Which functions in a mid-sized company are best automated with AI agents first?

Start with finance and accounting, then procurement and vendor management, then customer support. These are the functions where work arrives as documents, follows written rules most of the time, and gets retyped from one system into another. Most companies start with marketing and sales instead, and the evidence says that is where the money goes in, not where it comes back.

Most companies start in marketing and sales, and most see no return

Eurostat (2025) found that about 20 percent of EU enterprises used AI in 2025, and 30 percent of medium-sized ones. Among the users, the most common purpose was marketing or sales (34.7 percent), followed by business administration (31 percent).

The return has not followed the spend. IBM's 2025 survey of 2,000 CEOs (IBM Institute for Business Value, 2025) found that only 25 percent of AI initiatives had delivered the expected return, and 16 percent had scaled across the enterprise. MIT NANDA's 2025 report on generative AI in business, as reported by Fortune (2025), found that more than half of generative AI budgets went to sales and marketing tools while the largest returns came from back-office automation.

Our view: companies choose the reverse of the order that pays. Sales and marketing feel urgent and the vendors are loud. The money is in the back office.

Three tests decide the order

Does the work arrive every week in a similar shape? Would two people with the same rulebook reach the same answer most of the time? Does a human retype the output into another system? A function that passes all three is ready for an agent. A function that fails two needs process work first.

BCG (2024) surveyed 1,000 executives and found that the leaders put 10 percent of their resources into algorithms, 20 percent into technology and data, and 70 percent into people and processes. The three tests find the functions where the 70 percent is already done.

Finance and accounting passes all three tests almost everywhere

Every invoice, expense claim and bank line is already a document. The approval rules are written down, because the auditors asked for them. The manual work is matching, coding and retyping, which an agent handles well.

Start with accounts payable intake and three-way matching, then month-end reconciliations. The finance team can measure the result on numbers it already reports: days to close, invoices per person, overdue receivables. The CFO owns the number, so nobody argues about whose saving it is.

Procurement is second because the money sits in contracts nobody rereads

Procurement is about leakage more than volume. An agent that reads every supplier contract, flags renewals and price changes, and drafts the renegotiation note does simple work. Its advantage is persistence.

The limit is ownership. An agent can flag a renewal, but someone must be allowed to renegotiate it. If that mandate does not exist, build the mandate before the agent.

Support is third, and only the repeat share

The repeat questions (order status, invoice copy, password reset) are worth automating and easy to count in tickets closed without a human. The rest is judgment, and agents handle it badly. Automate the repeat share, route the rest to a person, measure both.

Sales and marketing ops, IT and compliance come after, for the reasons in the table. Compliance in particular should wait until the earlier agents have a year of trust.

Who should not start here

A company under about €20M in revenue usually has one person doing each of these jobs, and the person is cheaper than the build. An agent placed into a broken four-day approval chain produces a faster four-day approval chain. Map the process first, then build into the map.

Questions people ask next

Should we start with the biggest cost centre?

No. Start with the function that passes the three tests. The biggest cost centre is usually operations or sales, where rules are weakest.

Is HR a good place to start?

Rarely. HR has volume in two places, onboarding paperwork and policy questions, and judgment everywhere else. Automate those two and leave the rest alone.

How long before the first agent runs in production?

In finance, weeks rather than months, if the approval rules are written down and the finance team owns the measurement. Longer usually means the process was not ready.

Do we need to fix the process before automating it?

Yes, when the process fails the rules test. An agent does its task correctly and the broken step downstream still waits. Fixing that step is the larger part of the value.

Lightbloom AI works in this order: hear the whole company first, write up every problem with a number, then build with your team. To see where the value sits in your company, book a demo.

References

1. Eurostat, Use of artificial intelligence in enterprises (2025), https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises
2. IBM Institute for Business Value, IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles (2025), https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles
3. Fortune, MIT report: 95% of generative AI pilots at companies are failing, reporting on MIT NANDA, The GenAI Divide: State of AI in Business 2025 (2025), https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
4. BCG, AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value (2024), https://www.bcg.com/press/24october2024-ai-adoption-in-2024-74-of-companies-struggle-to-achieve-and-scale-value