How to Work Out Which Part of Your Company's Work an AI Agent Can Do

Almost every company has already tried some AI tool, but few have changed how they are organised and how their work is organised.

When you decide to build agents into an organisation designed for people to do every task, a faster company is hard to come by: all you get is a company with two layers of work that aren't always in step. All you get is employees handing tasks to an agent, with no process for reviewing the output and no protocol for when something goes wrong.

The book Agentic Enterprise, by Kashi KS, Giridhar LV Vishwanath and Thiyagarajan M, sums it up in one clear idea: the organisations that win in 2028 won't be the ones that deployed agents first, but the ones that designed themselves best around them. And it places the window for deciding this deliberately between 2025 and 2026. In other words: now.

These are the four steps we recommend to get started.

1. Think in tasks, rather than in job titles

The wrong unit of analysis is the employee or the department. The right one is the tasks that make up each role: not the ones in the job description, but the ones in each person's real work.

No credit analyst is going to be "replaced by an agent". Their role will be broken down: part of it gets automated, part gets reinforced and part stays exactly the same. As a reference point, the book puts today's split into three blocks:

• Around 10% fully automatable tasks: the agent handles them end to end.
• Around 30% hybrid tasks: the agent handles 70-80%, and the person keeps the judgement and owns the outcome.
• Around 60% of tasks that will remain purely human.

A method that works very well is this: list the real work of a role and score each task from 0 to 10 on how well an agent would do it today. From 7 to 10, a candidate to move to the agent. From 4 to 6, hybrid. From 1 to 3, protect it and invest in it.

2. Choose the pilot on criteria, rather than on enthusiasm

The initial instinct is to start where someone is keen. There's logic to that, but another criterion works better. The processes with the most potential combine five traits:

• High volume of activity.
• An identified bottleneck.
• Clear success metrics.
• A good proportion of agent-suitable work.
• Low regulatory sensitivity.

And a principle that will save you months: pick one team, learn what happens when you put people and agents together with clear decision boundaries, and only then roll it out, rather than reorganising the whole company from the start.

3. Decide governance before deploying — afterwards is too late

When an agent makes the wrong decision, who answers for it? The provider that built the model, the company that deployed it, or the person who set its parameters?

If that isn't written down before deployment, it isn't a compliance problem: it's an organisational design failure. The minimum viable setup is five layers:

• Guardrails: what the agent can and cannot decide.
• Traceability: what it saw, which rules it applied and with what confidence level.
• Escalation protocols: when it stops and asks for human judgement.
• Kill switches: how the system is stopped if it drifts.
• Periodic output review: the early signal that something is changing.

And underneath, an explicit split of decisions:

• Delegated decisions: the agent decides alone, with clear parameters and an established track record.
• Assisted decisions: the agent analyses and recommends; the person decides and owns it.
• Reserved decisions: human only.

There's also an upstream decision almost nobody has made: who hires agents in your company? IT, the people function, or each business unit on its own? An agent is not a software licence. It's a decision about how work is shared out.

4. Talk to people before touching the processes

The technical redesign is the easy part. The organisation, as a political system, resists: whoever sees their authority shift will defend it, and whoever sees their role change will feel lost without clear, proactive communication.

This will affect middle managers above all. They've spent ten years learning to synthesise information, report when it matters and decide under ambiguity, and they may get the impression that an agent does part of that better and faster. It's essential to say from the outset that their role isn't redundant: it changes. It becomes the orchestration layer, the one that sets decision boundaries, watches that the system does what it should, and steps in when the organisation's values are at stake.

The conversation has four parts, and it comes before the project, not after:

1. Your role is changing.
2. This is what it becomes.
3. This is what we need from you.
4. This is how we'll measure that it works.

And one question that particularly concerns us

That automatable 10% is, in most organisations, exactly the routine work that junior profiles used to learn on. If you narrow the base today, where will the middle layer that holds all of the company's judgement come from in five years? How do we keep training our most junior people?

We'd be glad to have this conversation at your company.

How we do it

At DO'IN AI we work through exactly this sequence: we list the tasks in each layer of your organisation and classify them as automatable, augmentable and human; we choose the pilot process with you on objective criteria; we define the governance framework and who decides what; and we prepare the conversation with the teams before the redesign begins.

The savings from automating are undeniable, but the value of communicating, preparing your teams and reinforcing the importance of their judgement is three to five times greater.

Free diagnostic

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 →