensitivity range, while integration and operating costs stay visible.

Four things that look alike

The old record separates what a presentation likes to blend together.

A lower supplier invoice is one thing. An avoided future hire is another. Capacity redeployed to more valuable output is a third. A quality or risk benefit is a fourth.

All four can make your decision sensible.

They can’t all be booked as cash, and they can’t be added twice because a slide wants one big number.

That matters most where productivity has no agreed pass-through.

A provider promising to automate a fifth of your cases hasn’t yet said who gets the benefit, who pays for implementation, whether quality has to hold, or what happens when your process blocks the change.

The commitment can enrich the provider, cut your bill, or fund a shared improvement.

The percentage isn’t the outcome. The pricing mechanism is.

Enterprise AI has the identical temptation, because a saved minute shows up so cleanly in a demo.

Your financial bridge starts with eligible cases, adoption rate, median minutes saved after quality checks, fully loaded cost per minute, and the share of capacity you can actually demonstrate was redeployed or removed.

Then it subtracts review time, exception handling, supplier charges and source maintenance.

It doesn’t turn the worst imaginable risk into realized return. And it doesn’t call a fixed-staffing queue a labour saving because the drafts arrive faster.

The reserve nobody calls a cost

Outsourcing left one quieter lesson. The team that stays is not a failed transition.

It’s your ability to own strategy, performance, demand, exception policy, audit, data, access, and the argument when the provider says a change is out of scope.

Underfund that and the provider becomes the real designer of your service. The apparent saving then depends on your inability to see what you gave up.

Exit belongs on the same bill.

A credible fallback needs documents, portable data, access transfer, emergency capacity, successor onboarding and parallel operations. If the work comes back in-house, there has to be a receiving organisation before the supplier lets go.

The Sourcing Playbook puts the outgoing supplier’s exit and the incoming team’s mobilisation into one integrated plan. Milestones, roles, a joint risk register, asset transfers.

That’s your economic reserve in plain language. Portability isn’t a philosophical preference. It’s what stops replacement becoming a panic purchase.

AI has its own version of the same list. Data-use terms, regions, security obligations, change notice, support, portability, exit assistance, controlled source access, a held-out evaluation set, a rollback plan, a manual fallback.

Your vendor contract is one layer of the service. It can’t test your records, your workflow, your users or your consequences on your behalf.

You’re not missing the gain. You’re refusing to count the same capacity twice.

Old costs have a habit of surviving under new names. Governance. Review. Source maintenance. Duplicate systems. Exception work. Exit risk.

The service may well be worth buying. The draft may genuinely save you time.

But the saving only becomes real after the old costs leave. Or after somebody is honest about what you decided to keep.