How I would build a company today that needs no army of administrators.

Most companies do not hire administrative staff on purpose. They hire them because a gap was left between systems and somebody has to step across it every day, carrying the data in their hands. An order arrives by email, somebody retypes it into the stock system. A client changes address, somebody corrects it in three places and misses the fourth.

This is not a failure of people. It is a consequence of the order in which the company grew. Sales first, then a tool for sales, then stock, then accounting. Each department picked whatever suited it, and people filled the gaps in between. Human effort is, after all, the only integration that works immediately and needs no project.

My thesis: if we design the company around data and processes, most of that connective work never comes into existence. According to 2017 research by the McKinsey Global Institute, which analysed roughly 800 occupations in the United States, data collection has an automation potential of 64 per cent and data processing 69 per cent, while managing people sits at only 9 per cent. This is the technical potential of the technologies of the day, not a forecast of job losses. As a map of where a company spends its time, however, it holds up.

A company has five floors and eight layers

You can look at a company as a building with five floors: Brand, Experience, Systems, Automation, Intelligence. Brand states who we serve. Experience is everything the customer sees. Systems are the data and the platforms that hold the truth about the company. Automation is the rules that run without a person. Intelligence is the ability to understand what is happening.

The order is structural, not aesthetic. Every floor rests on the one below it. Automation on top of data that does not exist is simply faster chaos, and AI on top of a bad process only makes the mistake more expensive. These five floors break down into eight layers.

Layer 1, Brand: who we are and what makes us different

It sounds like marketing, yet it is an operational decision. If a company does not know who it serves, it cannot define good demand. And what cannot be defined cannot be recognised by a machine either, so a person has to assess it, enquiry by enquiry. Brand here is not a logo but a definition that turns, a few layers down, into a filtering rule.

Layer 2, Experience: the customer interface

The website, the app, the sales conversation, onboarding. The only layer the customer sees, and at the same time the point where data enters the company. A large share of administrative work is created right here, when the interface collects less than the process needs.

A form with a single message field looks friendly. It also means somebody has to pull the industry, the company size and the deadline out of free text and type them into the system. A form that asks three structured questions and fills in the rest by itself does not remove that work, it simply moves it from an employee to a machine. Interface design is therefore an economic decision, not an aesthetic one.

The Systems floor: layers 3, 4 and 5

Layer 3, data: the layer most often skipped

This is the layer companies skip most often. A company buys a CRM, then an online shop, then an invoicing tool, and each one creates its own idea of the customer. Three versions of the truth appear, along with one person who reconciles them. Single source of truth means that for every piece of data there is one owning system and the rest only read it. A data model is the written record of which entities the company recognises and how they connect.

Why this matters is shown by the Enterprise Integration Patterns catalogue. With six applications, direct point to point connections require 30 message translators, whereas a shared canonical data model needs only 12. The cost of connecting grows quadratically, and a shared model makes it linear. The translators a company never wrote exist anyway. They are simply executed by people, by hand.

Zhamak Dehghani describes how a centralised data platform usually ends up: disconnected source teams, consumers competing for a place in the backlog and an overstretched data team. She locates the cause in the structure, not in the people. It is a professional essay, not an empirical study, yet it supports the argument of this piece: connecting departments with people is an architectural decision.

Layer 4, core systems

CRM, ERP, commerce, operations. I tell clients the thing that occasionally costs me the job: buy off the shelf for as long as you can. Custom development only makes sense where the process is the source of your difference. Accounting, invoicing and standard CRM are solved problems, and ready made tools handle them more cheaply. Why more software on its own solves nothing is something I unpack in Your company does not need more software. It needs less chaos.

More important than the brand name is whether the tool can send and receive data, and in what mode: in a batch once a night, or continuously. With every new tool this leads to a single question: who owns the record and how are the keys matched. If nobody can answer that, you have just bought yourself a future job vacancy.

Layer 5, workflow: what happens next

Workflow is the answer to the question of what happens after a lead, an order, a complaint or an invoice. In many companies that answer lives only in the head of one experienced person. Workflow orchestration means the sequence of steps, the conditions and the dependencies are written down in one place and enforced by a machine, not by a coordinator.

Apache Airflow takes this principle all the way into code: a DAG is a model that encapsulates everything needed to execute a workflow, meaning the tasks, their dependencies and the schedule. You do not have to use Airflow, and for most companies the native workflows in a CRM are enough. What matters is the principle: a process should have one place where it is defined, and it should be able to say which step it is on. A company is, after all, a single product in which every bad process is technical debt, which I explore in the article on the company as one product.

Layer 6, Automation: rules and repetition

Only now does automation arrive. Not as a project, but as a consequence. On top of clean data and a documented process it is mostly configuration rather than development. I use the order Remove, Simplify, Automate, Augment. First remove the steps that exist only because it has always been done that way. Then simplify, for example by replacing three rounds of approval with one that carries a limit. Only then automate what remains, and finally augment the person with a tool where judgement is required. Companies tend to start at step three and set in concrete the work that should never have existed.

