FIELD NOTE / N12
Agent Workflow 1.0 to 2.0: How 6 People Do the Work of 16
'AI helps you do things faster' actually understates the whole picture. When AI moves from being a tool to being an agent, it can carry out multi-step work on its own. It then returns finished output for review. At that point, the productivity ceiling of a small team is no longer set by headcount. DKH is a six-person team. Our daily output spans strategy, content, design, websites and data. All of this is supported by an agent workflow that has been through two versions. This piece breaks down both versions so you can see where to start and how to evolve.
An agent workflow is a division of labour in which a process is broken into steps. AI employees produce first drafts and handle repetitive work. People own judgement and quality control. For a small team, the key is not the tools but how the process is broken down and how sign-off is designed. DKH has validated this approach within its own six-person team.
TL;DR
- Workflow 1.0: people lead, AI provides single-point assistance; saves time but does not change the capacity ceiling
- Workflow 2.0: AI employees lead execution, people provide judgement; output decoupled from headcount
- The dividing line between the two versions is the quality gate: without a quality gate, do not let the AI employee run autonomously
- Starting from 1.0 is the right path; jumping straight to 2.0 usually goes wrong
How does Workflow 1.0 work?
The first version was straightforward. Each colleague used AI as an assistant within their own part of the workflow. Writers used AI to produce first drafts. Data people used AI to tidy up reports. Designers used AI to generate roughs. The process itself did not change. It still ran person to person. Only each leg was faster.
The value of this stage is building judgement. The team starts to learn what AI does well and where it fabricates information. They also learn which outputs cannot be used without editing. All of this experience is the foundation for 2.0.
The limits of 1.0 are also clear. Because the process did not change, the bottleneck is still the handover between people. Much of the time saved gets absorbed by waiting.
From 1.0 to 2.0, what has changed?
The core change in 2.0 is workflow redesign. It is no longer "people do the work, AI helps". It is "AI employees run the entire process, people sign off at key points". For example, a piece of content can go from research, drafting and rewriting through to layout in one continuous run by an AI employee. Human effort concentrates in two places: a clear brief at the start, and quality judgement plus brand tone at the end.
To get there, three things must exist first.
First, documented workflows: if the process is not written down, the AI employee has nothing to follow. Second, quality gates: the line between outputs an AI employee can pass straight through and those that require human sign-off must be drawn clearly. Third, ownership: every workflow has one person responsible for maintaining and reviewing it, not "everyone keeps an eye on it".
How is a quality gate actually designed?
The principle is risk tiering.
For low-risk internal output, such as data organisation and preliminary analysis, the AI employee produces it directly and a person spot-checks. For medium-risk external content, the AI employee drafts and a designated colleague signs off before it goes out. For high-risk matters, such as price commitments and sensitive points in client communication, a person always leads. This also applies to any claim that carries numbers. In those cases, the AI employee at most prepares a draft.
This tiering is not designed perfectly in one go. It emerges through practice. Each time an AI employee makes a mistake, we review it and then upgrade the checks on that type of output. Over a few months, the line settles naturally.
How should a small team get started?
Do not try to build 2.0 from the start. The sensible path is this: pick the highest-frequency workflow in your team and run it well in 1.0 mode first. That means human-led and AI-assisted. Document the workflow along the way. Once the team can judge the quality of AI output, hand the lowest-risk steps in that workflow to an AI employee to run autonomously. Then expand step by step.
Skipping steps has a real cost. Letting an AI employee run autonomously without that judgement means leaving quality to chance.
For how team structure should support this shift, see AI Team Restructuring: Grow Output Without Growing Headcount. For a training route that starts from zero, see SME AI Transformation Step 1: Train Your Team, Not Buy Tools.
Next step: hand it to us. Free 15-minute consultation, leave your details on the contact page and we will WhatsApp you within one working day.
FAQ
FAQ
- Does an agent workflow require coding skills?
- Not necessarily. Plenty of agent tools now let you set up workflows in natural language. What matters more than the tool, however, is the ability to break down the process: you need to write each step clearly enough that a new colleague could follow it and get it done, and only then will the AI employee run smoothly.
- How small does a team need to be to use it?
- It works for teams of two or three, and the return is even more pronounced for small teams, because everyone's time is a bottleneck. The only difference is lighter execution: no formal system is needed, but process documentation and quality gates are two things you cannot skip.
- What happens if an AI employee makes a mistake?
- Mistakes will happen, which is precisely why quality gates are the core of the whole system. The principle: an error caught by a quality gate is the system working as it should; an error that reaches the outside world is an incident. After every mistake, review it and adjust the level of gatekeeping accordingly.
- How long does it take to go from 1.0 to 2.0?
- Depending on the team's starting point, the transition is typically measured in months. The signal to watch for is when the team starts spontaneously writing repetitive tasks into workflows and proactively asking, "Can this step be handed to an AI employee." Before you see that signal, do not rush to push 2.0.
KEEP READING
Keep reading
NEXT STEP