The dilemma that stalls many COOs and CFOs
You have an ERP running. Your team knows it. Processes are more or less defined. And yet, certain tasks still depend on people: consolidating data from multiple sources, preparing management reports, cross-referencing information between modules, answering internal queries about inventory or invoicing.
The question is not whether AI can help. The question is when to act and how to do so without taking on unnecessary risk.
Mid-size companies typically land on one of three answers: build in-house, hire someone to do it, or wait for the technology to mature. All three have a logic to them. And all three carry costs that are not always calculated.
Option 1: build in-house
The in-house build option makes sense when the technology team has genuine available capacity, when the problem is specific enough to justify proprietary development, and when the company has the tolerance to iterate over several months.
The common problem: the IT team is running at 100% maintaining the ERP, managing incidents, and executing already-committed projects. There is no room to explore new technologies with the depth that a production-grade AI agent requires.
Building an agent that integrates with an ERP, handles errors, has observability, and is genuinely adopted by the business team is not a two-week project. It requires architecture decisions, prompt management, API integration, testing with real data, and an adoption process. If the team lacks the time and the methodology, the project drags on or gets abandoned.
When in-house build makes sense: when there is a dedicated technical profile with real allocated time, and the problem is scoped narrowly enough to deliver something functional in under six weeks.
Option 2: hire a partner
Hiring an external partner accelerates implementation, but not all proposals on the market are equivalent. There is an important difference between a partner who builds the agent and hands it over, and a partner who builds alongside your team and leaves the capability embedded.
The first model creates dependency: every modification, every new agent, every adjustment requires going back to the vendor. The second model leaves your team autonomous.
For a mid-size company, dependency is a real operational risk. If the vendor shifts priorities, raises prices, or simply fails to respond with the necessary speed, the system is left without internal support.
What to evaluate when hiring: does the partner work alongside your team or instead of your team? Is there an explicit knowledge transfer? Can your team operate and modify the agents without external intervention once the project closes?
When hiring makes sense: when the team does not have time to build in-house, when the problem has a clear economic impact, and when the partner guarantees autonomy at project close.
Option 3: wait
Waiting has an apparent logic: the technology improves every month, costs come down, and more documented use cases emerge. Why act now?
The problem is that waiting also has a cost. Every month your team spends 20 hours consolidating data manually is a real cost. Every report that reaches management late is a decision made on incomplete information. Every manual process that depends on a specific individual is a continuity risk.
Furthermore, competitors already implementing are not waiting. This is not an artificially manufactured urgency argument — it is a competitiveness variable that compounds month after month.
When waiting makes sense: when no process has a clear economic impact, when the team is in the middle of an ERP migration, or when an internal reorganization makes any new implementation unstable.
How to assess whether the time is now
Three concrete questions to reach a decision:
1. Is there a process that consumes more than 10 hours per month and whose logic is repeatable?
If the answer is yes, there is a viable use case. Common examples in companies with an ERP: consolidation of financial close data, generation of sales reports by region, reconciliation of invoices against purchase orders, responses to internal inventory queries.
2. What is the real cost of that process today?
A reasonable estimate: if the process consumes 20 hours per month from a profile with a fully loaded cost of 3,000 euros per month, the monthly cost of that process is approximately 375 euros. Over twelve months, 4,500 euros. If an agent can reduce that time by 70%, the annual saving on that single process falls between 3,000 and 3,500 euros. Multiplied across three or four similar processes, the ROI hypothesis becomes concrete.
3. Does the team have the capacity to adopt something new in the next six weeks?
This is not about technical capacity — it is about real availability. An agent nobody uses generates no ROI. Adoption requires that someone on the team is willing to change the way they work.
A concrete example
An industrial distribution company running SAP was spending between 25 and 30 hours per month preparing the month-end close report for management: extracting data from multiple modules, cross-referencing it in Excel, validating with the accounting team, and formatting the final presentation.
With an agent connected to the ERP via API, that process became automated. The accounting team stopped doing the extraction and cross-referencing work and focused on final validation. Time spent dropped to between 5 and 8 hours per month. The report reached management two days earlier than under the previous process.
The project cost was recovered in under four months.
Conclusion
Build, hire, or wait is not an ideological decision. It is a financial and operational decision that depends on concrete variables: team capacity, problem impact, and the opportunity cost of inaction.
If you have a repetitive process with measurable impact and a team with minimal availability to adopt something new, the time to evaluate is now — not next quarter.
If you want to assess whether there is a viable use case in your operation, we can run that analysis in 15 minutes.
[→ Request a free diagnostic]
