"We give our employees Gemini accounts, too." Some companies may respond this way when the conversation turns to AI. They may already use it to draft emails or summarize documents, and feel that it has made their work easier.
The next question is how to make AI part of the flow of work. In agentic operations, AI selects the information and steps needed for a goal, brings in human judgment along the way, and proceeds through authorized actions and verification of the results. Let us look at what changes through the example of a request to adjust a delivery date.
Asking AI to write still leaves the work before and after it
Imagine a company that sells parts. A customer asks for an order to arrive earlier than planned. The person handling the request can ask AI to draft a polite reply in moments. But there is still work to do before they can decide which delivery date to put in it.
They need to find the order in the sales management system and check the inventory. They need to find out when parts already on order will arrive, and ask the purchasing team about any shortfall. If part of the order can be delivered first as a split shipment, they must also consider whether that suits the customer and who will pay the additional shipping cost.
After replying, they need to update the order record, notify the warehouse, and confirm that everything is in place to deliver as promised. Writing the reply faster still leaves the person responsible for gathering information, coordinating with others, and checking that the work is complete.
This is where there is room to consider the next use of AI. It is not a comparison of products such as Gemini, or a claim about the limits of their features. Even with the same AI, the scope of what we can delegate changes depending on whether we use it to write, or connect it to business systems and internal rules so it can move the work forward.
Agentic operations bring AI and people together to complete the work
In this article, "agentic AI" means AI that chooses steps and tools according to a goal, then uses the results to decide what to do next. "Agentic operations," as we use the term here, means a way of running work that delegates information gathering and actions to that AI and combines them with human judgment and approval.
An article by Genpact describes operations that combine execution by AI agents with human oversight, exception handling, and improvement. A Deloitte report also discusses the need to rethink roles and the flow of work so that people and AI can collaborate. Both share the view that adopting AI calls for designing how work gets done.
Traditional automation, which follows predefined conditions and steps, is another way to connect business processes. Where the procedure is clear, it can continue to serve that purpose. Agentic AI can take on the role of investigating further when information is missing, or considering another approach when conditions change along the way.
Running several processes in sequence does not make all of them agentic. Anthropic's explanation also sets out this distinction and how the approaches can be used together.
How handling a delivery change could work
With more of the work delegated to AI, the person responsible can join the process once the information and options needed for a decision are ready. To make that possible, we first need to define which information AI may consult, which actions it may use, and when it must seek approval.
The following illustrates how this could work once those connections and permissions are in place. It does not describe an implementation at a particular company or benefits that follow simply from adopting AI.
Compare orders with inventory and identify shortages
When a customer requests a delivery change, AI checks the relevant order, current inventory, and expected arrivals. It treats inventory already allocated to other orders separately from stock that is available to use.
For example, if an expected arrival date in an email differs from the date in the system, AI cannot simply give the customer a delivery date. It shows the discrepancy and the source of each piece of information, then checks with the relevant person within the communication scope agreed in advance. Until that is confirmed, it must not assume the delivery will be on time and move ahead.
The aim is to let the person address missing information and points that need confirmation, instead of gathering every piece of information themselves each time.
Compare options and bring the decision to a person
Once the necessary information is available, AI develops options such as delivering the whole order at a later date, or shipping the available stock first and the rest after it arrives. It compares the customer's preferences, whether split shipments are acceptable, shipping costs, and the effect on other orders, then sets out the conditions for each option.
It then makes clear where human judgment is needed. For example: "A split shipment would let us deliver part of the order on the requested date. May we cover the additional shipping cost?" The person responsible decides, taking the customer relationship and profitability into account.
Having AI produce options alone can increase the amount of checking people have to do. When handing the work over, it is important to include the order and inventory records behind the proposed decision, along with anything that remains uncertain.
Carry out approved actions and verify the results
Once the person responsible approves an approach, AI prepares a draft reply to the customer and the proposed order changes. It obtains the approval required to send an external message, sends the reply, and waits for the customer's agreement on any terms that require it, such as a split shipment. Only after that agreement and the necessary internal approvals are in place does it change the order through authorized system operations. If the conditions change, it returns to the person responsible for a decision.
Attempting an action does not make the work complete. AI reads the updated order again to verify that the delivery date and quantities match what was approved. If sending the message or updating the record fails, it tells the person responsible which steps remain incomplete and what needs to happen next.
Finally, it reports what it checked, who approved what, and which records it updated. The person responsible can trace the results and take over work that stopped partway through. Only when these steps are included does handling the delivery change become a complete, connected piece of work.
Find work that needs attention before someone asks
Work does not always have to begin with a request or notification from a person. If we define what to check and how often, we can also consider delegating the search for inconsistent information or external changes to AI.
For example, internal product documentation may have been updated while sales materials still contain old specifications. AI compares the documents it is allowed to access and shows the differing statements and their update dates. After confirming with the person responsible which information is correct, it prepares revisions and updates the approved documents. A difference between two sources must not automatically make one of them the correct version.
In security work, one example is comparing new information about software vulnerabilities with a register of the products and versions used within the company. AI identifies systems that may be affected, notifies the person responsible with supporting evidence, and prepares possible responses. Changes to production systems or taking them offline require that person's judgment and approval.
In either case, AI does not automatically know everything inside the company. Connecting it to the sources it should check, granting access permissions, and setting up regular checks are what allow it to find work that needs attention and help move it forward.
People decide the goal and how much to delegate
For AI to act autonomously, it needs a target outcome and conditions to follow. People define the purpose, scope, and rules. They also handle important decisions, approvals, and exceptions that the rules cannot resolve.
In the delivery example, the goal may be to meet the customer's request as far as possible. Alongside that goal, we set conditions: do not reallocate stock reserved for other orders; consult the person responsible about additional costs; communicate commitments to the customer only after approval. AI works out how to investigate and what options to propose within those boundaries.
We also need records that people can inspect and a way to stop the work when necessary. Who should AI return to if it makes an incorrect judgment, or if the information is too old for it to be confident? If that person is unavailable, at what point should it wait? These arrangements are part of designing the process, too.
We also discuss how to decide what to delegate in "Which Work Deserves Our Best Intelligence?". Not every task needs the same degree of freedom.
Redesign one business process using existing systems
To begin, we recommend choosing one business process and reviewing its flow through to completion. A process like handling a delivery change is a candidate: gathering information takes effort, and the points that need human judgment and the conditions for completion can be explained.
Start by writing down the following with the people who do the work:
- What triggers the work, and what state counts as complete?
- Which systems and documents do they consult, and what do they confirm with whom?
- Which actions can be delegated to AI, and which require human judgment or approval?
- If information is missing or something fails, who takes over, and how?
This review should also consider whether repeated checks or unnecessary copying of data can be removed, rather than simply transferring the current steps to AI. It means focusing on one business process while looking at it from start to finish.
Next, check how the existing systems can be connected and what permissions are available. If APIs, interfaces for reading and writing data, or other connection methods are available, we can consider keeping the current systems as the place where business records are stored and letting AI handle the work between them. If there is no way to connect, or the data is not ready, that preparation needs to come first.
We explain these interfaces in more detail in Giving AI "hands": why internal system APIs matter now.
During a trial, we can start by delegating information gathering and the preparation of options, with a person checking the results. Evaluate the time from request to completion, the burden of checking on the person responsible, corrections, and missed actions, as well as how quickly text is produced. Include the cost of AI usage and the effort needed to operate it, then adjust how much to delegate.
Enabling employees to use AI is one part of that preparation. Next, choose one process in your company and write down where people currently connect information and where they need to exercise judgment. That will help you see what needs to be put in place to carry the work forward together with AI.