The future of work, seen from inside a small business founded in 2013

In December 2025 there were 32 of us at btodigital. By September 2026 there were 28, and we did not replace everyone who left. That is the number people ask for when I talk about the future of work. I want to give it in full, because the half almost nobody tells is the half that matters most to someone who runs a team.

I have spent more than 20 years working with teams, and I have run btodigital, a digital agency in Medellín, Colombia, since 2013. In September I wrote that AI will not take your job, someone who knows how to use it will. That piece spoke to the person worried about their own job. This one speaks to the person who has to redesign other people’s jobs: team leads, managers and HR people who know something has to change and do not know where to start.

What actually happened to the size of the team?

The numbers first, with their context.

When	People	What explains it
Early 2024	37	Starting point
Late 2024	29	A drop unrelated to AI: we were not using it in operations yet
December 2025	32	We grew again
September 2026	28	Departures we did not fully backfill

Any conference speaker would be tempted to put 37 next to 28 and say AI let us run with almost a quarter fewer people. That would be false. The big drop happened in 2024, before AI entered our operations, and I am not going to credit it to AI.

What I do credit to AI is smaller and more useful. Between December 2025 and September 2026, when someone left, we did not always open the position: part of that work was already done by a tool, and the rest was absorbed by the people who stayed. Today we are 28 people in Colombia: 22 in operations, 3 in sales and 3 in administration.

The other half of the change does not show up on payroll. Much of the internal software we use today I built working with Claude, Anthropic’s AI, as if it were part of the development team. My estimate is that three years ago this would have meant hiring about five more engineers. That money stayed where people are actually needed: strategy, judgment and client relationships.

What has to change in people’s heads before it changes on screen?

Before a tool there is a team, and before a team there is a mindset. These are the three we had to change, and they are the same three I see stuck in almost every company I visit.

One: iterate instead of redoing. The old line was “I don’t like how it came out, let’s do it again.” The new one is “refine this version with these three changes.” Starting over throws away the part that was already good. Iterating keeps it and puts the effort where it is missing.

Two: describe the problem well. The question used to be “who knows how to do this?” Now it is “how do I explain this to the AI so it can do it?” It sounds minor and it is the deepest of the three, because it forces people to understand their own work with a precision almost nobody needed before.

Three: work in drafts. From “this has to be perfect before I show it” to “this is a draft, what do you think?” When producing costs almost nothing, waiting becomes the expensive part. An example from our own shop: the monthly audience report for a regional news outlet now builds itself, but it comes out as a draft. Someone on the team reviews it and approves it with one click, and only then does it reach the client. The machine produces; a person answers for what goes out.

What conditions does a team need for AI to work?

I presented these five at a workshop with MBA students in Costa Rica, and I hold every client to them. Without them, AI does not work in any company, however good the tool.

Condition	The question that tests it	The sign it is missing
1. Your own data	What information do we have that competitors do not?	Everyone runs AI on public information
2. A human owner	Who answers for this use case?	”We’re exploring it” for three months running
3. A metric before starting	Which number should move in 30 days?	”We saved time”
4. A draft culture	Do we accept 70% and let a person take it to 95%?	Nobody shows anything until it is perfect
5. Minimum governance	Which data may leave the company and which may not?	People uploading payroll to personal accounts

Two of them deserve one more line.

“I saved time” is not a metric. It is a feeling. Hours per month, receivables recovered or appointments not lost are metrics. If you cannot write the number down before you start, you will not be able to defend the project when someone asks whether it worked.

The 70/95 rule. AI gets the work to 70%; a person takes it to 95%. That split is the basis of the role redesign below: what you ask of people is no longer to produce 100%, but to spot quickly what is missing from that 70% and complete it with judgment.

What does the evidence say about who wins and who loses?

What we lived through matches what others have measured, and what they measured carries a warning HR should read slowly.

Study	Who was measured	Result
Brynjolfsson, Li and Raymond, Quarterly Journal of Economics (2025)	5,179 support agents with an AI assistant	14% more cases resolved per hour; novices, 34% more
Dell’Acqua et al., Harvard Business School (2023)	758 consultants	Lowest performers improved 43%
Otis, Clarke, Delecourt, Holtz and Koning, Management Science (2026)	640 small business owners in Kenya with an AI mentor on WhatsApp	Those already doing well improved 15%; those doing poorly got 8% worse

The first two rows are hopeful: AI hands the newcomer what the best performer knows. That shortens the learning curve and changes how you should think about onboarding.

