AI agents that autonomously carry out tasks toward a set goal—first-line handling of inquiries, searching and summarizing internal documents, running routine work automatically. Requests to build an “AI employee” have surged since the second half of 2025.
At the same time, outsourcing in this field has its own particular pitfalls. This article organizes what to look for in a development partner, what costs to expect, and how to prepare before you commission AI agent development.
The most common failure is “stalling at the PoC”
With AI agent development, failure far more often means “the demo worked, but it never got embedded into actual operations” than an outright inability to build it. The causes are almost always the same.
- Started without setting the accuracy pass/fail line — Only later do you realize it was a task where “even 90% accuracy is meaningless if a human still has to check all the remaining 10%.”
- Doesn’t connect to existing systems — The agent can’t securely access business data (internal databases, groupware), so it ends up being run by copy-and-paste.
- No accountability design — When the AI gets something wrong, it’s undefined who notices and who fixes it.
Five questions to gauge a development firm
| Question | How a good firm answers |
|---|---|
| ”How will you decide the PoC pass criteria?” | They agree on business metrics (processing time, accuracy, volume) before starting. |
| ”How will you connect to our existing systems?” | They mention standard connection methods such as MCP, and go as far as permissions and audit-log design. |
| ”What’s your design for when the AI makes a mistake?” | They propose where to insert human checks (human-in-the-loop), tailored to the work. |
| ”What happens when the model evolves?” | They can explain a build that isn’t locked to one specific model, and how easily it can be swapped out. |
| ”Who looks after it once it’s running?” | They present an operations setup that includes accuracy monitoring and updating prompts and data. |
We recommend choosing not by how polished the demo looks, but by whether they can answer these five concretely. Firms that use AI in their own daily development give concrete answers.
A sense of the costs
- PoC (2–4 weeks): roughly ¥500,000–¥1,500,000 — validate accuracy and business fit on one task with real data.
- Production build: from ¥2,000,000 — includes connecting to existing systems, permission design, and monitoring. Varies widely with the number of integration points and permission requirements.
- Operations and maintenance: roughly ¥50,000–¥300,000 per month — accuracy monitoring, improvement, and keeping up with model updates.
Rather than a single “straight to production” contract, a two-stage approach—validate the pass criteria in a PoC, then get a production estimate—tends to be cheaper and more reliable in the end.
What to sort out internally before commissioning
- Narrow the target down to one task — An “AI employee that can do anything” is a recipe for failure. Start with one task.
- Check whether you have “correct-answer data” for that task — Past handling records and examples of judgments become the AI’s training material.
- Decide the kind of mistakes you can tolerate — Put the asymmetry of each task into words, e.g. “misses are a problem, but false positives are acceptable.”
- Decide the scope of data it may touch — Settle the handling of personal and confidential information with your information-systems department before commissioning.
Summary｜“A firm that gets it into operation” beats “a firm that can build it”
Now that building an AI agent itself has gotten easier, the difference is made in implementation into operations and running it. SHANNON practices AI agents and AI-driven development daily in operating its own SaaS, and offers support in the form of a PoC (2–4 weeks) where we design the pass criteria together with you from the start. Please start with the “prove it small, one task first” approach.