AI agent implementation, built like production software

An AI agent is software that uses a language model to complete a defined task end to end: it plans steps, reaches systems and documents through permitted tools, and asks for human approval before important actions. Unlike a chatbot, it doesn't just answer; it does the work and follows up on it.

Automation in n8n or Zapier runs fixed rules, and it fits when the input is structured and every case is known in advance. An agent also works with unstructured input, such as an email, a drawing or a fault report, and decides the steps itself. That makes it a fit for processes that are hard to capture in fixed rules and have an outcome a person can check.

A Discovery is the first step of every implementation: we map one process using real examples, measure the baseline, design the agent and estimate the return on investment. We deliberately take on your most complex process, to make it significantly better in handling time, effectiveness and running cost. It has a fixed fee agreed up front in a written proposal, and if you go ahead with setup, the fee is credited toward your first months of monthly care. You can cancel monthly care any time with one month's notice, with no minimum period.

AI consulting that starts with one process

It's priced per workflow: a one-time setup plus a fixed monthly care fee. The amount depends mainly on how many systems need connecting, the state of your data and the approval flow, so the fixed price for each workflow is set in writing after the Discovery, before any build work.

You pay the model usage directly to the provider, on your own account: you see every charge, there's no mark-up from us, and the data agreement is between you and the provider. Discovery estimates the expected usage; with a local model on your own servers, the cost is hardware instead of usage.

What makes up the cost of an AI agent

Before go-live, the agent goes through shadow mode to catch mistakes; after go-live, every important step waits for approval from someone on your team. In shadow mode it works on real cases alongside your team, but nothing it produces goes out: not to a customer and not to a system. Each draft is compared with your team's work, and the agent goes live only when the criteria agreed in writing are met.

Every step the agent takes is logged, so you can check what happened and why, and roll back to the previous version if needed. With your ERP, the agent gets a dedicated user with minimal permissions, read-only at first, and is built in a test environment before it touches live data.

Not necessarily. For sites where the data is sensitive, we can run on-premise models on your own servers, so the data never goes to an outside provider. Where cloud APIs fit, we use them. We make the choice between cloud and on-premise together during Discovery, based on how sensitive the data is, the performance needed and the cost.

Your data, your approved cases and every report we deliver are yours. The software we build and run for you remains ours, and you use it under the licence in your agreement. You can get full access to it and its environment at any time: on request we grant access right away, after you sign a short acknowledgement of what changes there can affect.

No single model fits every task, so each solution is tailored to the client and the choice is made during Discovery, then tested on the test set and in shadow mode. The framework, such as Dify, LangGraph or Hermes Agent, depends on how complex the process is, which systems need connecting and who will maintain the agent. For each task we pick the model with the best price-performance, for example Claude, GPT, Gemini, DeepSeek, Kimi, Qwen or GLM; a simple task doesn't need the largest, most expensive model. We design the agent so the model can be swapped later, if another one gives a better result at the same cost, or the same result for less.

An AI agent can connect to systems such as Priority or SAP Business One through their API. We start with read-only access, and every write action goes through human approval. Feasibility for your version and modules is checked during Discovery, before any commitment to setup.

Connecting an AI agent to Priority and your ERP

AI implementations are led by an engineer whose background is software that runs in production without close supervision: CI/CD and infrastructure automation, automated HIL tests that run on every change, performance analysis for safety-critical software and production-management software. Tests, monitoring and a record of every step are what make an agent safe to run, the same discipline as in our automation projects, where a system is tested in our lab before it reaches site. That's why every implementation starts with a Discovery and goes through shadow mode before go-live.

Industrial automation: the data that feeds the agent