AI agents for business — autonomous process automation
Not a chatbot on your website, but a program that runs a process end to end: it reads, decides, writes to your systems, and hands back to a human whatever needs a human.
Almost everyone has AI. Almost no one has agents
This is the most interesting number in the AI Index 2026 report: organizational AI adoption reached 88%, yet agent deployment stayed in the single digits across nearly every business function. Agents work — few companies know how to deploy them.
- 88%
- 69%
- 26%
- 62%
An AI agent is a process, not a chatbot. We take one task that happens hundreds of times a month in your company and hand it to a program together with guardrails, a log and a kill switch. Only once that works and can be measured do we add the next one.
How an agent differs from a chatbot
It has access to tools
A chatbot replies with text. An agent calls APIs, writes to your CRM, issues documents, sends messages and verifies that the operation actually succeeded. Without that, you get a polite conversation with no outcome.
It works in steps
It receives a goal, breaks it into stages and returns to the one that failed. The labour market moved in exactly this direction: in 2025, job postings stopped asking about chat tools and started asking about orchestrating task-oriented systems.
Its limits are in the code
What is allowed, what is not, which amount requires human approval, what happens on failure. Not a sentence in a prompt that can be talked around, but rules in the system and the smallest possible set of permissions.
It leaves a trail
Every decision and every tool call goes to a log. Without it you can neither prove the return nor find the cause when something goes wrong.
Where agents are already good, and where they fail
We would rather say this in the first meeting than halfway through a deployment. The numbers below come from the Stanford report, not from our deck.
Here they work
Structured work with a verifiable result: ticket handling, moving data between systems, preparing documents, code. Studies cited in the AI Index 2026 report show productivity gains of 14–15% in customer support, 26% in software development and 50% in marketing output.
Here they underperform
Tasks requiring judgment and context that exists in no system. Gains there are smaller or negative. The report adds a longer-term warning: heavy reliance on AI can slow down how fast a team builds its own skills.
Here they get it wrong
On τ-bench, which tests conversation with a user plus tool calls under policy constraints, the best model reaches 70.2% and none exceeds 71%. On OSWorld, agents fail roughly one attempt in three. That is why every project starts with the question of what happens on that one in three.
Where we start in your company
The percentage next to each function is the share of organizations using no agents there at all. The higher it is, the less competition — and the cheaper it is to build an advantage before someone else does.
- 71%
Customer service and tickets
Qualifying tickets, answering repeat questions, opening and closing tasks, escalating to a human by rule rather than by mood.
- 69%
IT and internal support
First-line requests, granting permissions, keeping assets and documentation in order. The most active function in the report — and still two thirds of companies without agents.
- 68%
Marketing and sales
Qualifying inbound leads, keeping the CRM complete, preparing quotes and content for a human to approve.
- 66%
Knowledge management
Answers based on your own documents — with a citation to the source and an honest "I don't know" when the source isn't there.
- 77%
Software and product development
Code review, tests, migrations, documentation. This is where productivity gains are best measured and show up fastest.
AI answers from memory until you show it your documents
Ask any artificial intelligence — ChatGPT, Copilot, any other model, meaning a program that assembles answers out of sentences it has read on the internet — about the warranty terms in your company. You will get an answer that is fluent, confident and untrue, because it has never seen your terms and is assembling the answer out of thousands of other people's. A new employee in that situation has two options: answer from memory, or get up and open the binder. We build agents that first get up for the binder, and then show the page they are quoting from. The mechanism is called RAG — retrieval-augmented generation, meaning answering on the basis of retrieved documents. If you hear the acronym from someone else, it means exactly what we describe below.
It looks it up first, then answers
Before any answer appears, the system searches your documents and picks the passages that relate to the question. Only those go to the model, together with the question. That is the difference between "I think the warranty is two years" and "the terms of sale, point 7, say two years".
The model does not learn your documents
This is the most common misunderstanding in these conversations. We do not train the model on your files and we upload nothing into it. Your documents stay in your systems, and we build a search index over them that pulls out the matching passages with every question. Only those passages — not the whole price list and not the whole contract — go to the model so it can compose an answer from them. Where that index and that model physically sit is agreed with you before we start. You correct a price on Friday morning — the agent quotes the new one as soon as the system has reloaded the corrected file, usually within minutes. Updating the knowledge means swapping a file, not running another project.
