Where the work lives
Most organisations think the AI work they are doing is the thing the assistant produced. The part that decides whether it lasts is where that work ends up.
Thanks for coming back for issue nine. This fortnight I sat in on a hosting handover for an automation build, a conversation about analytics across a growing subscription business, and a coaching session where a dashboard ran out of room. They look like three different problems. Each came down to the same question: when the AI does a piece of work, where does that work live afterwards?
Last issue was about the org chart nobody drew: who each assistant is for and who looks after it. This one sits underneath it. Even a well-defined assistant can leave its output in the wrong place.
The pattern: the work lives in the tool
The way it starts makes sense. Someone opens a chat, gets a good result and keeps going. A project fills up with files. An artifact turns into a working app. A few weeks later there is something useful that the whole team would like to use.
Then three things happen, usually in this order.
The first is a limit. Projects have a capacity, and one coaching client’s dashboard got to 98% of it. Every question was re-reading everything, so it was slow and it was using up the weekly token allowance. The second is a sharing problem. An artifact can only be opened by someone with an account on that platform, so the people who need it cannot get to it. The third is an ownership problem. The thing was connected through one person’s profile, and nobody else knows which account it sits under or where the data went.
None of that is the fault of the tool. The tool did what it was asked. The work was just never given an address in the places the business already keeps its records: the inbox, the shared drive, the CRM, the data warehouse.
The address test
The framing I keep coming back to is that every piece of AI-built work needs an address, and an address answers three questions.
Where does it write to? Into the system the business already treats as the record, or into a copy that only the tool can see.
Who can change it? Someone inside the business, or only the person who built it, or only whoever built it for you.
What happens when the person leaves? The work carries on, or it sits on a personal account until it quietly stops.
When those are answered, the work becomes part of how the business runs. When they are not, you have useful output that nobody can hand over, audit or change.
Option one: everyone builds their own
I heard the first version from a finance leader at a growing subscription business. The company has capable people who are keen to use AI. Her description of where it had got to was that everyone is “effectively doing their own analytics”, and her read was that nobody was using it efficiently yet. The people building assistants were doing it “just for their use case”, so there was already duplication across them. She also said she was not yet confident anyone was solving the commercial problems with it.
Nobody there was doing the wrong thing. The business had no shared place for the work to land, so each person made their own. The result was a growing pile of personal tools sitting next to a data warehouse the business already trusted.
The same pattern shows up in the smaller version. A project that holds everything slows down as it fills, an artifact stays inside the platform it was built in, and a personal connector holds the keys. The fix in that coaching session was to move the working parts into a code repository, a hosted database and a hosted front end, and to link the hosting to the company’s own account instead of a personal profile. The project dropped back to between 15 and 30 percent of its capacity. It also meant the person who built it was no longer the only one who knew where everything was.
Option two: build it into the record from day one
The clearest example of the other way was an accreditation team at a professional body.
Their automation does its work in the places the team already works. It reads the shared inbox, drafts and sends the replies, and files each document into the right folder in their own Teams. When their operations lead logged in to test it, her actions showed up for the other users straight away. The audit log recorded which person had done what. Nobody had to learn a new place to look.
It is not finished, and I would rather say so. The code still sits in our (Friyay’s) secure environment while the move to the organisation’s own infrastructure waits on a conversation about Microsoft 365 access. Until that happens, a change to the wording of an automated email is a code update, not something the team can do themselves. The test also turned up a few things: filing a document added stray signature images to an existing folder, and once an application is pulled in, it cannot be deleted because the tool holds its own copy.
So this is a better answer, not a finished one. What makes it better is the starting point. The work lands in the inbox and the folders that already exist, with the users and the logging the team recognises, and the open items are all about moving it, not about finding it.
The proof of concept comes second
One more thing from the subscription business. Her instinct was to avoid a long, expensive programme, and the approach we talked through was a standalone proof of concept first, with the measure of success agreed upfront (for example, how many more fraudulent visits an automated assistant picks up). Integration with the core systems comes only after that has worked. That sounds like the opposite of “build it into the record from day one”, but it is the same discipline. Decide where the work will eventually live and who owns it before you scale it. Do not decide it after twenty people have built their own versions.
What it costs
The cost does not show up on an invoice, which is why it is easy to miss.
It shows up as work that has to be rebuilt once it outgrows its first home. It shows up as an assistant that cannot be shared with the person who needs it, and as access that nobody can switch off because nobody is sure where it was set up. It also shows up when a team asks for a small change and finds out that the change needs the person who built it.
An organisation that gives its AI work an address gets the opposite. The work belongs to the business, the changes happen in the business, and the people who built it can move on to the next thing.
Enjoy the read and see you in a fortnight.