What are AI agents in RevOps?

AI agents in RevOps are AI systems that carry out routine work on the boundary between sales, finance and account management: updating CRM fields, investigating discrepancies between order and invoice, preparing renewals, following up on payments. They use a language model to decide, situation by situation, which step is needed, and carry it out through connected systems. Their value lies in the work that currently falls between departments, and their risk in the permissions they are given to do it.

Why RevOps in particular?

Revenue operations is about the seams between departments. Sales closes a deal in Salesforce or HubSpot. Finance invoices from Exact, AFAS, Xero or NetSuite. Account management handles renewals and expansions. Each department does its job well. Things go wrong in the handover:

• A deal is closed with a non-standard payment term that never reaches the order system.
• Additional work is approved verbally and never entered as an order line.
• A customer expands from 40 to 55 licences, but the invoice stays at 40.
• A contract expires, nobody prepares the renewal, and the customer carries on using the service on the old terms.

This is work nobody fully owns. It is repetitive, but just too varied for a fixed automation. Exactly the kind of work where an agent adds something. The wider context of why these seams cost revenue is covered in why silos cause revenue leakage.

Which RevOps tasks suit an AI agent?

Not every RevOps task suits an agent. These five do, because they involve a lot of look-up work, have a clear end point and produce an outcome that can be checked.

1. Investigating discrepancies between CRM and billing

A reconciliation shows that the deal value in the CRM is EUR 48,000 and the invoiced annual value is EUR 42,000. The agent looks up the deal, the order lines, the invoices and the relevant notes, and writes an explanation with sources. An employee decides what should happen. This is the agent-driven part of CRM-to-billing reconciliation.

2. CRM hygiene

Missing contract data, outdated close dates, deals that have stalled for months, duplicate accounts. An agent can make proposals: this deal has been in the "negotiation" stage for 140 days without activity, proposal: mark as lost. The account manager confirms with one click.

3. Preparing renewals and indexations

Ninety days before a contract end date, the agent prepares a file: current terms, usage, open tickets, indexation clause, proposed new price. For contracts with an indexation clause it adds the calculation.

4. Accounts receivable follow-up

Reminders for outstanding invoices according to a schedule, with exceptions set aside: a customer with an ongoing dispute, a customer who has just placed a large order, an invoice with a credit note against it.

5. Passing on signals

A support ticket about a feature that is not in the customer's licence is an expansion opportunity. An agent can recognise such signals and create a task for the account manager, with context.

Which tasks should an agent not do?

What you do not let an agent do is at least as important, particularly at the start:

• Setting prices or discounts. Commercial decisions stay with people.
• Changing master data in the accounting system. A wrong VAT number or bank account number is hard to reverse and can cause damage outside your own company.
• Creating or sending credit notes. Money leaving the business needs a person.
• Negotiating or making commitments on someone's behalf. An agent that emails "agreed" creates obligations.

The distinction between proposing and deciding is the core of it. It is worked out in AI recommendations vs AI decisions.

How is a RevOps agent built?

An agent is a language model with tools. For a RevOps agent the tool list might look like this:

What stands out: almost everything is read-only. Writing is limited to places where a mistake is visible and does no harm outside the company. That is not a coincidence but a design choice. How the model calls these tools is explained in what is tool calling, and the basics of agents in what is an AI agent.

Worked example: licence expansions that are not invoiced

Worked example: suppose you are a software vendor or MSP with 250 customers on a licence model. Usage is measured in the platform, invoicing runs in the accounting system, and updates are made by hand at every invoice run.

Each month an agent compares measured usage with invoiced quantities and sets discrepancies aside for the account manager. Suppose it finds 18 customers using an average of 6 more licences than they are invoiced for, at EUR 35 per licence per month.

• Per month: 18 x 6 x EUR 35 = EUR 3,780.
• Per year: EUR 45,360.

That amount did not appear in any report. It sat in the gap between two systems that were each correct on their own. These figures are assumptions; the mechanism itself is very common in practice. More examples of this type are in revenue leakage from wrong subscriptions.

Where does it go wrong?

The agent is given an employee's account. It can then do everything that employee can, and you can no longer tell in the CRM what the agent did and what the person did. Give an agent its own account.

Too many tasks at once. An agent that does "everything in RevOps" cannot be tested or checked. Start with one task, measure the outcome, then expand.

No owner for the outcome. An agent making a hundred proposals a week that nobody looks at only moves the problem. Every proposal must land with a person.

Poor source data. An agent updating the CRM from an outdated price list spreads the error faster than a person ever could.

No monitoring. An agent that makes the same mistake for a week is something you want to see on day one, not at the quarter close. Why that matters so much is covered in AI monitoring explained.

How do you set up your first RevOps agent?

1. Choose one leak or one task that involves a lot of manual work and whose outcome can be measured in euros. Discrepancies between order and invoice are a good first candidate.
2. Measure the starting point. How many cases a month, how much time per case, how many are left untouched.
3. Make the tool list as narrow as possible. Read-only by default. Write only where necessary, and then in a limited form.
4. Decide who approves. For each type of action: who sees the proposal, how quickly, and what happens if nobody responds.
5. Run a trial period in which the agent only proposes. Compare its proposals with what your team would have done itself.
6. Only then let it carry out actions where that is safe, with a step-by-step log.

Where do agents fit in the bigger picture?

Agents are the executing layer. Before them there has to be a layer that connects the data and finds discrepancies, and a layer that prioritises by euro impact. Without that foundation an agent does fast work in the wrong place. The full picture is set out in how does AI work within Revenue Intelligence.

Frequently asked questions

Do AI agents replace the RevOps team?

No. They take over the look-up and preparation work. Decisions, customer conversations and agreements about processes stay with people. In practice a team gets time back for the work that is currently left undone.

Which systems does a RevOps agent need to work with?

At a minimum the CRM and the billing or accounting system. Often also the place where contracts are stored and the mailbox, to read arrangements. Every connection is an extra permission, so add only what the task needs.

How quickly does a RevOps agent deliver results?

That depends on the task and the quality of your data. An investigating agent that only reads can be deployed safely sooner than an agent that writes. Measure what the task costs now, so you can see afterwards what it delivers.

May an agent email customers?

Technically it can. Start with drafts that a person sends. Only once you know how the agent behaves can you consider letting it send by itself for a narrow, predictable category, such as a first payment reminder with a fixed text.