Every answer comes with a citation
The answer arrives with the name of the document and the passage it came from — you can click through and read the original. That is the whole difference between an answer you have to believe and an answer you can verify on the spot by clicking through to the source. An answer without a source has to be confirmed by hand anyway, so it saves nobody any time.
It limits invention, it does not remove it
The system can pull the wrong passage, or land on a document that expired long ago — and it will back the wrong answer with a citation. That is why we set the agent up to say "I don't know, it is not in the documents" and hand the case to a person: no source is a correct outcome here, not a failure. And with the source shown, the mistake is visible straight away, rather than only once the customer sees it.
The knowledge sits in two people's heads, not in your documents
Every company has someone who gets asked about everything. When they go on holiday, work slows down. An agent has nobody to ask. It answers only from what you gave it — so a more expensive artificial intelligence will not fix bad documents. Rubbish in, rubbish out, only phrased more elegantly. That is why a deployment starts with your documents, not with the choice of model.
Three versions of the price list mean three different answers
If a March price list, a June price list and "the right one" on a sales rep's desktop are all in circulation, the system will quote one of them, and not necessarily the one in force. A person would ask the colleague next door. A program will not ask; it will answer in a confident voice. Order in your files does more here than the choice of which company's artificial intelligence we use.
What sits in people's heads has to be written down
The most valuable knowledge in a company is usually recorded nowhere: why this customer is billed differently, what must never be promised over the phone, what to do with a complaint filed after the deadline. We collect it in short conversations with your team and turn it into documents a few pages long. As a side effect, the company stops stalling when those two people take their holiday. This is the first stage of the deployment, not a precondition: if the audit shows that the tidying up alone is a bigger job than the agent, we will say so plainly and start there.
One current document per topic
Price list, warranty terms, complaints procedure, discount rules — one binding document for each, with a review date and a named person who keeps it up to date. Without that, a year from now the agent answers the way things were a year ago — only very convincingly. The log helps here, meaning the record of every answer: you can see which file an answer came from, so you can also see which file is spoiling the results.
Permissions travel with the knowledge
Not everyone in the company is meant to see margins, contracts and HR records. The system first checks who is asking, and searches only the documents that person could open anyway. A sales rep will not see margins through the agent if they cannot see them in the system. A deployment without this step is not an assistant, it is a leak with a convenient search box.
The procedure card that runs itself
Every company has that card pinned up somewhere: step 1, step 2, step 3, the same way every time. An invoice arrives, check the seller's details, enter it into the system, send it for approval, file it. Nobody thinks about the order — and the result still depends on who had time and whether anything was forgotten. A card like that can be rewritten as a program that walks through it on its own. This is called a workflow — the service is simply called process automation in our offer, and it is usually a cheaper answer than an agent.
A workflow follows the path, an agent picks it
A workflow does exactly what was written down, always in the same order. The same invoice takes the same route on Monday and on Friday, no matter who happens to be at work. An agent is given a goal, not a list of steps — it decides for itself where to start and which stage to return to when something fails. You pay for that freedom with a longer deployment, more supervision and less certainty that two identical cases will take the same route.
A test you can run yourself before you call us
Try writing the process out as a list of points. If you can, and it does not end with "it depends", this is work for plain automation rather than for an agent — cheaper, faster and more reliable. If the case first has to be read and understood before you know which route it takes, then the conversation about an agent begins. We say this even when you come to us asking for an agent.
The sensible deployment is usually both
You do not have to choose between "no AI" and "an agent deciding everything". Usually the workflow carries the case from start to finish. AI does the one thing that cannot be written down as rules: it reads a message written in plain human language and pulls the data out of it. Fixed steps do the rest. That one step has to be measured like an agent: the same message may be understood better one time and worse the next. That is why the pilot tests that step rather than the whole workflow. Cost and risk then grow in one place instead of across the whole process.
