Three Gears for Entering a New Market

Almost every capability programme I am asked to review arrives with its ambition level already chosen, and it is almost never the modest one. The paper commits to building the thing that does not exist yet, because the version that merely improves a decision the firm already makes does not read like a strategy. Some months later that programme has a working model, an unresolved argument about who operates it, and nothing the business can sell.

I keep returning to a market-entry appendix I worked on for a European composite insurance group, because it set out the alternative across three consecutive pages and then made, almost in passing, the argument I now use in every readiness conversation.

The subject was connected sensing in the home, and the drivers were ordinary claims economics. Installed home sensing had risen fourfold over five years. Retention was poor: 36 percent of customers intending to switch their home cover within the coming twelve months, against an industry average net promoter score of minus six. The appendix was blunt about why: nothing in the product the customer valued, no loyalty earned, so the renewal came down to price. A survey of who would own the connected-home ecosystem ranked seven candidate groups, and insurers came last, named by 6 percent of respondents against 67 percent for the consumer device company at the top.

That last number is the one people remember. It is not the useful one. The useful pages came next, where the appendix stopped describing the market and started describing routes into it.

Three routes, described rather than named

The first route pointed new data at a decision the insurer already made every day. Keep the household policy as it is, take a feed from sensors in the home, and let it either sharpen the price or head the loss off before it becomes a claim. Occupancy and security behaviour as pricing inputs. Whether the smoke alarm works, whether there is water where there should not be, as mitigation inputs. The hardware came through a partner who already made it, and the worked example was a customer installing a detector or moisture sensor and getting a lower premium for it.

The second route turned the sensor from a pricing input into a service sold alongside the policy: catch the problem early, stop it where stopping is possible, send somebody where it is not, all bundled with standard cover. The distinguishing detail is not the device. It is where the alert goes, to the customer and to a monitoring operation that dispatches assistance. Two European propositions already worked this way, one pairing a telecoms operator's home package with an insurer's emergency assistance service, the other a packaged telematics device from a bank-owned insurer that watched three perils and put a manned centre behind every alarm.

The third route covered the household and the people in it: information shared with care providers, remote monitoring so an intervention could be triggered when the pattern warranted one, and tools letting people manage their own conditions and lean less on formal care. It leaned on partners and on hardware that already existed, while at the same time asking the firm to sell something it had never sold, plus three families of monitoring device the insurer had never touched. Its market-example column carried one line: no existing insurance proposition.

The ratings do not move together, and that is the finding

If the five headings rose in step, the appendix would be a price list and you could pick the ambition you could afford. They do not.

From the first route to the second, the cost and internal-resource headings did not move at all. What moved was the offer itself and the work of taking it to market. From the second to the third, cost and resource finally shifted up a band while the offer and the capability heading both went to the top.

Read as a snapshot, that pattern says something reassuring and wrong: ambition is cheap, capability is the tax. Read as a sequence, it says something more useful. The second route's cost rating stays low only because the first route has already paid for the ingestion, the integration into pricing and policy systems, and the partner supplying the data. The third route's cost stays merely middling only because the second has already built the thing that decides what to do when an alarm fires and dispatches somebody to do it.

Take the first route out and the second route's rating is fiction. Take the second out and the third is not a proposition, just a slide.

A ladder drawn as three panels on one page will be read as a menu. It is not a menu. The second route's rating is a lie unless the first route's plumbing already exists, and the third route's rating is a lie twice over.

What this looks like on a machine learning programme

Change the vocabulary and the ladder is the shape of every analytics capability build I see.

The first gear points a model at a decision the business already makes, using data it can already get. Fraud scoring on a claim a handler was going to open anyway. Nothing about the product changes, and the whole difficulty is technical: getting the feed in, giving each feature a lineage anyone can trace, a feature store that means the same thing in training as in production, and a deployment path that makes a retrain an ordinary event rather than a project. That work looks cheap on a business case and is not, which is exactly the shape of a rating low everywhere except capability.

The second gear wraps a service around the model and sells it. That needs everything the first gear built, plus an operating function that did not exist: a queue, an owner, a service level, straight-through processing for the cases the model can settle alone and a routed human path for the rest. Process mining the existing flow is the cheapest way to size that first category, and it is usually smaller than the business case assumed. This is where cost ratings quietly lie, in that appendix and in almost every plan I read. A monitored service that responds is a staffed function, and staffed functions do not cost what a pilot costs.

The third gear sells something the firm has never sold, to a market with no comparator. It needs both lower gears working, plus governance for data classes the organisation has never held: in the appendix's case, health and care data about vulnerable people.

Programmes fail at gear three not because it is wrong but because it was funded while gear one was still a slide. The result is capability nobody can sell: a model with no route into an operating decision, or an operating decision with no lineage anyone will sign off.

The claims arithmetic underneath all three

The appendix started at claims cost rather than technology because all three routes are a bet that prevention beats indemnity, and the market moves it cited were priced. One insurer offered a discount band of 7.5 to 12.5 percent on contents cover to customers taking a partnered alarm system. Another gave a quarter off the home premium and threw in the hardware: a leak sensor with a touchpad, three window and door sensors, a motion detector. Real commercial terms, setting up a question none of those players published an answer to. Does the avoided claim beat the discount plus the subsidised hardware?

That question is answerable, and it is a first-gear question. It needs claims history joined to whatever the sensing layer reports, a defined counterfactual, and enough clean lineage that a pricing actuary accepts the result. A firm that cannot answer it has no business funding the second gear, whose economics assume the answer came back positive.

4x

Rise in installed home sensing solutions over five years

36%

Customers intending to switch home cover in the coming twelve months

-6

Industry average net promoter score

6%

Respondents naming insurers as likely owners of the resulting ecosystem, last of seven

What we ask before funding a gear

Four questions open a RealAI Platform readiness pass, and none of them is about the model.

What decision does this change, and who makes that decision today. If nobody makes it yet, the programme is at gear three whatever the paper calls it.

What already runs. Not what is planned, not what sits in a backlog. If the ingestion, lineage and deployment path from the gear below are not in production, this gear's rating was written against a stack that does not exist.

Who operates it on a Tuesday. Every route past the first implies an operating function, and its cost belongs on the same page as the build cost, not in a later phase.

What would have to be true for us to stop. A gear you cannot abandon is a commitment, and commitments made before the plumbing works are how a programme ends up defending a model instead of using one.

None of this argues against ambition. The third route was the most interesting thing in that document, and the absence of any comparator was an argument for building it rather than against. It sat third for a reason, and the reason was invisible on a page showing all three at once.

Tooling has moved a long way since. Pipelines are close to assembly work, deployment is largely solved in a decent stack, and the first careful language model pilots I am seeing stand up in days rather than quarters. Gear one is cheaper than it has ever been. That does not make the gears above it any less dependent on it, and the temptation to skip straight to the interesting one has gone up rather than down.

Detail is as recorded in a market-entry appendix written for a European composite insurance group: its three ordered routes, their ratings, the statistics it cited and the priced market moves in its notes. That appendix produced routes and recommendations, not delivered results, and called its own ratings illustrative. Reading the routes as a ladder rather than a menu, and the middle route's cost rating as optimistic, is ours.