A planning job splits into six layers and the middle ones are where scoping goes wrong, because a statement of work written against activities lets the judgement layer drift across the boundary without anyone agreeing to it. Williamson's asset specificity test sorts them: the statistical baseline, data preparation and first-pass exception detection are general work at high frequency, while judgement, commitment and accountability are specific to your business and stay. An accuracy SLA is a trap because the provider controls none of the drivers, so score the stage they own against a benchmark they can beat. Barthelemy's 2001 categories of hidden cost (contracting, transition, ongoing management and reversal) are the ones missing from the business case, and the retained management layer belongs in that case as named positions before any rate card is discussed.
The statement of work said the provider would produce the statistical baseline, maintain planning parameters and clear the exception queue. Three roles came off the internal headcount. Twelve months later the internal team spends Monday and half of Tuesday reviewing what the provider produced, which is more time than the three roles used to spend producing it.
Nobody is at fault in the usual sense. The provider delivered every line in the contract. The internal team is reviewing because the output arrives without the reasoning attached, and reviewing an unexplained number takes longer than making it.
The error was upstream, in a statement of work written against activities. Activities are what a planning job looks like from the outside, and they cut across the layers the job is actually made of, each of which transfers to a third party on different terms.
Six layers in a planning job
Take a demand planner's week apart and it decomposes into six things that happen to sit in the same seat.
- Data preparation. Pulling history, correcting known errors, handling censored periods, aligning calendars, chasing missing master data. Mechanical, high volume, procedural once the rules are written.
- Statistical baseline. Fitting and running models, checking residuals, re-fitting on a cadence, applying reconciliation across the hierarchy. Technical, general, improves with concentration.
- Exception detection. Applying thresholds, ranking what broke, assembling the supporting evidence for each item so the next layer can act. Rules-driven at the front end and interpretive at the back.
- Judgement. Deciding what to do about the exception. Which override to make, whether the sales figure is credible, whether this customer's odd week is signal.
- Commitment. Turning a number into a promise somebody else will act on. Agreeing supply to a market, signing an expedite, telling an account manager what they can sell.
- Accountability. Owning the consequence when the number is wrong, in front of people who can do something about it.
Most planning outsourcing conversations start at the top of that list and stop halfway down without saying where. The first two layers transfer cleanly and everybody agrees. The sixth cannot transfer at all, and everybody agrees about that too. The argument, and the money, sit in layers three, four and five.
Which layers survive the hand-off
Williamson's transaction cost framework, set out in 1985 and built on Coase's 1937 question about why firms exist at all, gives three properties that decide whether an activity belongs inside a firm or across a contract: asset specificity, frequency, and uncertainty. High specificity, meaning the knowledge or investment only has value in this one relationship, pushes an activity inside. High frequency at low specificity pushes it out, because a market can supply it repeatedly and cheaply.
Run the six layers through that.
Data preparation is high frequency and low specificity once the rules exist. The knowledge required is how to apply a documented rule, and that knowledge is general. It transfers.
The statistical baseline is the cleanest case in the whole list. The skill is general, the data is transferable, the work is repetitive, and concentration helps because one team running models across many businesses builds practice faster than one planner running them across one.
Exception detection splits down the middle. Detecting and ranking exceptions is rule application. Assembling the evidence behind each one requires knowing which fields lie in your system and which ones nobody maintains, which is specific knowledge that takes months to acquire and does not appear in any document.
Judgement is the layer where the transfer fails, because the inputs to a good override live outside every system you would give a provider access to. Whether the sales director's number reflects a signed listing or an aspiration. Whether the distributor asking for double is stocking up ahead of a price rise. This is specific knowledge in Williamson's sense, and a contract is a poor instrument for moving it.
Commitment fails for a different reason. A promise is only worth something when the person making it carries the consequence, and the provider's consequence is a service credit.
The delivery shape changes the answer for layers three and four, and the distinction is usually collapsed in the conversation. A captive shared service centre keeps the specific knowledge inside the firm, so it can hold exception triage and, over time, some judgement. A third-party provider cannot, because the knowledge walks out with the account team. Staff augmentation, where contracted people sit inside your process under your management, is a labour arrangement rather than an outsourcing one and should be evaluated as such. A captive centre is a centralisation decision wearing a cost label, and it should be tested with the decision-rights question rather than the rate card.
The SLA a provider can actually control
Almost every planning outsourcing contract is scoped on activity: forecasts produced, tickets closed, cycles completed, hours delivered. Activity metrics are chosen because outcome metrics are contested, and they create the exact incentive you do not want, which is a provider paid by the exception to have more exceptions.