A failure is visible immediately
When a workflow stops, you know at which step and why: you fix that one step and carry on. With an agent the diagnosis takes longer, because you have to reconstruct why it chose that order rather than another. That is why more than one process we are asked to hand to an agent ends up as a workflow — cheaper to build and easier to repair.
How we deploy
- 01
Process audit
Two to three days with your team. We look for tasks that are repetitive, measurable, and where a mistake is not a catastrophe. We leave with a shortlist and an estimate of the time you can win back.
- 02
A business case for one process
What this task costs today, what it should cost after deployment, and how we will know it worked. "Unclear business value" blocks agent deployment in 32% of organizations — here it is on paper before the first line of code.
- 03
A pilot on one process
Four to six weeks. The agent works next to a person, not instead of them: it proposes, the human approves. We measure accuracy on your real cases, not on a benchmark.
- 04
Guardrails, oversight and security
Smallest possible permissions, data in your infrastructure or with a provider you choose, a log of every decision, a threshold that requires human approval and a kill switch within reach. Security is the most frequently named barrier to deployment, so it is a separate stage here, not a clause in a contract.
- 05
Deployment and integrations
Connecting to the systems you already run: CRM, ERP, mail, inventory, spreadsheets. We train your team and leave documentation good enough to maintain it without us.
- 06
Maintenance and growth
Models and APIs change every few months. We monitor accuracy, refine the rules, and add the next process only after the pilot is stable. Covered by Wataha Care.
The barriers most deployments stall on
Order and percentages from the McKinsey survey cited in the AI Index 2026 report. For each barrier we say what we do about it — because these are exactly the conversations deployments run into.
Security and risk
Smallest possible permissions, data under your control, a log of every operation, tests on synthetic data before production, and a kill switch that stops the agent instantly.
Technical limitations
We do not promise an agent can do everything. We match processes to what models actually do well — hence the section above, with numbers instead of assurances.
Regulatory uncertainty
We work with GDPR and the EU AI Act in view: risk classification, clear disclosure that the user is talking to a system, an operations register and deployment documentation. The AI Act's first prohibitions took effect in 2025.
Gaps in oversight tooling
A panel with decision history, accuracy metrics and token cost ships with the agent, not as a paid add-on a quarter later.
Budget
We start with one process and a calculated return, not a company-wide platform. A single pilot is a reversible decision.
Unclear business value
The business case comes before the code. If the audit shows no return, we say so plainly and turn the project down.
What it costs
Honestly: an agent cannot be priced off a list, because the cost depends on how many integrations are needed and how well organized your data is. So we work in one of three modes.
Running costs (tokens, model provider APIs) are separate and visible — you see them in the panel from day one of the pilot, not on an invoice a quarter later.
Before agents come in
An agent runs on data and processes that are already in order. When they are not, we start one step lower — and we say so in the first conversation.
Questions we hear most often
How is an AI agent different from a chatbot on a website?
A chatbot talks; an agent acts. An agent has access to tools (APIs, CRM, mail, databases), splits a task into steps, retries the ones that failed, and logs every decision. It also has limits built into the system: a permission scope, a threshold requiring human approval, and a kill switch. A website chatbot will answer a question about opening hours; an agent will take the request, create the task, assign it to the right person and check that it was closed.
How much does deploying an AI agent cost?
A custom deployment starts from PLN 10,000 — the same as a custom web system, because it is the same class of work: integrations, permissions, an oversight panel and documentation. The figure depends mostly on how many systems the agent must connect to and on security requirements. A pilot on one process is quoted after the audit, because only the audit reveals how much work sits on the data side. Care and development afterwards run under Wataha Care, from PLN 100 per month. Token and provider API costs are separate and shown in the panel.
How long does an AI agent deployment take?
The process audit takes two to three days. A pilot on one process usually closes within four to six weeks, and it is the pilot that decides whether to go further. Full deployment depends on the number of integrations: the more systems and the messier the data, the longer it takes. We do not quote a single timeline for everything, because the difference between one system and five is a difference of months.
Is my data safe? Will it be used to train a model?
