How we worked in 2025 and how we work now: what changed inside an agency when AI stopped being an open browser tab

btodigital before and after AI, process by process: from 3 days to 45 minutes per piece, what we learned, and a 30-day test you can run in your own company.

In 2025, one piece of content took btodigital three days. Today we produce five variants in 45 minutes. We did not switch tools to get there: we were already using ChatGPT in 2025. What changed was where AI sits inside the company. It stopped being a tab open in each person’s browser and moved into the processes themselves, with a memory of each client and working on our own data.

This is the first article in a series about how btodigital, the marketing agency I founded in Medellín, Colombia, in 2013, changed with artificial intelligence. In AI will not take your job I wrote about what happens to people. This one is about what happens to the company: what we used to do, what we do now, what we learned along the way, and how you can test it in your own business in 30 days.

It is written for the manager of a small company or an agency who uses ChatGPT now and then and suspects they are not getting what was promised. I was in that exact spot less than two years ago.

What is the difference between using AI and operating with AI?

Using AI means opening a chat, asking for something, copying the answer and pasting it somewhere else. Everyone does it their own way, the chat does not remember the client, and the output depends on who wrote the prompt that day.

Operating with AI is something else. The AI goes inside a process that already exists: the monthly report, WhatsApp customer service, the content calendar, the sales commission payout. It knows which client it is dealing with because it holds that client’s memory, it works on company data and not just on what it knows about the world, and one person is accountable for what comes out.

A quick way to tell which one you are doing:

Signal	Using AI	Operating with AI
Where it lives	In each person’s browser tab	Inside the process
What it knows about the client	Whatever you paste in at the moment	Their history, brand and data
Who is accountable	Nobody in particular	A named owner
How it is measured	”I saved time”	A number defined before you start
When the person leaves	The knowledge leaves too	It stays in the system

Many of the companies I visit sit in the left column and believe they are in the right one because they pay for a subscription.

How did we work in 2025?

On December 19, 2024, I finished our strategic plan for 2025. It focused on SEO and paid media, stated in so many words that we would “outsource graphic design and web development to partners,” and proposed training the team in AI for SEO and paid media. AI was a training topic. It was not yet a way of operating.

Day to day looked like this:

Loose tools. In May 2025 I put together a document for internal classes covering ChatGPT, Gemini, Gamma, NotebookLM, Sora and CapCut. Each one solved something and none of them talked to the others.
Reports built with Metricool and templates. Someone pulled the data, pasted it into the template and wrote the commentary. Every month, for every client.
Third-party bots on WhatsApp, with their own rules and limits.
The buyer persona in a Word file, written at kickoff and then forgotten in a shared drive.
Team commissions calculated by hand in spreadsheets.

We were not a backward agency. We were a normal one. But there was a signal we were not reading properly.

In November 2024 I ran a survey of managers and executives at Colombian companies and got 114 responses. I asked those who worked with an agency what they would improve. The most frequent answer was meeting deadlines, followed by proactivity. They were not asking for more creativity. They wanted us to deliver on time and to get ahead of problems.

An open tab fixes neither. It speeds up an isolated task, but it does not get the whole process delivered on time, and it does not make anyone spot the problem before the client has to raise it.

What changed, process by process?

In 2026 AI stopped being an SEO aid and a training topic. On March 2 we launched our own custom-built website, and we built everything else on top of that foundation:

Process	2025	October 2026
Content	3 days per piece	45 minutes for 5 variants
Client knowledge	Buyer persona in a forgotten Word file	Cerebro IA de btodigital, our per-client memory platform (since July 8)
Social media comments	By hand, network by network	Community module: classified automatically, replies approved by a person
Reports	Metricool and templates	Automated; the audience report sends itself since October 2
WhatsApp	Third-party bots	Atendio, our own platform (since June 16)
Commissions	Spreadsheets	Calculated against billing data from our ERP

Two examples so this does not stay in a table.

Per-client memory. Before, anyone opening a chat to work on an account started from zero: explaining the brand again, the tone, what the client had already approved. Today the Cerebro IA de btodigital (our “AI brain”) keeps each client’s documents, brand kit, buyer persona and past output. Someone new to an account does not start from a blank page.

Comments for an industrial brand. We connected its social accounts to the community module, and the first sync pulled in 85 real comments. The AI classified them as 72 positive, 12 neutral and 1 negative, and drafted a reply for each one in the brand’s voice. None of them went live until a person approved it.

That is operating with AI. The machine does the mechanical part (reading, classifying, writing the first draft) and a named person decides what gets published on the client’s behalf.

What happened to the team?

This is where it is easiest to dress up the numbers, so I will be careful.

