15 November 2026
Paper, PDFs and copy-paste: collecting the data is the real bottleneck
People talk about models, agents, automation. Then you open the file and find a folder called “final final 3”, a scanned PDF that leans, a sheet where someone pasted two columns by hand, and an email that says “I’ll send the rest when I find it”. In a great deal of serious work, the bottleneck is not the calculation. It is the collection.
The data has to come in clean
I say this as someone who has spent time getting numbers into one place. At To Be, as a project manager, part of the job was making steps repeatable and having planning and reporting tools, on photovoltaic and agrivoltaic projects totalling over ten megawatts. If the data does not come in clean, the most elegant report tells an estimate. At Atlas, the benefit company I have led as CEO and co-founder since February 2025, the subject is the same on a different perimeter: energy diagnosis, carbon footprint, sustainability reporting. The company site is www.atlascarbonneutral.com. I am not writing the brochure here. I am describing the point where the work jams, because it is the same point I find when I design custom AI for another firm.
The folder, from January 2027
From January 2027 Atlas is preparing a service set up like this. The client receives a folder that is already structured, and a guide. Not a vague invitation to “send the documents”. A structure: what goes where, under which name, for which purpose. The documents serve three precise jobs: the energy diagnosis, the carbon footprint, the sustainability report. The client uploads there. An AI agent extracts the data from those documents. Atlas’s software prepares the working documents. A human professional validates before anything is treated as done.
Notice the four beats, because they are the same as the method I use outside Atlas as well. There is an initial listening: which documents are actually needed, not every document in the world. There is a map: the folder is the map, visible, dull, effective. There is a build: the agent that extracts and the software that prepares. There is a person who checks. If the folder is missing, the agent fishes in the pile. If the person is missing, the extraction becomes a fact only because a machine produced it.
Copy-paste is the symptom
Copy-paste is the symptom, not the trade. When someone moves a number from a PDF into a sheet, they are translating a document into a datum. It is humble work and very easy to get wrong: a row, a comma, the unit, the wrong version of the file. Doing it by hand a thousand times does not make it nobler. Having a machine do it without a folder and without a check only makes it faster. I do not want the speed of the wrong datum.
Imagine a firm — again: an example, not a client — that wants a sustainability report and sends twenty-three attachments in one email, with different subjects, some repeated, one locked by a password written in the previous email. Any “intelligent” agent put in front of that email will do one of two things: it will stop, or it will invent an order. Stopping is honest. Inventing an order is dangerous. The pre-structured folder exists so you never have to make that choice. The guide exists because nobody has time to guess what the meter file is called.
This is not a result I have measured on a list of clients, and I am not publishing one. It is the design of the service, said by the person who leads the company. Atlas’s figures, patents, and the measure-and-report paths live on the company site. I do not repeat them here to make them look like my personal AI cases. This site is about another trade: going into a firm, finding the point where data is collected badly, and building a tool for that point.
Copy-paste, outside Atlas, has the same face. A salesperson who retypes price-list lines by hand. An accounts desk that recopies the same records. A site — imagine again that construction firm of eight people, still hypothetical — where delivery notes become a table only if someone has a free afternoon. The solution is not identical to Atlas’s folder. The principle is: first the place where the document must sit, then the extraction, then a person.
A simple test
There is a test I use to see whether collection is the bottleneck. I ask: “If tomorrow the person who usually recopies is away, does the work stop?” If the answer is yes, you do not have a process. You have a patient person. AI can take the compulsory patience off them, not the judgement. The judgement stays, and in the Atlas service it is what the validating professional does, and in other companies it has to have a name.
Paper and PDFs are not the enemy. They are the material. The enemy is expecting a model to understand them at once, in random order, without a guide and without a check. From January 2027, at Atlas, the answer to that mess is a folder, an agent, software and a person. In the rest of my work the answer has the same shape, even when the documents are different. You start from how the data comes in. Not from how you wish it would come out.