The research behind Upkeep

Vaskehjelp is the intermediary: cleaners are self-employed, hold their own organisation number and HMS card, and set their own rate above the tariff minimum, while the platform handles matching, payment, VAT and insurance. Freska and CityMaid employ their cleaners directly, with sick pay, holiday and insurance attached — the difference is visible in the register, where Vaskehjelp is filed under intermediation services and the other two under building cleaning.

These are driftsresultat — operating profit or loss, i.e. EBIT. Norway’s free public accounts data does not break out depreciation, so no EBITDA figure is quoted here. Vaskehjelp’s book equity remained positive at NOK 2.78m despite the loss; Freska’s was marginally negative at year end despite the profit, its accumulated losses not yet cleared.

The pure intermediary is the one losing money. The two that employ their cleaners are the two in profit, and the larger of them has been doing it since 1987. That is the opposite of the usual marketplace story, and any plan that routes around employment has to explain why it expects a different result from the one company in this table that tried.

OSCAR advertises more than 200 fixed-price home services across 12+ categories with a 30-minute arrival option and a 15-day guarantee, operating in 33+ cities across Portugal, Spain and the UK — funded by a EUR 6M pre-Series A in May 2025 on top of a EUR 6M seed the year before. TIDY markets itself as “Your AI Property Manager” and says it “maintains a digital twin of every property” — company copy, not a technical disclosure, and evidence that the property-model idea is already claimed. Turno advertises 126,000+ vetted short-term-rental cleaners. Taskrabbit, owned by IKEA’s Ingka Group since 2017, runs in eight countries and 240+ IKEA stores — distribution attached to the moment a product creates a service need. Norway is not one of the eight.

The working model subtracted an ex-VAT cost stack from a VAT-inclusive price. A platform that sets the price, assigns the operator, defines the method and guarantees the outcome is acting as principal rather than agent, so it charges and remits VAT on the whole ticket. Once NOK 210 goes to the state, the NOK 300 gross becomes NOK 90, and the NOK 200 contribution becomes minus ten. Volume makes it worse rather than better.

To actually earn NOK 300 on the job, the ticket has to be NOK 1,312.50 including VAT — NOK 525 per productive hour. That is not a discount to the Norwegian market. That is the Norwegian market. Correcting the arithmetic deletes the price advantage that was the entire consumer proposition, which means the proposition has to be something other than price, or there is no proposition.

The same mistake sat in the headline number, and it is visible as a coincidence. Five per cent of Norway’s 2,762,504 dwellings buying 26 services a year at NOK 1,050 gives about NOK 3.77bn of gross volume, and a 20% take on that is NOK 754m. The VAT inside the same NOK 3.77bn is also NOK 754m — exactly, because a fifth of a VAT-inclusive price is the VAT. A headline revenue figure that equals the tax bill is a reliable sign the model was built gross. It is not published here as a target.

Uber One costs about USD 9.99 a month, reached 46 million members by the end of 2025, and buys $0 delivery fees over a basket threshold, member pricing, credits back on rides, and priority access to higher-rated drivers. Uber has said members spend several times what non-members spend and retain materially better, and that a new member starts out loss-making and pays back across their lifetime. The mechanism is precise: subsidise the customer at the moment they feel the charge — the fee, the surge — and earn it back in frequency and loyalty.

A delivery fee is a small charge on top of a basket, so waiving it costs the platform little. In cleaning, roughly 89% of the ticket is somebody’s wage. There is no fee sitting on top to hand back — the platform’s entire gross margin on the illustrative job is NOK 90, and it is already negative after costs. Any perk denominated in free labour is being paid for out of a margin that does not exist. So the transferable parts of Uber One here are the ones that consume no hours: priority when supply is tight, guaranteed continuity, free rescheduling, a covered re-clean, and the property record itself.

The property model was framed as a pricing asset — predict the time, quote a fixed price. But predicting time does not make money. Reducing it does, and a subscription is the only structure that allows either lever to exist. Continuity needs the same person returning to the same home; density needs control of when the visit happens. A one-off booking market gives you neither.

