Someone picks up

What the co-operative actually does for you on an ordinary Tuesday.

Before you read this

The rest of this series is about governance, ownership and infrastructure. All of it is scaffolding. This part is the thing itself, and four questions decide whether any of it is worth doing.

What happens when I need help? — Not in theory. At twenty past four, when the thing that always works has stopped working and you have a client waiting.
Why would the answer be any good? — Every organisation has been told that AI will help. Most have found that it helps with things they did not need help with.

Before you read this (cont.)

Who is on the other end? — A script reader in another time zone, or someone who knows your situation.
And what does it do that a subscription does not? — If the answer is nothing, buy the subscription.

An ordinary Tuesday

A three-partner accountancy practice. A question comes up about how a particular client entity should be treated for a filing that only arises every few years. Nobody in the office has done one since 2021.

First: they ask their own system — A model running on the practice’s own machine, which has read the practice’s own files — every previous filing, every note, the correspondence, the way this practice does things. It answers in seconds, cites the two occasions this came up before, and shows what was done.

An ordinary Tuesday (cont.)

Most days that is the end of it. This is the part that does the volume, and it is not remarkable technology — it is a modest model with access to the right material, which is a different thing from a large model with access to none of it.
Nothing left the building to make that happen.
Second: it is not sure, so it says so — The 2021 case is not quite the same. The system does not guess. It marks the question as one it cannot settle and offers to escalate.

An ordinary Tuesday (cont.)

Third: someone picks up — A video call to a person in the co-operative’s service desk. Not a script reader. Someone who has the same knowledge base in front of them plus everything the co-operative’s members have collectively resolved about this kind of filing — with the practice’s own records visible only if the practice chooses to share them for this call, and not otherwise.
They work it out together in eleven minutes.

An ordinary Tuesday (cont.)

Fourth: the answer stays — The resolution is written back into the practice’s own records, signed and dated. The next time this arises, the practice’s own system knows. And with the practice’s explicit permission, the pattern — not the client, not the file, the shape of the problem and how it was resolved — joins what the co-operative collectively knows, so that the next practice does not start from nothing.
That is the whole service. Everything else in this series exists to make those four steps safe.

Three tiers, and the economics live in the second one

FIG-14 A question answered by the member's own system, escalating to a person only by exception, and the answer written back so it is known next time.

Three tiers, and the economics live in the second one

Tier one — your own system, on your own hardware — Routine questions, drafting, summarising, finding things, triage, filling forms, checking work against your own precedents. It knows your organisation because it has read your organisation. Nothing leaves.
Tier two — a person, by exception — When the local system is unsure, wrong, or out of its depth. The person is not a call centre operator working from a decision tree; they have the same material the member has, plus the co-operative’s accumulated resolutions.

Three tiers, and the economics live in the second one (cont.)

Tier three — advice, not support — The same relationship extended: how should we structure this, what are we not seeing, what have comparable organisations done. Delivered by people with genuine domain expertise, often member-contributed and member-accredited rather than employed, and coordinated by the co-operative rather than staffed by it.
The bet is the ratio between tiers one and two, and it is a bet — If the local system handles most routine work and people are needed by exception, the cost per member is affordable. If it handles less than assumed, the model inverts and the co-operative is a very expensive help desk.

Three tiers, and the economics live in the second one (cont.)

We are not going to dress that up. It is the largest single unknown in the whole design, it is measurable, and it should be published as a number from the first month.

Part C — The blueprint

Why the answer is better, and it is not the model

FIG-04 Federation is an enrichment layer over complete clusters, never a dependency of them. A cluster can be severed and keep operating.

Why the answer is better, and it is not the model

The tempting claim is that this is better AI. It is not. The models are ordinary and often smaller than what you would get from a subscription.

What is different is what the model has read.
A general assistant knows a great deal about accountancy and nothing about your practice. It cannot tell you what you did in 2021, because it has never seen it. It will give you a competent generic answer to a question whose answer was specific.

Why the answer is better, and it is not the model (cont.)

A model that has read your own records gives you your own answer. For most of the work most organisations actually do, that is the difference between useful and merely impressive — and it requires the records to be somewhere the model can read them, which is exactly why the ownership arrangements in this series matter. You cannot have a system that knows your organisation without giving something access to your organisation. The only question is who, on what terms, and whether you can withdraw it.

And someone stays in the loop

There is a failure mode in every escalation-based service and it is worth naming because it arrives slowly.

If people only ever see the exceptions, they lose the sense of what normal looks like. The exception path degrades exactly as the routine path improves, and nobody notices until a bad answer goes through unchallenged.

And someone stays in the loop (cont.)

So a defined proportion of routine work is checked by a person regardless of how well the system is performing. The proportion is published and it does not fall as accuracy rises — that would be switching off the alarm because it has stopped sounding. It is a permanent, deliberate cost, and it appears in the cost model as one.

Who does the work

Employed people at the desk, in the co-operative, at ordinary pay, without a large spread between them and anyone else.
Local subcontractors for hardware — installation and repair by people who already service equipment in that district, because a national field-service operation is how organisations like this die.
And members themselves — Contribution earns membership on the same terms as payment: accredited support hours, domain expertise, a resolved procedure another member uses. A member organisation with deep expertise in one area and no spare cash can join by supplying what it has.

Who does the work (cont.)

That last one is not charity, and it is not only an economics fix. It is what stops the knowledge base being something taken from members. Where members are customers, contributing what they know is extraction with a friendlier name. Where they own the thing and contribution earns entitlement, it is investment in an asset they hold a share of.

What this does not do

It does not run your existing software for you — Some of what you already use will work against a local system. Some will not, because it needs its own offshore inference and cannot be pointed elsewhere. Finding out which is which is genuinely useful — and it is a finding, not a fix. What to do about the ones that refuse is not settled, and we are not going to pretend otherwise.
It is not a frontier model — For the hardest reasoning problems a large remote model is better, and the design assumes some work escalates outward. What it does not assume is that your routine work depends on that.

What this does not do (cont.)

It is not free, and it is not the cheapest option — Holding capability locally and keeping people available costs more than a subscription that keeps nothing and answers nobody.
And it is not fast to build — A cluster is a year of work before anyone is answering a call.

How you would know this is wrong

Two things, both measurable in the first months.

1. The tier-one ratio — If the share of work a local system handles unaided is materially below what the cost model assumes, the service economics fail and no amount of governance rescues them.
2. Whether members value a person at all — Perhaps a good local system plus a forum is enough, and the desk is an expensive answer to a problem that has largely gone away. If the escalation rate is far lower than modelled, that is not a triumph — it is a signal that tier two should be smaller than planned, or member-contributed rather than employed.

How you would know this is wrong (cont.)

Measure both, publish both, and change the design if they say to.
Published under CC BY 4.0. The organisation in the example is illustrative. The arrangement is not.