I calculate the economics through Cost of Friction: time times frequency times number of people times cost. A model calculation: a two minute retype of one data point, done by five people twenty times a day, comes to more than three hours of work every day. That is not a detail, that is a role nobody approved and nobody put in the budget.

The Intelligence floor: layers 7 and 8

Layer 7, AI: interpretation, not magic

AI can do four things that rules cannot: interpret unstructured text, classify, provide a language interface to data and assist with preparing documents. It can, for instance, pull the subject of a complaint out of an email.

What it cannot do is fix a badly designed process. Without one place holding the truth, a language interface will answer confidently from the wrong data and speed the error up. I write about that in AI will not fix a broken process. It will just do the wrong thing faster and cheaper.

Layer 8, management intelligence

Dashboards, reporting, alerts. A report requires somebody to look at it, whereas an alert arrives by itself when a value crosses a threshold. Companies tend to have reporting and no alerts, which is why a problem only surfaces at the next report. The minimum: time from enquiry to first contact, the share of orders handled without human intervention and the age of the oldest stuck case.

One lead step by step

Let us walk through a model lead step by step and ask where a person is genuinely needed.

1. A visitor arrives on the website. No person. The source, the campaign and the pages viewed are recorded automatically.
2. They fill in the form. No person, but with a design decision: every extra field lowers the number of submissions, every missing one creates work later.
3. Data enrichment. No person. Lead enrichment is the automatic completion of company details from external registers: industry, size, registered office. Typical work that somebody does in a browser today, even though it is entirely mechanical.
4. Writing to the CRM. No person and no duplicates. If the system cannot match a contact to an existing company, this is where future administration is born.
5. Scoring. No person. The HubSpot documentation describes lead scoring as assigning values based on record properties, meaning job title, company size or revenue, and on behaviour such as visiting the website or opening an email. The score is then usable in segments, workflows and reports. Configuration on top of a data model, not manual work.
6. Assignment to a salesperson. No person. The rule belongs in the system, not in a meeting.
7. First contact. With a person, and quickly. The authors of a study published in Harvard Business Review in 2011 audited 2,241 US companies with a test enquiry: 37 per cent replied within an hour and 23 per cent never replied at all. Across a sample of 1.25 million leads, companies that contacted a prospect within an hour were almost seven times more likely to qualify the lead than those that tried an hour later. Among the causes of slowness the authors list the practice of pulling leads from the CRM once a day instead of continuously. That is not laziness on the part of salespeople, that is process design.
8. Preparing the proposal. With a person, but assisted. Pricing and risk assessment are judgement. Assembling the document and formatting it are not.
9. Signature. No person on the company side. An electronic signature returns a signal to the process and triggers everything that follows.
10. Invoice. No person. If billing details are keyed in by hand again, layer 3 is not finished.
11. Onboarding. Mixed. Setting up the project, the access rights and the tasks is mechanical. The first conversation about expectations is human and cannot be replaced by a template.
12. Reporting. No person. If somebody is assembling a report by hand, the company is paying for incomplete layers 3 and 5.

Of the twelve steps, three need a person. And in each of them the value comes from their judgement, not their hands.

When a person should step into the process

The test goes like this: a person should step into the process when their judgement creates value. Not every time information has to be moved.

If you can describe that step as a precise set of instructions for a new colleague, you do not need a colleague. You need a rule.

Human in the loop does not mean improvisation. In the AWS Step Functions documentation, waiting for a human decision is a proper state of the process, with its own token and a time limit. The process pauses, waits for approval and continues when the token comes back. If it does not come back, the system reports it as an error. Compare that with an email nobody knows arrived.

The note that usually gets left out

None of this means administration will disappear. According to Cedefop, roughly 8.5 million people worked as office clerks in the EU in 2022, and the forecast expects a decline of about 720,000 jobs by 2035. A 2021 OECD study found no country level evidence of net job destruction, yet between 2012 and 2019 employment in the half of occupations at highest risk grew by 6 per cent against 18 per cent in the least exposed half.

My conclusion is an opinion, not a fact: this is not about redundancies, it is about whether a company grows by adding people to move information or people to make decisions. The first route has a ceiling, because moving information grows with the volume of business. The second does not. If somebody in your company spends their days shifting data around, the problem does not sit in their performance or in a missing tool, it sits in one of these eight layers. Finding that layer is cheaper than a solution bought before the problem has been named.

Key takeaways

• Before you buy another tool, write down which entities your company recognises and which system owns which piece of data. Without that, every integration is merely an agreement about who will do the retyping.
• Walk one process from start to finish and at every step write down whether a person is exercising judgement or moving information. The steps of the second kind are your list for the coming year.
• Calculate the Cost of Friction for at least one step. Most companies discover that their most expensive activity has neither an owner nor a line in the budget.
• Automate only once you have removed and simplified. An automated pointless step is more expensive, because nobody questions it any more.
• If you do not have capacity for everything, start with layer 3. Data cannot be caught up later without rebuilding everything above it.