Between July 2024 and January 2025 the team went from 37 to 29 people. That happened before AI entered our operations. It has nothing to do with AI and I will not use AI to explain it.

What does coincide with AI is this: in December 2025 we were 32, and today we are 28 people in Colombia, 22 in operations, 3 in sales and 3 in administration. Some people left and we did not replace all of them. I cannot separate exactly how much of that is AI and how much is other factors, but saying AI played no part would be dishonest.

With those 28 people, in September we invoiced 62 companies and closed 1,031 tasks on our project board. So far in 2026, 117 different companies have bought from us.

What rules did the change leave us with?

Three lines I repeat at every workshop. The most recent was with an MBA group at FUNDEPOS in Costa Rica, in August.

“AI gets you to 70%, the person gets it to 95%.” The first AI draft is not better than before. The difference is that it arrives in minutes, and the person spends their time taking it to 95% instead of writing from a blank page. If your team expects the machine to deliver 100%, they will get frustrated and drop it.

“The team is not an AI user, it is a co-designer.” The people doing the work every day know where it hurts, and the tool has to fit them, not the other way around. The story behind this rule is below, under what we learned.

“‘I saved time’ is not a metric.” Hours per month, delivery days, tasks closed, replies on time: those are. If you do not define the number before you start, any result will look good at the end.

What we learned

The tool gets better when the team designs it with us. The first version of the Cerebro IA de btodigital did not do what the team needed. When our content team tried it, the feedback was blunt: it was too extractive, it quoted sources but proposed nothing. Within 24 hours we switched it to an “ideation mode” that proposes new ideas from the client’s memory. The version in use today came out of that criticism, not from my original design.

Before you blame the AI or the software, check the data and the rules. When the module calculated commissions against ERP billing, the total came out 9% below the manual calculation. The formula was the same; the gap came from configuration: we were applying today’s client assignments to past months, and those assignments change every month.

Take the baseline before you start. The time comparison I have well documented is the content one. For reports or commissions I did not keep a 2025 measurement that can be compared, so I will not give you savings figures I cannot back up. That is why the test below starts by measuring.

How do you run the 30-day test in your company?

Before you open any tool, answer these four questions in writing:

Data: what information do you have that your competitors do not?
Process: which specific process will you tackle? One, not ten.
Owner: who is accountable for the result?
Metric: which number will change, and how will you measure it in 30 days?

If you cannot answer one of them, you are not ready yet. In most cases I have seen, AI projects that fail in small businesses fall apart in those four boxes, not in the technology.

With the four answers in hand, the month looks like this:

Week	What you do	What you deliver
1	Pick a process that repeats at least 20 times a month and measure how it performs today	The baseline: time, errors, on-time delivery
2	Build the memory: documents, good examples, house rules	A project in Claude, ChatGPT or Gemini Notebook (formerly NotebookLM) holding that information
3	Run the process with AI alongside the normal one; AI gets to 70% and the person takes it to 95%	A list of every correction the person made
4	Compare against the baseline	The decision: scale, adjust or switch it off

Three rules for those four weeks:

Do not upload personal data to free plans. Payroll, medical records or customer data go in anonymized and on business plans. Check the data protection law where you operate; in Colombia it is Law 1581.
The week 3 corrections are the most valuable output of the month. Each one is a rule your memory is missing. Add them.
Switching it off is also a good result. If the number did not move in 30 days, you learned cheaply.

If you want to run this test with support, this is how I work as an AI consultant for companies: one process, one owner and one number, and the decision follows from there.

Frequently asked questions

What is the difference between using ChatGPT and operating with AI? Using ChatGPT means asking for one-off things from a browser tab, with no memory of the client and nobody accountable for the result. Operating with AI means placing it inside an existing process, with your own data, per-client memory, a named owner and a metric defined before you start.

Where should a small business or an agency start? With a process that repeats at least 20 times a month and is done from scratch every time. In an agency that is usually reporting or content. In a small business, it is almost always customer service on WhatsApp.

Did AI shrink the btodigital team? There was a reduction from 37 to 29 people between July 2024 and January 2025, before AI, which has nothing to do with it. Between December 2025 and September 2026 we went from 32 to 28 without replacing everyone who left, and that does coincide with AI. Today we are 28 people in Colombia.

How long does it take to see the change? A well-run test gives you an answer in 30 days. Transforming the whole operation takes longer: in our case, from the December 2024 plan to our first in-house modules in March, June and July 2026.

Do I need to hire developers? Not for the 30-day test: projects in Claude, ChatGPT or Gemini are enough. Building your own tools does require technical capacity. I built a large share of what we have today myself with Claude Code, without putting together a large engineering team.