Custom AI agents for business
Last updated: September 2026
Sapio develops AI agents that work over your company's systems and documents: they search your ERP and CRM, read contracts and procedures, draft replies and carry out approved steps. We build them to go into production: least-privilege permissions, an evaluation set drawn from your real cases, an audit trail for every action and a classification under the EU AI Act.
What AI agents do we build for businesses?
We build agents with controlled access to your data and systems that do one precise job: answer with a source, prepare a decision, or carry out a step a person would otherwise do by hand. An agent is not a chatbot with a personality. It is software that uses an AI model to decide which tool to call, and in what order.
| Type of agent | What it does | Example |
|---|---|---|
| Assistant over company documents (RAG) | Answers from procedures, contracts, legislation or technical documentation and cites the exact source | A lawyer asks what the law says; a technician looks up the procedure for a fault code |
| Agent over systems | Reads from ERP, CRM or helpdesk, combines the data and prepares the answer or action | "Where is order X and when does it arrive?" without opening three applications |
| Customer-support agent | Answers frequent questions, checks status in your systems, escalates with the full context | Chat, email or WhatsApp, with hand-off to a person |
| Monitoring agent | Watches an external source and flags what matters to you | Public tenders filtered against your company profile |
| Document-preparation agent | Assembles a quote, report or reply from templates and data, for approval | The first draft of an offer, checked by a person |
If your process is a fixed sequence of steps with no decisions, you probably need automation, not an agent. The difference is explained in how to choose between an AI agent and RPA, and we build automation on the AI process automation page. For the basic definition, Vlad Tudor has written separately about what an AI agent actually is.
Which technologies and systems do our agents work with?
We build on or integrate with the models, systems and channels you already use. If a system has an API, a database or an export, an agent can usually reach it, with the permissions you set.
- AI models
- OpenAIAnthropic ClaudeGoogle GeminiMistralOpen-source models on-premise
- Business systems
- SAPOracle NetSuiteMicrosoft Dynamics 365OdooSalesforceHubSpotPipedrive
- Channels and support
- Microsoft TeamsSlackWhatsApp BusinessEmailWeb chatZendeskFreshdesk
- Documents and data
- SharePointGoogle WorkspacePostgreSQLn8n
What makes an AI agent production-ready?
An agent is production-ready when you know exactly what it is allowed to do, how often it is right, and what it did yesterday at 2 p.m. The demo shows it can; production requires proof that it is under control. Every agent we deliver has:
- Least-privilege permissions. The agent can reach only the systems and actions it needs, with the rights of the user it acts for. Reading is separate from writing; writes go through approval until the agent has earned them.
- An evaluation set. Real cases from your company, with the correct answer marked by the people who do the work today. We re-run it on every model, prompt or tool change, to see whether the agent got better or only different.
- Cited sources. A document assistant shows where the answer came from. If it cannot find a source, it says so rather than improvising.
- An audit trail. What the user asked, what the agent searched, which tools it called, what it answered and who approved.
- Role limits. The agent knows what it does not do: no discounts promised, no accounts changed, no legal or medical verdicts. It leaves its role by handing the case to a person.
- EU AI Act awareness. We classify each use case. For agents that talk to people, the Article 50 transparency obligations have applied since 2 August 2026: people must be told they are interacting with an AI system (Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744). General information, not legal advice.
- Data stays where you decide. Public cloud, your own cloud region or on-premise, depending on how sensitive the data is.
Why we insist on this: Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 (Gartner, 25 June 2025). That is a forecast, not a finding. The best-known public example points the same way: in 2024 Klarna said its AI agent was doing the work of 700 staff; in 2025 its CEO said they had gone too far, quality had dropped, and it began hiring people again (Forbes, 18 May 2025). The agent hit its volume target and missed on quality.
How does AI agent development with Sapio run?
We start with a paid, fixed-price sprint on your data. Only then do we build.
- Scoping or audit sprint (a few days to two weeks). We pick one use case, define what the agent may do, build the first evaluation cases and test on a sample of your documents or systems. You leave with a written specification and a fixed price for the build.
- Fixed-price build (usually 6–12 weeks). Integrations, permissions, evaluation set, audit trail, interface (chat, email, Teams or inside your application). The agent starts in "propose, a person approves" mode and earns autonomy only on the case types where it has proved it is right.
- Optional: an AI engineer inside your team. If this agent is the first of several, a senior Sapio engineer can work inside your team, paired with one of your engineers, until your team can run it alone. Ask us about it.
For a support agent, the concrete steps are in how to deploy an AI agent for customer support; for a document assistant, in how to build a chatbot that knows your company data (RAG).
Which AI agents have we already built?
- ai-aflat.ro, built by Sapio: an AI assistant (RAG) over 220,000+ Romanian legislative acts, updated daily, used by more than 15,000 people. Every answer cites the law and links the official source, and the assistant deliberately refuses to give verdicts. Those are the rules we apply to every agent: a cited source and a clear role limit. How we built it.
- An AI customer-support agent for retail, integrated with ChatGPT, delivered for a client that went on to raise a $4.7M seed round.
- A public-tenders agent that reads notices on e-licitatie.ro, Romania's procurement portal, and matches them against a company profile built from its past projects. In testing it identified the relevant tenders. It is not yet in production.
- Sapio was founded in 2021 by Vlad Tudor (Forbes Romania 30 Under 30, 2025), who leads every engagement personally.
What drives the cost of a custom AI agent?
We do not publish prices: an assistant over 50 procedures and an agent that writes into your ERP are different projects. The cost depends on:
- How many systems the agent touches, and whether they have APIs or need another route in.
- Whether it only reads or also acts. Writing into systems needs permissions, approvals and stricter tests.
- The volume and quality of the documents. Clean and structured, or scanned, versioned and contradictory.
- How bad a wrong answer is. An internal answer about a procedure, or an answer to a customer about money.
- Where it runs. Cloud, your region or on-premise; plus the monthly running cost, which grows with the number of conversations.
- The AI Act classification and the documentation it requires.
Frequently asked questions
What is an AI agent for business, in short?
A program that uses an AI model to understand a request, search your company's data and carry out steps through a set of permitted tools, with a person approving what is sensitive. It differs from a chatbot, which only answers, and from classic automation, which follows fixed steps.
Do I need an AI agent or a chatbot?
If you only need answers from documents, a RAG assistant with cited sources is enough. If the answer needs data from your systems or an action, such as an order status or a new ticket, you need an agent.
How long does it take to build a custom AI agent?
The scoping sprint takes a few days to two weeks. Building an integrated agent usually takes 6–12 weeks, including the period in which the agent proposes and a person approves.
Should we build a custom agent or buy a ready-made platform?
If a platform covers your process and integrates with your systems, use it. A custom agent pays off when your systems have no connector, when sensitive data cannot leave your infrastructure, or when the process is where you compete.
Can an AI agent make decisions on its own?
Yes, on the case types where it has proved on the evaluation set that it is right and where a mistake can be reversed. Payments, commercial promises and decisions that affect people stay with a person.
Who owns the agent after the project?
You do: the code, prompts, evaluation set and documentation stay in your repository, and the agent runs on the infrastructure you choose.
A good agent does not impress by how much it can do. It impresses by how clearly you know what it will not do.
Tell us what your agent should do