A box of twelve parts
Imagine a purchasing team comparing two quotes. One supplier offers a box of twelve parts for €48; another offers the same component for €4.30 per part. Leaving shipping, tax and other conditions aside for this example, the first quote works out at €4 per part.
An AI agent comparing only the numbers printed on the quotes could recommend the wrong offer. A person corrects the comparison and the purchase moves forward. The company can also gain something useful for tomorrow: a method that identifies the unit of sale, checks the pack quantity and makes prices comparable before recommending a choice.
The difference becomes visible in the next request. A quote arrives at €39 per box, with no pack quantity. Does the system retrieve the criterion and ask for the missing information, or start comparing incompatible figures again? This second case helps establish whether the correction actually changed the work.
The example is hypothetical, but it makes a business question tangible: what capability remains available after an agent finishes a task?
What are we delegating?
An AI agent can use tools and information to pursue an objective: finding documents, comparing quotes, preparing a reply or updating a connected system. What it can do depends on its available tools, the context it can reach and its actual permissions.
The delegated work needs a recognisable outcome. In purchasing, that could be a reasoned comparison with its sources and outstanding questions. Preparing the comparison and placing an order are different activities, with different consequences and permissions. The first can be delegated while the second retains an approval step.
The arrangement can also vary within an activity. Routine cases may follow agreed criteria; an anomaly or missing information may call for a person. These arrangements should fit the work, without assuming every process should eventually become fully autonomous.
Italy provides useful context for this discussion. Istat reports that 16.4% of enterprises with at least ten employees used at least one AI technology in 2025, up from 8.2% in 2024. These figures measure AI adoption. They do not measure autonomous agents or describe all microbusinesses.
Resuming work with its reasons intact
A business case rarely lives within one session. A colleague is waiting for a reply, a supplier updates a quote, or someone takes over from another team member. Continuing requires the objective, relevant sources, decisions already made and the reasons supporting them.
In the purchasing example, “prefer supplier A” would be fragile memory. It is more useful to preserve that, on a particular date and for a particular quantity, its unit price was lower while delivery still needed confirmation. A new quote or a different deadline could change the choice.
This continuity requires knowledge maintained where the work can retrieve it, with responsibility for updates and appropriate access. Saving every conversation can increase the available material without making the important parts clearer. A useful handover makes it possible to see which reasons still hold and which need reconsideration.
When a correction becomes expertise
Here, reusable expertise means a reachable combination of knowledge, sources and criteria that changes how a situation is handled. The correction about pack pricing can become part of the method for comparing quotes, together with its rationale and conditions of use.
That improvement can involve updating knowledge, instructions and tools while retaining the same AI model. Technical documentation describes mechanisms for memory across sessions and reusable specialist resources. The existence of those mechanisms does not establish that a particular application uses them effectively.
The useful test remains a new case. Is the criterion retrieved when needed? Is it applied within its scope? Does the system recognise missing information? We also need to check that the correction has not made a method too rigid to consider quality, availability and delivery time.
Maintaining this knowledge is part of the work. An outdated procedure or invalid source can make a previously useful correction harmful.
Expertise can change the question
A purchasing comparison may involve price, quality, availability and technical constraints. Bringing these perspectives together may reveal that the cheapest component requires different maintenance, or that a proposed substitute is unsuitable for the equipment.
The initial question can then change: which alternatives deliver the required operational outcome? One agent or several agents may contribute to that understanding. The technical arrangement should reflect the usefulness of their contributions and the cost of coordination.
Customer support offers another possibility. If many requests concern the same installation step, better individual replies are one useful outcome. Investigating the causes may suggest a preventive guide, a change to onboarding or an assisted installation service. These are possibilities to assess against the company's sources and circumstances; observing recurring requests does not by itself establish which solution will work.
A trial that reveals the difference
A useful starting point is a bounded activity with an owner, accessible sources and a comparable outcome. Familiar cases help expose errors and remaining human work; different cases help show whether the method holds beyond the examples used to correct it.
Measure the total effort needed to reach a correct result, including human checks, clarification and rework. Other useful observations include the recurrence of corrected errors, the time needed to resume a case and the ability to flag an exception. Two agreeing answers may rely on the same incorrect source: consequential decisions need evidence capable of contradicting them.
Access should fit the mandate. If the job is to compare quotes, permission to place orders may be unnecessary. Permissions enforced in connected systems remain decisive even when the agent has received appropriate instructions. OWASP describes this relationship in its excessive agency risk. If an operation is interrupted, establish what already happened before repeating it.
The person reviewing the work needs time, sources and the ability to intervene. Formal confirmation of hundreds of proposals does not make supervision effective. The NIST AI Risk Management Framework provides a reference for connecting responsibilities, evaluation and ongoing monitoring.
What remains in the company?
At the end of the trial, consider both the result and the capability it produced. Was the case completed correctly? Is the remaining work sustainable? Can its reasons be recovered? Does a useful correction change a subsequent case? Have possibilities emerged that deserve a different experiment?
Time released is available capacity: its value depends on how the organisation can use it. The calculation should include integration, checks and knowledge maintenance. The accompanying Italian report develops these points through three business cases, an explicitly illustrative economic scenario and a worksheet for choosing the first activity.
A concrete next step is to select a familiar case and identify the outcome to delegate, the decisions that remain with people and the knowledge that should be available next time. From there, observe whether the work delivers a useful result and makes the company better able to continue.
Report and references
Download the Italian report: AI Agents in Italian Companies, R2, 10 pages.
The business examples are hypothetical. The possibilities described are design and trial criteria, rather than observed results from a company.
Primary sources are linked where relevant. A further reference on delegation and supervision is Evan Kormos's presentation, How to Know When an AI Agent Can Work Unsupervised, published by AI Engineer. This article develops its own argument about business work.