The conversation about AI agents has got stuck on the tool. Which model, which integration, how much a licence saves. And the question a CEO actually has in front of them is a different one: what happens to my organization when part of the work is done by something that decides on its own.
Agentic Enterprise, by Kashi KS, Giridhar LV Vishwanath and Thiyagarajan M, answers exactly that. We came to this reading through David Boronat's reflection on the book, and these are the five ideas that carry most weight for business owners.
1. AI moved from tool to colleague
A program gets executed. A colleague gets asked for something. That shift looks like a nuance and it changes everything: how work is shared out, how decisions get made, and who answers for the result when something goes wrong.
The book sorts it into three phases. In late 2024, automations appear that can complete one specific task from start to finish with nobody on top of them. In spring 2025 the focus moves to the whole flow, with several coordinated steps, which is the orchestration stage. And today we are in the third: the agent that receives an assignment decides what to do with it, solve it, pass it to another agent, or ask a person to resolve it.
With that comes a structural condition rather than a luxury: the organizations that come out ahead will be the ones that actively decide to protect their teams' sense of purpose in the work.
How is an AI agent different from ordinary automation? An automation runs the steps somebody programmed. An agent receives an assignment and decides for itself how to resolve it, including when to stop and ask for human judgement. The practical difference is that one gets configured and the other gets directed.
Before you read on, if you want a snapshot of where your company stands today: the free AI Readiness assessment is 15 questions and 3 minutes, and it ends with the three priorities worth looking at first.
2. The unit of analysis is the task
Looking at the employee is looking in the wrong place. Looking at the department, too. What needs opening up is the role, a container whose name matters far less than the real tasks it holds inside. The tasks are what count.
Open it up and the split that organizes the whole book appears. Around 10 % of the work is automatable today: invoice processing, structured data analysis, scheduling and planning. Another 30 % is hybrid, where the agent resolves seventy to eighty per cent and the person keeps the judgement and answers for the outcome. And the remaining 60 % stays entirely human: judgement, relationship, empathy and decision. That is where a company's real competitive advantage lives.
The credit analyst is still there. What breaks apart is the role: one part gets automated, another gets reinforced and another stays exactly as it was.
The method is simple to explain and laborious to apply. List the real work of a role, rather than what the job description says, and score each task from 0 to 10 according to how well an agent would do it today. From 7 to 10 it is a candidate to move to the agent with defined supervision. From 4 to 6 it gets shared, with the agent proposing and the person validating. From 1 to 3 it is preserved and invested in.
That is where the distance between automating and redesigning shows. Automating asks how to run the existing process faster. Redesigning asks how it would be built starting from the end result.
Where do you start when working out what an agent can do? With the tasks, one by one, rather than with whole roles. You list the real work of a role, score each task from 0 to 10 according to how well an agent would do it today, and the result splits that role into three blocks: agent-ready, hybrid and human.
Before deciding on any task, three questions: what removing it is really worth, counting the actual saving rather than the theoretical one; what gets lost by taking the person out; and what happens to whoever was doing it. And a fourth that almost no leadership team can answer: who hires agents in your company, IT, the people function, or each unit on its own. An agent is a decision about how work gets shared out, more than a software licence.
3. The middle layer gets thicker
The idea going around is that agentic organizations will be flatter. The book argues the opposite: bringing in agents concentrates the need for human judgement, rather than removing it.
A flat organization without agents is a stressed organization. With agents and without governance it is an organization in chaos, with people and agents fighting over the same decision authority. What appears in the middle is an orchestration layer that sets each agent's decision limits, watches that the system does what it should, and steps in when values come into play.
Hence the change of shape. Yesterday's pyramid had execution at the bottom and decisions at the top. Today's diamond has a narrow base and a thick centre. There is a lot of talk about reskilling workforces, and the talk should be about rebuilding the middle layer.
Do AI agents flatten the organization? Quite the opposite. They automate execution at the base, and in doing so they thicken the middle layer, which takes on orchestrating the work between people and agents, setting decision limits and stepping in when values are at stake. The pyramid turns into a diamond.
4. Governance is the real bottleneck
When an agent gets it wrong, who answers for it? The model provider, the company that deployed it, or whoever set its parameters? Writing it down afterwards turns the problem into a design failure, rather than a legal one.
