One procurement automation proposal from the summer stayed with me. The price was over EUR 30,000, the build was expected to take three months, and ongoing support would cost EUR 100 an hour. For the same work requirement, we reached a usable solution in about five hours.
Five hours. That difference made me ask what the much larger proposal was actually selling.
Before commissioning AI automation, ask what will be delivered, why that approach has been chosen, and how it will be tested with your data. Look beyond the development price to running costs and whether you will be able to operate the system yourself. The title “AI specialist” or “AI trainer” does not answer those questions on its own.
Over EUR 30,000, three months and paid support
The proposal included a demonstration. In my view, the approach resembled how I might have solved the same problem about a year earlier. My own tools and capabilities had moved on in the meantime. That made the three-month development plan difficult to understand.
I was also concerned about what would happen after handover. If only the person who built a system understands it, every change means going back to them. Support in this proposal cost EUR 100 an hour. For me, that meant considering the potential ongoing dependency on the supplier as well.
The procurement automation solution we needed was ready in approximately five hours and could be used immediately. We did not need to commission the proposed three-month development project for that work.
Five hours describes the time taken to reach a usable solution. A complete price comparison would also need to consider scope, integrations, user numbers, reliability requirements and agreed responsibilities. None of those can be inferred from a single time comparison. It was, however, enough to make me ask: could we try a much smaller solution first?
An AI solution’s price needs a clear explanation
As a business owner, I want to know what work I am paying for. If a proposal includes three months of development, the supplier should be able to explain what will happen during that time and which part of my problem each stage will solve.
Substantial work might involve connecting legacy systems, cleaning up data, managing multiple users’ permissions, recovering from errors or handling heavy usage. If so, that work should be visible in the proposal. The word “AI” does not explain it by itself.
The same applies to the age of the technology. A year-old tool may still be an excellent choice. My question is whether the chosen approach makes sense for today’s task and whether the supplier has considered simpler options.
Does this task need an AI agent at all?
AI automation usually means that AI performs part of a workflow, such as reading a document, classifying text or preparing a summary. An AI agent can also choose subsequent steps and use the tools it has been given. Not every recurring task needs that degree of freedom.
If the work follows fixed rules, an existing software feature, a simple script or a direct connection between two systems may be enough. If different kinds of text need interpretation, AI may be useful at that particular step. The rest of the workflow can remain rule-based.
This principle also appears in Anthropic’s guide to building agents: start simple and add complexity when needed. OpenAI’s practical guide to building AI agents likewise recommends checking whether a task needs an agent and beginning with a limited scope.
When connecting systems, find out what your existing software already supports. An API, or a suitable MCP connection, may simplify part of the work. Neither automatically solves access permissions, data quality or every exception. The supplier should be able to explain what will work in your situation.
Five questions before commissioning AI automation
1. What must work when the job is finished?
Ask the supplier to describe the result through one specific work task. Where does it begin? Which file, record or action marks its end? Who checks the result, and what criteria determine whether the work can be accepted?
“We will automate procurement” leaves too much open. Ask which stages are included, which remain a person’s responsibility and which are excluded altogether. That also gives another supplier enough information to prepare a comparable proposal.
2. Could an existing tool do the job?
Ask to see at least one simpler option. Does your current software already handle part of the task? Could you start with a narrower stage of the workflow? Why was a separate system chosen?
There may be a sound reason. But “this is how we always do it” or a long list of complex tool names is not an answer. You should understand the benefit of each necessary component.
3. Does the solution work with my data?
Ask for a test using examples that resemble your company’s work. Anonymised or fictional examples can be used initially instead of confidential data. Alongside a straightforward case, include situations with missing information, different file formats or conflicting data.
Inspect the outcome: did the right data reach the right place, was anything omitted, and did the system notice problems? Anthropic’s evaluation guide also distinguishes between what an agent says and what actually happened in the system. One successful demonstration is not enough to assess an entire workflow.
4. What costs come on top of development?
Ask for software subscriptions, AI usage, hosting, maintenance and changes to be listed separately. What happens if a connected service changes? Which support is included, and which work is charged by the hour?
If support costs EUR 100 an hour, I want to know in advance when it will be needed and how much support ordinary use is likely to require. Ask for an estimate with its assumptions, not just one particularly inexpensive example month.
5. Can I operate the system myself?
Whose accounts does the solution run on? Who can change its settings? What instructions and access will be handed over? Could another developer continue the work? How can you retrieve your data if the relationship ends?
These questions should be addressed in the proposal. Handover could also include a practical walkthrough: someone from your company starts the task, checks the result and knows what to do when a typical error occurs.
A small test can inform a bigger decision
If a proposal’s price or scope raises questions, ask for a limited test first. Agree on one stage of the work, suitable examples, the expected result and an evaluation method. The test may be paid; its purpose is to give you evidence for the next investment.
Measure the time spent on the work, including checking and corrections. Record the cases where the solution fails as well. This shows the benefit of what already works and where further development is needed. I discuss this in “Measuring the benefits of AI”.
Our five-hour experience made me think about exactly this. Before commissioning three months of development, it is often possible to investigate how far a smaller step will take you.
Get a second opinion, including from a university
If nobody in your company can assess a technical proposal, get another opinion. Give the reviewer the same task description and requirements. Ask which work is necessary, what could be simplified and what the proposal leaves out.
One option is to explore research and development collaboration with the University of Tartu’s Institute of Computer Science. The institute works with businesses. You can send its contact person a short description of the problem, available data and desired result.
Manufacturing companies can also look into AIRE demonstration projects, which offer opportunities to test the suitability of AI or robotics before making a larger investment. Eligibility and conditions need to be checked; this does not mean every company can obtain a complete system free of charge.
My advice to a business owner is straightforward: ask, check and compare before ordering. A proposal should make sense to someone who does not build AI systems. If you cannot understand what you are paying for, the supplier needs to explain it.
To help your team understand which tasks to give AI and how to check the result, explore AI team training. We begin with your work tasks and practise evaluating the solution together.