Team Perspectives

Once AI joins the workflow, the things that matter most are the easiest to underestimate

·8 min read

Once AI enters the workflow, the problem shifts from tooling to governance

About three or four years ago, I was invited to help a large conglomerate select a consultant for a development process improvement engagement. I came in through a shareholder's introduction, so I was not there as an outside consultant. My job was to look at the people and the methods from a practitioner's point of view, and judge whether their recommendations had any real chance of being put into practice.

Several large SI firms and some management consultancies were competing for the job.

What stuck with me was that some of the people who looked the most professional, with polished decks and well-presented frameworks, had spent very little time on an actual development floor since graduating. They could explain the methods and write the reports, but much of what they said was essentially recited from the textbook.

With process, if you have never actually built a product, struggled through collaboration or carried delivery pressure, what you end up with is easily just a nice way of describing things.

So whether I am looking at a development process, team collaboration, or now AI workflows, there is one thing I pay particular attention to: can it actually reach the front line, or does it stay in slides and documents?

That is also why, while helping several startups and groups adopt agentic development recently, I have found myself caring more and more about things that look basic but are the easiest to underestimate.

The hard part has never been whether an agent can be plugged in. It is which things can no longer be left vague once an agent starts taking part in the workflow.

Conclusion first. These are the five things I think managers should watch when bringing AI agents into a full SDLC:

  • Whether decision rights have been clearly assigned first.

  • Whether finishing the execution can be treated as finishing the process.

  • Whether the workflow, for all its flexibility, keeps a basic skeleton.

  • Whether multiple agents are adding capability or clarifying responsibility.

  • Whether the whole process can be tracked, understood and handed over.

Let me go through them one by one.

More than automation, I now care about how decision rights are divided

When teams first talk about agentic development, the natural reaction is: can we automate a bit more? Can we cut a bit more manual work? Can the agent just carry on by itself?

That is a perfectly normal way to think. We went through it ourselves.

But once you actually do it, the first thing to deal with is not automation at all. It is decision rights.

  1. What AI can do directly.

  2. What AI can recommend but not decide on its own.

  3. What still needs a person who is accountable for the final call.

If this is not settled first, more automation later often makes things more dangerous, not more stable.

Once an agent is handling more than single-step execution, and starts moving the process forward, handing work between roles, drawing conclusions from verification, or even deciding whether the next step goes ahead, you are no longer managing a tool. You are managing a system that affects how the organisation makes judgements.

If you try to stay vague about that at the start, you usually pay a bigger price for it later.

The agent being done does not mean the process is done

This is something I have felt strongly about lately.

AI easily creates an illusion: it works fast and produces a lot, so it feels as if things are getting finished fast too.

That illusion is exactly what a workflow should fear most.

An agent putting together an analysis for you does not mean the analysis holds.

An agent finishing a piece of implementation does not mean the requirement has actually been met.

An agent completing a round of verification does not mean this stage is really ready to move on.

I have seen "there is output" and "it is complete" get mixed up many times recently. And honestly, this is one of the places where teams most easily slip after adopting AI.

When people were slow, everyone knew each step had to be checked.

Now that AI is fast, people are more likely to skip the check and treat "done" as "good to go".

Those two are very different.

So I increasingly insist that execution complete and process complete must be kept separate.

Someone still has to be able to define whether this step actually passes and whether it can move forward.

Otherwise the whole process will run very fast, but not necessarily accurately.

A workflow can be flexible, but it cannot lack a basic skeleton

I do not really believe the claim that one standard agentic workflow can fit every team.

In the real world, tasks vary too much.

Some are research.

Some are hotfixes.

Some are just small adjustments.

Others are large projects spanning teams and systems, with a lot of collaboration involved.

These could never all follow exactly the same process.

So I agree that workflows should be flexible. But I do not agree that a workflow can do without a skeleton.

In other words, you can have different templates, adjust the order by task type, and let different organisations grow their own ways of working. But however it changes, a few basic questions cannot disappear:

What exactly is this round supposed to do?

Who does it?

Who checks it?

After the check, does it move forward, go back, or change direction?

If these are not spelled out, a lot of so-called agent workflows end up being a pile of prompt chains and automation scripts in nicer packaging.

It looks like a process, but it is not one.

The real value of multiple agents is not more roles but clearer responsibility

Another common misconception in this wave of agentic development is that people readily equate "multi-agent" with "more advanced".

