Bring the invoice that took three emails to approve. It will tell an AI agency more about your company than a slide headed “Our AI vision”.
Perhaps the amount was difficult to read. Perhaps it was clear from the start, and all three emails were spent finding the person who had ordered the work. The same request to “automate invoices” can hide two very different projects.
For a first agency meeting in Leipzig, follow that one invoice. We write from Pharen’s perspective in Naunhof, in the Leipzig district. The example should make our proposal easier to judge too.
Show the detour as well as the destination
“We want to automate invoices” leaves the problem open. Perhaps capture takes too long. Perhaps capture is already quick and nobody knows who should approve the result. Those require different work.
Walk through a suitable case from arrival to outcome, using anonymised material where appropriate. Pay attention to the spaces between tools: a copied amount, an extra message, a call to a colleague. They reveal work that otherwise remains invisible.
An agency should be able to explain where it would intervene and what it still needs to know. If the answer immediately consists of a product name, the investigation has not happened yet.
Give an on-site visit a purpose
A visit can help when the process needs to be observed. The list on screen may look tidy while a sheet of paper beside it contains the cases that require different treatment. Those exceptions belong in the process too.
Other steps may work well through screen sharing. Development, reviewing a proposal and many tests can be organised remotely. The useful question is what people need to see together to understand the work.
Use that to agree which meetings happen on site and how travel appears in the quote. Our Leipzig provider comparison explains why a location page alone does not establish a local office.
A small pilot can still complete a whole job
For incoming invoices, the first goal might be modest: a findable document, visible missing information, a reviewer and a recorded approval. That completes a useful journey without rebuilding the finance function.
A boundary makes the proposal assessable. You can see which inputs are included, which systems are connected and where the result goes. You can also see what is deliberately outside the first phase.
Include the difficult case. What happens to an incomplete invoice? If this is left until after commissioning, the pilot is not yet adequately described. The article on AI in invoice intake shows why that case reveals so much.
Your team provides more than test data
An agency needs access to people who know the work. They can explain why an apparently redundant check exists or why an exception cannot be decided automatically.
Allow time for this. A process owner needs to answer questions, review drafts and judge whether the proposed route will work in practice. Involve IT and the people responsible for data processing early when their decisions affect the setup.
If nobody can make those decisions internally, reduce the scope or wait. More development hours cannot substitute for missing knowledge of how the business should operate.
Leave with something you can inspect
Request a written description of the proposed journey: its start, expected outcome, approval and handling of an unresolved question. Include implementation cost, ongoing cost and what your team must contribute.
Agree how progress will be assessed. Quicker capture is different from quicker approval. Without a baseline, you cannot later tell whether work was removed or transferred to someone else.
What we could build in Pharen Hub for that invoice
For an invoice project, we can use Pharen Hub Lists for the invoice record and Workflows for its checks and approval. The practical work includes mapping your intake and accounting handoff, not merely enabling AI. Bring the three emails into that mapping: each should become a clear question, decision or source in the proposed process.