Operator pay held constant at NOK 300 per productive hour in every row. Visit-length reduction is an assumption, not a measurement — see the clutter problem below, which is the same question asked from the other side.

The two levers pay different parties, and that is the whole point. Continuity is the only thing that moves the platform’s margin — a cleaner who knows the house does the same job in less time. Density is the only thing that moves the operator’s income — less unpaid travel means more billable hours in the same day. Continuity alone slightly reduces what an operator earns per day; density alone leaves the platform exactly where it started. Only together is it a trade both sides should accept, and only a subscription makes both possible at once.

Membership revenue carries no labour, so it is close to pure margin after VAT — which makes the design question sharp. A membership priced at roughly NOK 249 a month yields about NOK 2,390 net a year. One included clean costs about NOK 850 all-in. So the membership funds two or three free cleans a year at the absolute limit, and one is the only responsible headline. Any offer implying meaningfully free cleaning is writing cheques against a wage bill.

The bigger trap is subtler. If the membership is used to discount the visits — take NOK 249 at the door and give it back in the per-clean price — the whole exercise nets about fifty kroner a year and is pure theatre. It only works if the membership sells something genuinely additional: priority, continuity, the guarantee, the record. That is the difference between Costco and a loyalty card, and it is the single most important design decision in this section.

A subscription over a negative unit is a direct debit for losses — retention makes it worse, not better, and frequency makes it worse faster. The order of operations is not negotiable: the unit has to clear first, and the levers above are the only route to that which does not involve paying somebody less. Two further constraints follow from Norwegian consumer law: a fourteen-day right of withdrawal applies, and cancellation is set to have to be as easy as signing up. So there is no lock-in to hide behind — the membership is retained on value or not at all.

Same-day booking is the part of the Uber comparison most likely to be over-read. Norway is a low-density, high-wage market, and the ten- and fifteen-minute fulfilment models funded in India in 2026 do not transfer to it — a 2.5-hour job cannot be summoned like a car, because the supply pool is thin and the unit of work is long.

What does work is the inverse framing. Recurring subscription visits are the base load that puts operators in a neighbourhood at a known time; on-demand requests are the overflow that fills the gaps around them, priced higher because they are scheduled against whatever capacity is left. That is also why member priority is the strongest perk in the set: it costs the platform nothing to give members first call on scarce same-day capacity, and it is worth the most at exactly the moment the customer feels the shortage. Without the subscription base there is no route for on-demand to sit inside, which is the real reason the two have to launch together.

Sweden’s RUT deduction refunds 50% of the labour cost of household work, up to SEK 75,000 per person per year in 2026. Norway has no equivalent. The only comparable instrument is the tax-free threshold for paid work in a private home — NOK 6,000 per person per year, above which the payment must be reported, and above NOK 60,000 in total the household itself becomes an employer with employer’s national insurance to pay. Twenty-six visits at around NOK 1,050 is roughly NOK 27,300 a year: far past the threshold, with no relief on any of it.

Set that against what software actually saves. Travel might be a fifth of an operator’s day in a dense cluster, and excellent routing might take it to an eighth — a couple of percentage points of job cost, call it fifteen kroner. Coordination savings are real but they are the platform’s own overhead, not the operator’s. So the efficiency argument brings about NOK 15 to a NOK 210 problem. This is a large part of why Sweden has a big legal household-services market and Norway does not, and no amount of routing closes it.

There is an obvious escape. Businesses reclaim input VAT and consumers cannot, so employers buying a staff benefit, insurers buying a claims service, property managers and short-term-rental operators all deduct the 25% that a household simply pays. On the arithmetic alone, a business-to-business wedge is the easier one.

Upkeep is not going that way, and the decision is a product decision rather than an oversight. Upkeep is for the person who owns the home or the apartment — someone booking a clean for the place they live in. Routing around that to sell to property managers would be a different company with a different customer, and the consumer product would never get built. So the tax stays in the price, and the price has to carry it honestly. Recorded here because it is the strongest counter-argument the research produced, and it should be visible rather than quietly dropped.