But having watched this closely for a while, I am increasingly convinced the key is not the number of agents or the amount of work (time) they get through. It is whether responsibilities have been clearly divided.

Analysis, implementation, verification and coordination are different jobs to begin with.

If you just copy the same generalist agent many times, the system is not necessarily better. Often it is just busier.

Multi-agent setups that really add value are not ones where every agent can do everything. They are ones where it becomes clear what each role should look at, what it should do, and which outcomes it is accountable for.

The benefit is not just that it "looks more like a human organisation".

The real benefit is that when something goes wrong, you finally have a chance to trace back which stage the problem came from.

Otherwise, if analysis, implementation, verification and coordination all end up mixed together again, you have only moved the vagueness from people onto the system.

If you cannot track how the process moves forward, it is not a workflow yet

Many agent demos look great, because the agents really can do things.

But once they enter a real team, what people ultimately care about is not whether it can do the work, but:

  • How far has it got?

  • Why is it stuck?

  • Who decided this step could go ahead?

  • Did this failure come from analysis, implementation or verification?

  • If someone has to take over, where do they pick it up?

If these questions cannot be answered, I would not call it real workflow management yet. It is more like an agent running in the background.

That is why I now put a lot of weight on tracking and observability.

Not because it is cool, but because if a process cannot be tracked, understood and handed over, it will struggle to become part of a team's daily work.

Many people think the most important thing after adopting agents is to "let it run by itself".

What really matters is whether people can understand what it is doing while it runs.

Teams Hub is the name we gave it

All of the above is distilled from experience. We started building our own agentic development framework in Q3 last year, and it is now on its eighth version. Each change came from finding a problem and adjusting step by step. After Lunar New Year, everyone felt it was time to give it a name, and the vote came out as Teams Hub. If your team wants to build one of its own, here are some suggestions.

Records matter

Teams Hub

What our RD engineers look at most is how work is split between Agent (PM) and Agent (RD), and the decisions behind each stage transition. We took a few detours early on, and eventually settled on a principle: work can go wrong and can be redone, but you must always find a way for that work to get on the right track next time.

People are essentially the requesters and the supervisors

teamshub-task.jpg

That requirement was actually only three lines, under 50 characters in total, yet our agents drew up all those requirements and the implementation plan based on the product, existing specs and conventions. The key is the adaptive workflow and the retrospective mechanism. Of course, you also need a strong model behind it, or you will not even notice the problems.

Planners must all be SOTA models

For the agents themselves you can use Codex, CC or Gemini CLI, via API or subscription. For roles like Planner / Orchestrator, we only use whatever the SOTA model is at the time, and really can only use those. Right now Opus with 1M context is well worth trying for planning.

Designers can have agent teams behind them too

Our designer went from directing several agents at once to building several workflows, and now spends most days reviewing design output together with the agents (in practice, the agents keep asking him to endorse the drafts). Most of our design files have also moved from Figma to Pencil. It is not as powerful yet, but sometimes good enough is enough.

Building an agent framework that fits your organisation

This is a new line of business we started recently. Using this methodology, together with specs and interviews, we can stand up a practical agent framework that fits an organisation's internal processes in one to two days. The framework also grows and corrects its own workflows, so it does not get frozen once built and become hard to change.

In fact, the hardest part of launching an agent framework is straightening out internal processes: knowing where to put people and what they should do, and most importantly, how to verify that what the agent did is the same thing you wanted.

If you and your team are still looking for an agent platform, having one of your own may be the better choice. If you are not sure how to choose, you are welcome to talk to us.

P.S. If you are in financial services, you can contact the DE team at EY Taiwan. Alongside our agent framework, they also provide services covering compliance and information security requirements.

Need Professional Assistance?

Our team are senior engineers and former executives with 20+ years in commercial product development — from apps serving hundreds of millions of users to B2B SaaS platforms serving tens of thousands of merchants. In recent years the focus has been enterprise AI adoption, from mapping the opportunity through to tuning what is live. Whether that means:

  • Mapping AI opportunities and sequencing them
  • Preparing data flows and integrating existing systems
  • Building forecasting and optimisation models
  • Monitoring and tuning after go-live
  • Test strategy and process review
  • Building and adopting test automation

We can provide professional advice and concrete solutions for teams of different scales and types. If you're facing similar challenges, feel free tocontact our consulting team.