The book proposes five layers: guardrails, the limits the agent carries built in; traceability, with a record of what it saw, which rules it applied and with what confidence; escalation protocols, defining when it stops and asks for human judgement; kill switches, to stop the system if it starts drifting; and periodic review of outputs.
On those five layers rests the split of responsibility, which gets classified rather than settled by general trust. A delegated decision is taken by the agent alone, such as approving a 500 euro transaction with established history and clear parameters. An assisted decision is analysed and recommended by the agent, and signed off by the person. And a reserved decision is exclusively human: anything touching people, values or the client relationship at a critical moment.
The deployment rule is to measure the agent's reliability, rather than compare it with a person: what level that task demands, and whether the agent can reach it in a reasonable time. The expensive and frequent mistake is deploying a phase one agent as if it were already in phase three. It fails, the technology gets blamed, and the whole project gets pulled.
Who is responsible when an AI agent makes a mistake? The company that deployed it, and the way to contain that is to write it down beforehand: decision limits, a record of what it did, when it stops to ask for human judgement, and how the system gets stopped. Deciding it afterwards turns it into a design failure.
5. Almost everyone is optimizing for the small prize
This is the figure that is hardest to swallow in a leadership meeting. The saving from the automatable 10 % is visible, quick and easy to defend, which is why it is the only thing being measured. The augmentable 30 % is worth, according to the book, three to five times more, and almost nobody models it, because it is bigger, less visible and slower to materialize.
The real return adds both together, and that second term shows up as speed in understanding what is going on, capacity to adapt and organizational learning.
From there come the three company models the book describes. The Efficiency Enterprise, the most common one, uses agentic deployment as a cost reduction lever: it wins in the short term and can leave structural damage in the long. The Augmented Enterprise takes reinforcing its teams seriously and builds its operating model around the relationship between people and agents. And the Adaptive Enterprise, which is the ceiling, captures all three levels: the saving, the augmentation and the new capabilities that only appear when people and agents work together.
The book leaves the winner open and argues that the difference lies in having chosen deliberately, rather than by inertia. Its horizon is 2028, and it says the winner will be whoever chose their model consciously between 2025 and 2026 and executed it with discipline. We are in 2026.
What is the most common mistake when deploying AI agents? Measuring only the saving on automatable work, which is the small, visible part. The big value sits in the work that gets reinforced, where the agent resolves most of it and the person keeps the judgement, and that is worth three to five times more.
How to start
Five steps, in this order. Start with a single team, which is the counterintuitive part, and learn there what happens when you put people and agents together with clear decision limits. Build the agentization map layer by layer: the most augmentable layer is your opportunity and the most human one is your advantage. Choose the pilot process well, with high volume, an identified bottleneck, clear metrics and low regulatory sensitivity. Talk to people before touching processes, because the technical redesign is the easy part and whoever sees their authority move will resist. And diagnose your starting point honestly across data, systems, processes, culture and governance: it is enough to know what you have and what you lack.
Three questions the book leaves open
These are the ones that come up most in projects and have the least answer today.
The pipeline paradox. If that automatable 10 % is precisely the routine work that used to train junior profiles, and the base of the pyramid narrows, where will the thick centre of the diamond come from five years out?
How performance gets measured now. How do you evaluate somebody whose output comes partly from an agent? Objectives, variable pay and appraisal are all still designed for a world where the whole output was the person's.
What a team is made of. The ratio of people to agents, who sits in the meetings, how the work gets coordinated. A great deal has been written about tasks and roles, and almost nothing about the team as a unit.
All three stay open. What does exist is a way of working them through inside a real organization, with its roles, its data and its governance. That is what we do at DO'IN TALENT.
If you want a quick picture of where your company stands before deciding anything, the AI readiness assessment takes three minutes and you leave with your three priorities.
A reading of Agentic Enterprise, by Kashi KS, Giridhar LV Vishwanath and Thiyagarajan M, prompted by David Boronat's reflection on it. The 10 / 30 / 60 figures and the three to five times multiple are from the book.
AI Readiness
If you want to see where you stand, start here: 15 questions on data, tools, culture and governance. It takes 3 minutes.
Take the diagnostic →Book your session, no cost and no commitment.
Book your session →