The stated principle was that price comes down through software and routing, never through underpayment. Check it: NOK 750 for 2.5 productive hours is NOK 300 an hour, but add half an hour of travel and it is NOK 250 door-to-door — and from that a self-employed operator funds their own holiday pay, pension, insurance, equipment, admin and unbilled travel. An employed cleaner at the statutory floor costs roughly NOK 319 an hour once employer’s national insurance, holiday pay and pension are loaded on, and that covers every paid hour rather than only the productive ones. So the model’s price advantage over an employer comes substantially from moving employment cost onto the operator. The principle does not survive its own numbers, and it has to be fixed at the price, not in the wording.

Continuous recording inside a house captures children, documents, screens, medication, intimate spaces and visitors who never agreed to anything — and the moment the operator is, or might be, an employee, it is also workplace surveillance. Norwegian camera monitoring at work requires an objective justification that is not a disproportionate burden on the worker, discussion with employee representatives before it is introduced, advance information about purpose and duration, and periodic review. That is the correct answer for a reason beyond compliance: a service whose trust model is a camera has not solved trust.

The proportionate version is duller and works: geofenced arrival and departure, one-time or smart-lock access codes, room-by-room checklist completion, before-and-after images of agreed surfaces only, on-device analysis with prompt deletion of raw media, customer confirmation, structured exception reports, and fraud detection from patterns rather than from continuous private video.

Having reasoned carefully about cameras, the analysis then skipped the same reasoning for the more intrusive artefact. A property model is a persistent, structured, queryable map of the inside of a private dwelling — household composition, children’s rooms, mobility aids, valuables, which door opens with which code — held by a third party and distributed to rotating operators. That is systematic large-scale processing using new technology, which puts a data protection impact assessment before the first capture rather than after the first incident. And the value side is unmeasured: if a six-question form gets close to the scan’s accuracy, the scan takes the largest privacy exposure in the design for the smallest gain.

Everything above rests on one claim: that a property model predicts how long a job takes accurately enough to sell a fixed price. But the most likely source of variation is not floor area or surface type — it is how much stuff is on the floor that day and whether anyone tidied before the operator arrived. Clutter is not in a scan, it is not a property of the property, and it changes between two visits to the same house a fortnight apart.

If clutter dominates, the entire architecture is aimed at the wrong variable, and the product is a photograph taken on arrival. That is a cheap thing to find out and an expensive thing to assume.

A fixed price averages across homes, and the error is one-sided. Customers whose homes run long keep the subscription; customers whose homes run short work out that they are subsidising it and leave. Adverse selection is not a rounding error in this model — it is the failure mode, and a median accuracy target hides it, because the median is exactly the statistic that ignores the tail doing the damage.

Scoping twenty or thirty homes once each and comparing estimate to actual is a cross-sectional study. It measures variation between properties — the part nobody doubted — and cannot say anything about variation within one property between visits, which is the number the business depends on. The test has to be repeated measures: the same homes across several consecutive visits, an arrival clutter rating recorded before work starts, the same operator on most visits and a different one crossed in on some, and the scan tested head-to-head against a short questionnaire to find out whether it earns its place at all.

Property data does not improve time estimates after repeated visits. Customers turn out to prefer negotiating hours to buying an outcome. Repeat jobs stay contribution-negative after routing and support are optimised. Operator earnings are not competitive once travel and admin are counted. Privacy requirements make the scan unacceptable to a large enough share of customers. Or the legal model cannot scale without an ordinary cleaning employer’s economics — in which case the honest answer is that the two profitable companies in the table above already found it.

Before any of this: one property’s register, assembled from paper that already exists. It costs nothing, it is the test this page has promised since it was written, and it happens to answer the scan question for free — if a usable register can be built from the sale condition report and the boligmappe, that is a strong hint about how much a LiDAR sweep was ever adding.

The test is whether a register for one property can be assembled from paper that already exists — the sale condition report, the electrical inspection, the boligmappe — without paying a surveyor to walk the building. If it can, the cadence and the log are software problems. If it needs a professional visit, this is a services business wearing software, and we would rather learn that with a notebook than a codebase. The test has not been run.