Your data stays where you decide: in your own infrastructure or with a provider you choose, on terms we review before signing. The agent gets the smallest possible permissions — access only to what a given process needs. Every operation goes to a log. Security and risk is the most frequently named barrier to agent deployment (62% of organizations in the survey cited in the AI Index 2026 report), so we treat it as a separate project stage, not a contract clause.
What happens when the agent gets it wrong?
It will get things wrong, and that has to be designed for rather than promised away. On τ-bench, which tests conversation plus tool calls under policy constraints, the best model reaches 70.2% and none exceeds 71%. On OSWorld, agents fail roughly one attempt in three. So every deployment keeps a human in the loop on irreversible operations, a threshold requiring approval, a log that lets you reconstruct any decision, and the ability to roll changes back. The pilot exists to measure that failure rate on your own cases.
Will an AI agent make my employees redundant?
We do not sell deployments as a way to cut headcount and we will not pretend otherwise. From the AI Index 2026 report: one third of organizations expect to reduce their workforce over the coming year, but almost half expect no change at all. The visible effect so far is narrow, concentrated in the employment of the youngest developers. We start where people are buried in repetitive work, because that is where the return is fastest and nobody has to lose a job for the deployment to pay off.
Is this compliant with the EU AI Act and GDPR?
The AI Act's first prohibitions took effect in 2025, so compliance is part of the project from the start rather than after it. We prepare a risk classification for the deployment, clear disclosure that the user is talking to a system, an operations register, a description of the data and the grounds for processing it, and technical documentation. We are not a law firm and will not pretend to be one: we supply the documentation and work with your lawyer or data protection officer.
Which models do you work with?
We do not lock you to one provider. The AI Index 2026 report shows that leading models are now nearly indistinguishable from one another and that open-weight models are more competitive than ever — at that rate of change, choosing one forever is an architectural mistake. We pick the model per task and build the deployment so it can be swapped without a rewrite. Where data cannot leave the company, we work with a model running on your side.
Does an agent deployment make sense for a small company?
It does, if there is a process repeated often enough to be measured. The report shows smaller organizations are clearly behind: among companies with revenue under $100 million, 39% are still only experimenting with AI and just 5% have fully scaled deployments. For a smaller company that is good news, because the entry cost today is low: one process, one pilot, a calculated return. You do not need a company-wide platform to start.
Will the AI learn our documents, price lists and procedures?
Not in the sense that usually implies — and it does not need to. We do not train the model on your files and we upload nothing into it. Your documents stay in your systems. With every question the system searches them for matching passages and only then composes an answer. It names the document it used. The mechanism is called RAG. In practice that means two things: a price list corrected on Friday works once the file has been reloaded, usually within minutes, with no retraining, and every answer can be verified at the source instead of taken on trust. When the answer is not in the documents, the agent is meant to say "I don't know" and hand the case to a person.
Our documents are a mess. Does that rule out a deployment?
It does not rule it out, it moves the starting point. An agent answers from whatever it is given: if three versions of the price list are in circulation and none is marked as binding, it will quote one of them, and it will do so in a very confident tone. That is why the audit is the first stage and not a formality — we check what exists, where it lives, which version is in force and what nobody ever wrote down. Part of the knowledge sits only in people's heads, and we capture that in short conversations with your team. If the audit shows the documents have to be put in order first, we say so plainly instead of launching a pilot that ends in disappointment on both sides. That stage has value for your team even if you ultimately do nothing with an agent.
How does a workflow differ from an AI agent, and which is cheaper?
A workflow is a written procedure that runs itself: the steps are fixed in advance, so the same input produces the same run. An agent is given a goal and decides the order itself, because what arrives at the input cannot be predicted. A workflow is cheaper, faster to deploy and easier to verify. The test is simple. Can you write the process out as points, without the sentence "it depends"? Build a workflow. Does the case have to be read and understood first, to know where to route it? Then we talk about an agent. Very often the best deployment combines the two: the workflow carries the process, and the model handles the one step that requires reading with understanding. When plain automation is enough, we say so and point you to the cheaper service.
Let's start with one process
Fifteen minutes on a call is enough to tell whether your company has a process an agent would do better — and whether it is worth touching at all.