The Last Mile Is a Human Being

The question behind the question

Ask any answer engine what to do about a lump you found last night. It will tell you, accurately and instantly, that most lumps are benign, that you should see a doctor, and roughly what a doctor will look for.

Now notice what you actually wanted. You wanted a doctor.

This is the shape of almost every important question a person asks a machine. The machine gives a good answer. The person still has an unanswered question, because what they needed was not information. It was someone who would be responsible for the outcome.

We have spent thirty years making the answer part extraordinary and almost no time on the part where a human closes the loop. The gap between a correct answer and a solved problem is, in nearly every case, one person. Finding that person is still done the way it was done in 1994: ask a friend, or take your chances.

Why the last mile resists automation

It is tempting to think the human step is a temporary inefficiency, something a better model eventually absorbs. We do not think so, and the reason is not sentimental.

Three things a person supplies that a model structurally cannot.

Authority that is answerable. An adviser who tells you to roll over a pension has a licence that can be revoked. A doctor who is wrong can be sued. That exposure is not a bug in the system, it is what makes the advice worth acting on. A model has no licence to lose. It can be right far more often than a human and still not be responsible, and responsibility is the thing being purchased.

Judgement about the case in front of them, not the distribution. A model reasons over what is typical. A good professional notices that you are not typical, usually from something you did not think to mention. That noticing is the entire value of the visit.

The willingness to act. Answers do not file appeals, negotiate with a carrier, or sit with a family. Somebody has to do the thing.

None of that argues against the machine. The machine should absolutely answer, and should answer better every year. The argument is about where the work ends. It ends at a person.

What we are building

A recommendations engine that connects people who have problems to people who solve them.

Not a listing. A listing is a phone book that got a design refresh. What we mean is a system that takes the question you actually asked, understands enough about you to know which human would be right, and makes that introduction with your consent and on your terms.

The domains are the ordinary shape of a life. Money. Health. Fitness and well-being. Relationships. The parts of life people enjoy: how they shop, how they look, where they travel and what they want that trip to feel like. Providing for a family, while you are alive and after you are gone. And giving back, which is not a footnote. A great deal of human happiness comes from helping other people, and someone looking for the right way to give is asking a matching question exactly as real as someone looking for a tax adviser.

The order matters and we will say it plainly. Connections first. Then communication. Then commerce. Communication exists so the connection can proceed. Commerce exists so it can be completed. A product that opens with a transaction has the order backwards, and you can feel it as a user, because it is trying to sell you something before it has understood you.

The hard part is not search

Search is the easy half and we already have most of it. We hold public regulatory records for hundreds of thousands of licensed professionals across insurance, financial advice and healthcare. Given a postcode and a profession, returning the nearest qualified humans is a solved engineering problem.

The hard part is the second question: of the forty people who are qualified, which one is right for this person?

That is a matching problem over things we mostly do not know, and this is where an honest company has to be careful. Location and proximity we know. Speciality and credential we know. Stated preference we know, because the person told us. Everything past that gets speculative fast, and the speculation is where this category has historically gone wrong.

Personality and chemistry are real. Two competent advisers can be right and wrong for the same person, and everyone who has changed doctors knows it. We would like to model this. We are also aware that a personality model built on inference rather than on what someone actually said is a machine for encoding prejudice, and that a chemistry prediction with no feedback loop is a horoscope with a database behind it.

So: we will match on what people tell us, we will earn the rest from real outcomes, and we will not ship a psychometric claim we cannot substantiate. When we do not know, we would rather say so than guess in a confident font.

How it has to be built

The engineering shape follows from the problem rather than from fashion.

Candidate generation and ranking are separate stages. Generation is a geospatial and credential filter over an index that changes slowly and can be precomputed. Ranking is per person, cheap, and must degrade honestly. Conflating the two produces a system that is neither fast nor explainable.

Proximity is indexed, not computed per request. Distance between a person and forty thousand professionals is not a thing you calculate at request time and it never was.

Degradation is designed rather than discovered. There are places where we have no inventory in a given profession, and there are whole domains where we have none at all. The correct behaviour is to say so. A recommendation engine that always returns something is not confident, it is broken, and the user finds out the expensive way.

Every match writes a receipt. The person can see who was recommended, on what basis, and what was disclosed to whom. This is not a compliance feature bolted on afterwards. It is the only reason to trust the recommendation at all.

Serve people without exposing them

A thing we have had to learn: the information that would most help us help you is often the information you are most reluctant to hand over.

A debt. A diagnosis. A size. A dependency. A relationship you are trying to repair. People are judged for these, and the fear of being judged is exactly why they withhold the detail that would let anybody actually help. Any system that demands full disclosure before it will be useful has already lost the people who need it most.

So the rule is to serve without exposing. Disclose the minimum that accomplishes the task, to precisely the one party who needs it, for exactly as long as the task takes. Prefer a band to a number and a region to an address. Never let one party's grant become another party's inference. And never surface something intimate in a view that somebody could read over a shoulder.

If a design would work but would make a person feel judged, it is the wrong design. That sentence is in our engineering context file, not our marketing.

Why this is worth a decade

Here is the part that made us want to build it.

The cost of not finding the right human is enormous and almost entirely invisible. It is the family that did not know a benefits office could help. The small business that took the wrong insurance because the right broker was four miles away and unfindable. The person who managed a condition alone for two years before meeting the specialist who solved it in one appointment.

None of that appears in any productivity statistic. It shows up as a slightly worse life, several million times over.

A region where people reliably reach the right expert quickly is measurably better off than one where they do not. Better health outcomes, better financial decisions, more small businesses that survive their third year, more charitable capacity actually deployed. We think that is a real contribution to how a place works, and we think the mechanism is not a smarter answer engine. It is a shorter path from a question to a person who can be responsible for the answer.

That path is what we are building. It is going to take a long time and we would rather do it properly.

What is true today

We hold public regulatory data on hundreds of thousands of licensed professionals and it is searchable by proximity right now. Every listing is claimable and removable, free, without an account, and a removal survives our next rebuild, which is a harder engineering promise than it sounds and one we recently had to go and fix.

The matching described here is in design. The personality and chemistry layer is not built, and we have said above what would have to be true before we would ship it.

We would rather publish the plan and be held to it.

If you build recommendation systems, geospatial infrastructure, or consent architecture, or if you are one of the experts this is meant to reach, we would like to hear from you.