The third row is the warning. The Kenyan owners who were struggling asked the tool for miracles and followed generic advice. AI amplifies the judgment that is already there; it does not replace it. That is why redesigning a role is not handing out licenses. It is training the judgment needed to assess what the machine delivers.

How do you redesign a role without breaking the team?

This is the procedure we use. It works for operations, customer service or administrative roles.

1. Inventory a real week, not the job description. Job descriptions lie by omission. Ask the person to log everything they do for five days, and how long it takes.

2. Put each task in one of three columns. AI drafts and the person reviews. The person decides with AI’s help. Human only: the hard conversation, the negotiation, the client relationship, the decision with consequences.

3. Design with the person, not for them. When two people on our content team first tried the Cerebro IA de btodigital, our client knowledge system, their verdict was blunt: “this is too extractive, it’s just like NotebookLM” They were right. Within 24 hours we changed it so it proposed new ideas instead of only quoting sources. That afternoon I understood the line I now repeat at every workshop: the team is not a user of AI, it is a co-designer.

4. Give the freed-up hours a name. “AI didn’t come to replace my people, it came to give us back our hours.” I said it at the close of that workshop in Costa Rica and I stand by it, with one condition: you have to decide where those hours go. If they have no named destination (more time with clients, more analysis, the conversation that kept getting postponed) they evaporate or, worse, turn into fear.

5. Change the role’s metric. If the role now curates instead of producing, measure it by the quality of what it approves and the errors it catches, not by the volume it ships.

Which skill is becoming more valuable?

Describing a problem. Not coding, not writing clever prompts: describing precisely what is happening, what was expected and how the two differ.

One example: someone asked our system for three video scripts and got two, cut off. The report was precise: we asked for three, we get two, they get cut. With that we found three faults in one day: a cap on response length, a website that blocked our reader and a system that did not warn anyone when it could not read a page. A plain “it doesn’t work” would have cost us a week.

AI has pushed the cost of producing close to zero. The new barrier is curating, not producing. And to curate you have to be able to name what is wrong. A leader who can describe a process step by step, without ambiguity, is worth more today than three years ago. That skill is in no job description and it should be in all of them.

What we learned

Measure productivity before you start. I have the headcount, and I know the team closed 1,031 tasks in September 2026. But I have not yet built a comparable series for 2025. So when I say we are fewer and do more, I am saying it from what I see, not from a number I can show. It is the same “I saved time” mistake I point out in others.

Design the tool with the team, not for the team. We designed the first version of our client system without the team, and it came out extractive because we thought about it from the technology side and not from the work of the people who would use it. The fix was quick; the lesson was one of method.

Not backfilling a departure has a cost that does not show up on payroll. If you do not redesign the role of the person who stays, you have just added to their workload. It is the risk I watch most closely, and the one headcount figures hide best.

How can you apply this to your team this month?

One role, four weeks. Not the whole company.

• Week 1. Pick a role with repetitive tasks and ask the person in it for a five-day inventory.
• Week 2. Sort every task into the three columns together. Write down the metric that should move and who owns the pilot.
• Week 3. Run the pilot with the person as co-designer. Every time they say “this doesn’t work,” ask them to describe what they expected and what they got.
• Week 4. Measure against the week 2 number. Decide where the freed-up hours go and write it down.
• Always. One page with the data rule: what can be uploaded to an AI tool and what cannot, and under which account.

If the pilot fails, you learned cheaply. If you want to see why these projects tend to fail for the same reasons, I wrote about why AI projects fail in small businesses. And if you would rather have support training your team, this is how my Claude training for companies works.

Frequently asked questions

Did AI shrink the btodigital team? Partly, and it is worth separating carefully. The big drop, from 37 to 29 people, happened in 2024, before we used AI in operations. Between December 2025 and September 2026 we went from 32 to 28 because we did not replace everyone who left, and there the fact that tools already handled part of the work did matter.

Where do I start redesigning a role around AI? With an inventory of one real work week, not the job description. With that list, sort each task into three columns: AI drafts, the person decides with AI’s help, or human only.

What skill should I hire for? The ability to describe a problem precisely and to judge whether a result is right. Writing prompts can be learned in a week; the judgment to spot what is missing from a draft comes from knowing the craft.

Does AI help newcomers or experts more? Studies show it helps newcomers more: among 5,179 support agents, novices improved 34%. But in Kenya, small business owners who were already struggling got 8% worse with an AI mentor. The tool amplifies whatever judgment it finds.

How do I know whether AI worked for my team? With a number written down before you start: hours per month, receivables recovered, appointments not lost, errors caught. “I saved time” is not a metric, it is a feeling.