Checking open relocations as an agent’s first job

A chatbot answers the question you type. An agent takes an assignment, breaks it into steps, uses tools and hands back a result. It can work when someone asks, on a schedule, or when something happens in a system. Once it runs every day, it behaves like a digital co-worker with a narrow remit. The fastest way to learn what that means in your company is to give one agent one job and watch it closely. Checking open relocations is a good first job for an ISP.

Why relocations

Every ISP handles relocations, and they quietly go wrong. The customer reports the move, customer service places the order, and the wholesale partner comes back with a question: the floor, the apartment number, the ID of the fibre socket, a contact who can open the building. That question lands in a shared mailbox or as a status code in the partner portal. Nobody is waiting for it. The customer moves in, the line is not active, and customer service learns about the problem when the customer calls.

The case touches ordering, technical feedback from the network or wholesale partner, billing and sometimes fault handling. A person has to look in three or four places to see that it is stuck. That makes it a good candidate for an agent: the rules are clear, the result is easy to check, and an agent that only reads and reports can do very little harm.

The job, written down

Write the assignment as you would for a new colleague. For example:

Every working day at seven, before customer service starts:

• List every relocation order that is not yet completed.
• For each one, compare the order status in ordering with the latest technical feedback: partner status, confirmed activation date, field appointment.
• Flag the cases where the two disagree, where the partner is waiting for information, or where the move date is close and there is still no activation date.
• For each flagged case, state what is missing and who has to get it: the customer, the partner, or an internal team.
• Put the list into the customer service queue.

Then write down what the agent does not do in this first phase. It does not call or email customers, change orders, or touch billing. It reads, compares and reports.

A single entry on the list might read: relocation for the customer moving to Via Ronco, move date next Thursday, order status “submitted”, partner status “waiting for customer information” since Monday, missing the apartment number and the fibre socket ID, to be obtained from the customer by customer service. That is enough for a colleague to make one phone call and close the gap before the customer notices it.

Scope it before you build it

The first version has read access only. It prepares suggestions, and people act on them. This keeps the cost of a mistake at one wrong line on a list, and it gives you weeks of results to judge before the agent changes anything.

Give the agent one person who is responsible for it, usually the team lead in customer service. For the first weeks she reads the list every morning, marks wrong or missing findings, and corrects the instructions. Microsoft’s Work Trend Index 2026 frames this as three questions that every “Frontier Firm”, its term for the most advanced adopters, will need to answer: “Who reviews agent performance? Who has the authority to update the workflows that agents run?” and how a local win gets captured and scaled (Microsoft, 2026). In the same report’s survey of 20,000 AI users in ten countries, quality control of AI output was the human skill named most often as becoming more important. Write the owner’s name next to the agent’s name.

The agent needs the same information a careful person would use, from the system that leads for each part:

• Ordering: relocation orders, move date, old and new address, product.
• Technical: wholesale order status, partner feedback messages, activation appointment, line test results at the new address.
• Billing: whether the account is in good standing, open dunning that could hold up the order, the change of invoice address.
• Faults: open tickets at the old or new address that explain a delay.

Point the agent at the leading systems, never at exports or spreadsheets. If a stale copy of an order status survives somewhere, the agent will find it and report a problem that was solved yesterday. Whether a person or their agent handles a case, both should reach the same information, each within their own rights.

What done looks like after the first weeks

After the first few weeks, sit down with customer service and the technical team and go through the lists together. The pilot is working when:

• The owner trusts the list. Wrong findings are rare, and she can explain each one.
• Customer service starts the day from the list and contacts customers before they call.
• Stuck cases are visible by reason: missing apartment number, partner not answering, open dunning, missing appointment.
• The instructions have been corrected several times, and the current version is written down where the team can read it.

If the list is still noisy, the cause is usually the data. Two systems disagree about a status, or a partner message is not stored where the agent can read it. Fix that first. The same fix helps every person who looks up a relocation.

When to grant write rights

Grant write rights after the results have been checked, one right at a time. The first ones should be small and easy to undo:

1. Add a note to the order listing what is missing.
2. Create a task for customer service with the customer’s contact details attached.
3. Draft the email asking the customer for the floor and apartment number, for a person to review and send.
4. Later, send that request without a person, once the drafts have been right for long enough.

Cancellations, credits and contract changes stay behind an explicit release by a person. Every right is granted to the agent’s own identity, recorded, and can be withdrawn. The MCP authorization guidance recommends authorization whenever you need to audit who performed which actions, and advises least-privilege scopes, split per tool or capability (MCP, 2026). The Work Trend Index makes the same point for IT, describing agents as “managed entities with identities, permissions, policy enforcement, and lifecycle management”. Before you need it, test that revoking a right takes effect and that a running job can be stopped.

What to measure

Take a baseline from the relocations of the last few months before the agent starts, then measure the same things each week:

• Where cases get stuck: the number of open relocations by reason and by step.
• How often customers have to ask again: contacts per relocation after the first call.
• Processing time: days from the reported move to an active line at the new address, and hours from a missing detail appearing to someone requesting it.
• How well the agent does: the share of flagged cases the owner confirms as real.

Agree with your team, or with your partner if one runs the agent for you, how results are checked, who fixes errors, and who decides the next step. Extend the agent’s remit when the flow works reliably, starting with the next process that has the same shape, such as port-ins waiting for the donor operator.

What METAVSHN does for this

An agent like this needs defined access to the same record the team works on. The METAVSHN CLI gives an agent commands to look up customers, products, services, orders and invoices on the METAVSHN platform. Data commands can return JSON, and each command has a contract that lists its inputs, the type of operation and permission hints, so the agent can check what a command does before it runs it. METAVSHN checks access on every request, against the identity and rights behind the agent. Anything that deletes or sends something out needs explicit confirmation, and a dry run shows the request first. The CLI works with any agent that can run commands, so you can use the one your company already approved. It is in early access.

If you want to give an agent its first job on your own records, request early access to the METAVSHN CLI.

Sources

All sources accessed on 31 October 2026.

1. Microsoft WorkLab, “2026 Work Trend Index Annual Report: Agents, human agency, and the opportunity for every organization”, 5 May 2026. microsoft.com
2. Model Context Protocol, “Understanding Authorization in MCP”, protocol revision 2026-07-28. modelcontextprotocol.io