We have recently been spending time on agentic workflow analysis and design in large, complex organisations, and are learning lessons that we will try to share about the value of this work in advancing enterprise AI.
Many organisations start with product or customer functions as the quickest route to value, but if we accelerate the leading edge of the firm, whilst support functions and central services remain stuck in old ways of working, then this risks creating disconnects and coordination problems. Plus, there is a ton of value and ROI to be mined in what some people still think of as boring bizops functions.
But the detail of how we approach this challenge really matters, so here is some recommended reading on the topic that crossed my desk in the past couple of weeks.
Service design over process mining
When you make processes visible and digitally addressable, they also become programmable, and you can apply agentic AI to run them more efficiently and orchestrate them more effectively.
That might suggest that the quickest route to value is process mining to discover them, followed by process automation to speed them up. But many processes are defined by an accretion of workarounds and exception-handling, and it is rare to find a major example that does not need human oversight and judgement to handle these complications. Also, some processes are legacies of previous problems, or a response to prior stupidity, and do not necessarily need to exist at all.
Rather than start with process mining, it can be more enlightening to begin with capability definition - what is the superpower you want to create, for whom and why? - and service design to identify how that capability translates into value for its users and/or customers.
Developing a holistic view of a whole workflow helps to identify where agents and automation can help, and where human accountability, agency and judgement are still valuable.
From there, we can decompose the elements and components needed to make up the capability or product, and start to define the multi-agent architecture typically needed to make it work.
How to ‘read’ an organisation
Jennifer Houle recently wrote about How to read an organisation, and reminds us that what people say happens and what actually happens often diverge at every level, which implies that you can’t just build AI-enabled workflows from documentation or from asking people to describe their process. You have to follow actual work. If you automate the documented process without ever observing the real one, you get faster broken processes.
The business is what the organization is trying to do. The organization is the human system through which it is attempting to do it.
Houle argues that you need to follow the work. Pick up a thread and follow it end to end. Look at handoffs, workarounds, and bottlenecks. Find the exceptions. Watch it happen rather than asking people to describe it, and document behaviour before explaining it.
Some enterprise AI workflow projects start too late in this process, arriving after someone has already decided what to automate and just wants help with the implementation.
Workarounds are not always process failures to be corrected. Sometimes they are evidence of where the real design requirement lives. If you automate around them rather than into them, you might be designing for the documented process, but not the real one.
The design lesson here is that humans need to stay in contact with the work, not just above it; and this is an argument for keeping workflows observable, minimal, and reversible.
Workflow economics
Following on from recent debates about enterprise AI token costs and ROI, Ken Huang believes workflow economics is the new moat, not the AI model itself:
The model still matters. I would not argue otherwise. But the budget outcome now depends just as much on context shaping, caching, permissions, semantic definitions, approval boundaries, and stop rules. Those controls decide how often a workflow escalates, how much context it drags into the run, and how many steps it burns before a human sees the result…
This is where I think a lot of AI strategy will break over the next year. Teams that budget by prompt or by vendor rate card will keep underpricing autonomy. Teams that budget by successful outcome will see much earlier which workflows deserve more freedom and which ones need a cheaper lane or a harder boundary.
This is also one reason why we are seeing a closer relationship between CHRO and CFO functions, according to SiliconANGLE, with human talent, digital labour, technology investment and productivity increasingly managed as an interconnected system. When this becomes a three-way cooperation between people, finance and G&A function leads, the potential for savings whilst also improving employee experience really starts to come into focus.
But how do you define success? Internal functions running major workflows seek to become more efficient, but the bigger gain might come from marginal efficiencies for all of their internal customers using their services. This is also an area where service design is important, because it can define what a successful outcome looks like, which is what you then optimise workflow economics around.
Context, knowledge and documentation
Atlassian are an interesting company in this space, and are setting themselves up to be a key platform provider for context in the enterprise. They are best known for providing tools like Confluence, Jira and Trello for software and engineering teams to run their processes and documentation. But now that other areas of the enterprise are starting to think like designers and developers of what you might call human software - workflows, processes and codification of the workplace - they are widening their catchment area.
Co-founder and CEO Mike Cannon-Brookes recently spoke about their own experience of building and accelerating workflows, and whether their particular niche might be impacted by the so-called Saaspocalypse, in an interview with Nilay Patel:
We’re moving from information asymmetry to imagination asymmetry where your ability to create and think is going to be far more important as a competitive advantage for your team, your job, and your business than your access to information, your control of information, your understanding of information. That part is going to be cheap. That’s what LLMs are really good at. What they’re not good at is imagining things.
He also clearly sees human talent and systems as symbiotic, rather than competing approaches:
I ended up refereeing a bunch of, ‘Is this a talent problem or a system problem?’ As I say, that doesn’t make any sense. Usually the answer is this one’s 60 percent talent, this one’s 40 percent system... I’ve put them all together inside the business so that our system changes to apply AI internally and our talent changes to apply AI internally are in the same function.
This balance is important if we are to avoid some of the potential downsides of greater workflow automation. One of these is an area where Atlassian is very strong - documentation, codification and context. With so much contextual knowledge remaining either implicit, or being buried away in documents and decks, rather than connected knowledge systems, agents running workflows will struggle to have all the background they need to understand how work is managed.
Prioritise collective intelligence over super-agents
This brings us back to a perennial question with agentic workflows, which is the balance between human intelligence, machine intelligence and explicit knowledge captured as context. There are so many public debates around frontier models and their pursuit of AGI that some organisations are rightly nervous about the risk of agents re-creating unforeseen consequences like the story of the sorcerer’s apprentice.
We don’t need one super-intelligent agent running major workflows. We need the smallest, simplest and most observable agents owning specific outcomes or roles, and then we need to maximise the collective intelligence of these agents working together.
DeepMind recently published a useful essay on this topic, and reminded us how effective this pattern of collective intelligence within connected ecosystems is in nature:
As we face the prospect of AGI, the prevailing popular narrative of the Singularity - a single, titanic AI model bootstrapping itself to godlike, isolated intelligence - is likely the wrong vision. Every major evolutionary transition in the history of life on Earth, from multicellularity to symbolic culture, has been a highly social event. The next intelligence explosion will be no different; it will be plural, heavily social, and deeply entangled with the messy, complex reality of human culture, norms, and institutions.
Humans in/on/above the loop
And whether we put human oversight in the loop or above the loop, the choice we make about the role of human intelligence in agentic workflows should not be governed by efficiency alone. We need to retain enough tacit knowledge and involvement to maintain our agency and our ability to intervene when needed.
Another potential risk highlighted by Bertrand Duperrin recently, is Bainbridge’s paradox of automation, which can lead to a loss of human control when it is most needed:
In 1983, British psychologist Lisanne Bainbridge published an article that has become a seminal work in the field of automated systems: Ironies of Automation. Based on her observations of industrial control rooms and airplane cockpits (which were, however, far less automated at the time than they are today), she highlighted what is now known as the automation paradox, and her bottom line was that the more automated a system is, the more indispensable the human operator becomes, yet the less trained they are when their intervention is required…
Automation shifts employees’ roles from execution to supervision. This shift reduces daily practice, gradually erodes skills, and makes it more difficult to take back control when the system reaches its limits.
Reimagining work and redesigning workflows based on the affordances of abundant intelligence and agentic automation is reaching an exciting and productive stage. It is analogous to developing the human / organisational software that leads us towards programmable organisations, and the wonderful thing about it is the best people to lead this effort are not external consultants or AI forward-deployed engineers, but leaders of existing functions within our organisations.