An AI proof of concept usually takes four to eight weeks. That holds when you test one piece of work, you have the data at the start, and everyone agrees in advance how you will judge the result. Most delays come from waiting for access to the data, for security approvals, and for time from the people who can tell whether the results are right. For Boston Consulting Group (BCG), StrataEdge and two engineers built a working AI agent in four weeks, with structured outputs and human review built in.
What happens during a proof of concept
We build a working version of one part of the job and test it on real examples. The steps usually go in this order:
- Agree on what the software should produce, how the work performs today, and what result would justify going further, such as a lower error rate or a shorter processing time.
- Collect real examples where you already know the right answer, including incomplete and unusual ones.
- Build the simplest version that does the job. Use rules or ordinary code wherever they are enough.
- Test it against the examples, and have the people who do the work today review the differences.
- Estimate what production would take and cost, and decide whether to continue, change course, or stop.
Do the first step before any building starts. If the team has not agreed on what a good result looks like, nobody can tell at the end whether the proof of concept worked.
What makes it take longer
Data access is often the longest wait. Approval to read production records, especially records with personal information, can take weeks. Start those approvals before the build begins, or agree to work with masked or sample records while they proceed.
Scope adds time. Every extra document type, system, or decision the software has to handle means more building and testing. Pick one job that matters to the business and leave the rest for later.
Review time matters too. Someone who knows the work has to check whether the results are correct. If that person can give only an hour a week, the proof of concept will take longer. Agree on their time before you start.
What happens after it works
A working proof of concept shows that the approach works on your examples. Production takes more: integration with your systems, permissions, monitoring, exception handling, and training for the people who will run it. Treat that as a separate decision with its own estimate of build and operating cost.
For BCG, the agent was in production three months after we started, and we continue to improve it with them. Read the BCG case.
If you don't know what to test first
If you have several ideas and no clear first choice, start with a review. Our AI investment review usually takes two to four weeks. We go through the work, the data, and the full cost with your leaders, and you get a ranked list of what to build, what to buy, what to fix without AI, and what to stop. See our typical engagements.