From Excel to AI Agents

Image: AI-generated

What we did with spreadsheets is about to happen to agents — unless we change one thing.

Open the shared drive of any large company and count the spreadsheets that contain a single formula.

It will not take long. Most of them are tables with borders. Formatted lists. Meeting minutes with a grid, project plans coloured by hand, a column of names that someone sorts manually every Monday. Excel can calculate, pull data from half a dozen sources and run macros that do an afternoon's work in seconds. We used it for what Word does.

I bring this up because we are about to do the same with AI agents.

Look at where the attention goes. Every website, every feed, every newspaper is full of AI, and the examples are always the same: writing code, finding security holes, summarising research, drafting the email you did not want to write. All useful. All beside the point. The largest gain sits somewhere much duller — in the recurring business process nobody wants to own.

We all know these tasks. The report that has to be assembled from four systems. The list that needs cleaning before anyone can use it. The reconciliation that is not difficult, only long. We postpone them for a morning, sometimes for a week, and not because they are hard. Because they are boring, because it is unclear where to start, or because there are so many of them that choosing one feels like work in itself.

I am confident agents are the answer to exactly this kind of work. You explain a task once, and the agent does it every hour, every day, every week — whenever it is needed.

Here is the part that should embarrass us a little. Most of this work could have been automated years ago. The tools existed. What stopped us was the data: it sat in emails, documents, half-filled forms and the heads of the people involved. Before anything could be automated, someone had to structure it, and structuring it by hand cost more than simply doing the task. So we kept doing the task.

That obstacle is gone. Structuring unstructured data is precisely what these models are good at. The reason this work never got automated no longer applies.

You might think that teaching an agent a task is a project in itself. It is not. You describe what you want, the agent drafts, you correct it, it tries again. It takes time, and the first version will be wrong in places. Compare that with the time you spend doing the same task by hand, several times a week, for years.

The arithmetic is not close. Take an agent that goes through your calendar and your to-do list each morning, decides what belongs to today, plans the day and tells you when to move on. Setting it up takes half a day. If it saves you ten minutes a day — and I am sure your planning takes longer than that — it has paid for itself after 24 working days. Everything after that is gain, and that is before counting the reminders that stop a task from drifting into tomorrow.

But here is the catch. A powerful tool makes no difference on its own. Excel was on every desk for decades. The formulas were there, the macros were there, the connections to other data sources were there. And still most spreadsheets ended up as tables with borders. Nobody stopped people from using the power. Nobody showed them what it could do for their own work, either.

So what should a company do differently this time?

Make the agents available — to everyone, not to a pilot group of enthusiasts. Teach people how they work, so nobody gives up after the first wrong answer. And then do the thing companies find hardest: do not tell them what to use it for. The people who know which tasks waste their week are the people doing those tasks. Let them try, let them get it wrong, and make it easy to share what works. A good use case travels fast once someone shows a colleague how much of their Tuesday they got back.

There is a tension in that advice, and I do not want to skip over it. Training people on a tool without telling them how to use it is harder than it sounds. Every training is a set of examples, and examples quietly become the boundaries of what people think the tool is for. I have not seen anyone get this balance right yet.

Anyone working in risk or compliance will have stopped reading three paragraphs ago, with one word in mind: Excel. Uncontrolled spreadsheets are one of the most familiar control gaps there is — macros nobody can read, feeding reports nobody can trace back.

It is a fair objection. But look at what actually went wrong. Not that people built their own tools. That nobody knew what had been built until something broke. The answer to that is not to prescribe the use cases from the top. It is visibility: who built which agent, for which task, on which data. Inventory first, restrictions where the inventory shows they are needed. Let people automate the work they hate. Just not in the dark.

The real payoff is not the minutes. A head that is no longer busy avoiding a task has room for something else — a better idea, more time with clients, a conversation with a colleague that is not about the report. The companies that start early, and make this part of how they work rather than a project with an end date, will feel that difference first.

And in the spirit of the argument, a disclosure: the first version of this piece was mine. The editing was done with an agent. It took less than half a day.