The obvious alternative is worse. Forecast accuracy as a service level fails because the provider controls almost none of its drivers. Assortment churn, promotional intensity, a customer delisting, a competitor's price move and your own sales team's optimism all move accuracy, and none of them are in the provider's gift. An accuracy SLA either gets gamed, because the safest forecast for a provider is the one closest to whatever the internal team will accept without argument, or gets waived after two quarters when both sides recognise it is measuring the weather.
The version that works measures the stage the provider owns against a benchmark that stage can be scored against. If they run the statistical baseline, the test is whether their baseline beats a stated naive benchmark on a rolling-origin holdout, on a named segment, evaluated on the same items every month. That is a contract term a provider can be held to, because it depends on their model choices and nothing else. Scoring each pair of hands that touches a forecast is a method with its own literature and its own post; the point here is that the contract has to name which stage the provider is being scored on, in writing, before go-live.
Two more terms are worth insisting on. Cap or invert the volume-linked element, so that reducing the exception count is worth something to the provider rather than costing them revenue. And require that every action carries its reasoning in a structured field, because the review time that ate the internal team's Monday is created by output arriving without it.
Price the retained organisation before you price the deal
Barthelemy studied what companies actually paid on outsourcing arrangements and reported in MIT Sloan Management Review in 2001 on four categories of cost that sit outside the contract value: searching for and contracting with the provider, transitioning the work, managing the arrangement once it runs, and transitioning back if it ends. He returned to the theme in 2003 with a list of recurring mistakes, two of which are outsourcing activities that should have stayed inside and treating the personnel consequences as an afterthought.
The third category, ongoing management, is the one that decides planning deals, because it is a permanent line rather than a one-off.
Work an example. The deal replaces five internal planners with eight provider analysts at a rate that quotes a forty percent saving on the delivery cost. Against that, you need a service owner running the relationship, a data quality owner who is now full time because errors reach the provider as tickets rather than getting fixed at source, and the judgement layer stays where it was. Call the retained addition 1.5 full time equivalents at internal rates, which are the expensive ones.
If the internal five cost 500,000 a year fully loaded and the provider's eight cost 300,000, the headline saving is 200,000. Add 1.5 internal heads at 100,000 each and the saving is 50,000, a quarter of what was quoted. Add the coordination time the remaining planners spend answering provider questions, which nobody logs, and it moves again. The arithmetic still comes out positive in many cases. It comes out positive at a number that would not have justified the programme if it had been in the paper.
Do that calculation with your own rates before the negotiation rather than after, and put the retained roles in the case as named positions with names attached.
The knowledge that leaves and does not come back
The provider's attrition rate on your account is the single most useful number you can ask for, and it is rarely in the proposal.
Assume it takes fourteen weeks for a new analyst to be genuinely productive on your business, which is a modest estimate for anything with a complicated promotional calendar or a long tail of specials. On a team of eight with thirty percent annual turnover, 2.4 people a year go through that ramp, which is around 34 person-weeks against an annual team capacity of roughly 368. Nine percent of the team's year is spent in ramp, permanently, and perhaps half of that is lost output. That cost never ends, because it is a steady state rather than a transition.
There is a second effect that does not show up as capacity. Every departure takes undocumented knowledge with it, and the reason planning knowledge stays undocumented is that most of it is about exceptions to the documented rules. The internal team absorbs the gap by explaining things again, which is the cost that turns up as review time.
Two protections work. Name individuals in the contract with a minimum tenure on the account, and pay for the continuity if it needs paying for. And require the provider to maintain the runbook as a deliverable, reviewed quarterly by your service owner, with the review being an actual read rather than a sign-off. A runbook that a new analyst can be productive from in six weeks instead of fourteen changes the arithmetic above by more than any rate negotiation will.
Where this stops
The wage arbitrage is real, and the honest version of it is smaller than the rate card difference by roughly the size of your retained layer. There is no general figure for that, because the retained ratio depends on how much of your process is written down, which varies more between two companies in the same industry than it does between industries.
Outsourcing does force documentation, and that is a genuine benefit that people underrate. It is also an expensive way to buy documentation. If the honest driver is that nobody knows how the process works, the cheaper intervention is to write it down first, which also puts you in a far better position to negotiate.
Where master data quality is the underlying problem, a provider will surface it rather than fix it, because surfacing is in scope and fixing is not. Expect the ticket count to rise in month three and stay high, and settle the ownership question before signing rather than after.
The limit that matters most is reversibility. Bringing planning back in house costs more than sending it out, takes longer, and the cost is highest exactly when you most want to do it, because the trigger for reversal is usually that the arrangement is failing during a period when the business needs planning to work. Barthelemy counted this as one of his four hidden costs for good reason. Ask what a twelve month exit looks like in practice, in writing, and treat a vague answer as information about the deal.
Take one planner's last full week, split the hours across the six layers, and mark every activity that has a written procedure somebody outside the building could follow; the unmarked hours are the ones no contract will move.
More on this in organisation on the DemandLabs platform.
