AI Agents for Business Automation

Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls. The same analysts estimate that of the thousands of vendors selling "agents", only about 130 sell something that deserves the name.

That is the backdrop for every sales pitch about AI agents for business automation you will hear this year. Some of those pitches are describing a real and useful thing. Most are describing a chatbot or a scripted workflow with a new label on it. The difference matters, because the real thing costs more to build, more to run and more to supervise, and it only pays off for a specific kind of task.

This post sets out what an AI agent actually is, how it differs from the automation most businesses should be building first, where agents genuinely earn their keep, where they go wrong, and what a sensible first project looks like for a company in Bulgaria.

What an AI agent is, and what it is not

The cleanest definition we know comes from Anthropic's engineering team, which draws a line between workflows and agents. A workflow is a system where language models and tools are orchestrated through predefined code paths. An agent is a system where the model directs its own process and decides for itself which tools to use and in what order.

In a workflow, a developer has already decided the steps. Read the email, extract the order details, look up the customer, create the record, send the confirmation. The model does the reading and the extracting. The code does everything else, in a fixed order, every time.

In an agent, the model is given a goal, a set of tools and a budget, and it works out the steps. Find out why this customer's invoice does not match the delivery note. The agent might pull the invoice, pull the delivery note, notice a quantity mismatch, check the original order, find a partial shipment, and draft an email to the supplier. Nobody scripted that sequence. The model chose it, and on the next case it might choose a different one.

Everything you have seen sold as an agent sits somewhere between those two poles. A support chatbot that answers from your documentation is a workflow with retrieval attached, which is the pattern behind RAG chatbots. A Zapier chain with one model step in the middle is a workflow. Robotic process automation replaying recorded clicks is a workflow from a decade ago. None of them are agents, however the brochure describes them, and Gartner has a name for the relabelling: agent washing.

This is not a pedantic distinction. The two designs fail differently, cost differently and need different supervision. If a vendor cannot tell you which one they are selling, that is the first thing to find out.

Workflow or agent: the test that decides

Anthropic's advice to its own customers is blunt: find the simplest solution possible, and only add complexity when it demonstrably improves outcomes. For most business automation that means a workflow, and often a single model call with the right context in front of it.

The question that separates the two is whether you can write the steps down in advance. If a competent new employee could follow a checklist to do the task, with the model handling the parts that need reading or judgement, build a workflow. Workflows are predictable, cheap to run and easy to test, because the same input takes the same path every time.

Agents belong where the number of steps is unknown until you are in the middle of the task. Investigating a discrepancy. Researching a prospect across six sources before a sales call. Chasing a shipment through a courier portal, an email thread and a spreadsheet, where the next step depends on what the last one turned up. If the checklist would have forty branches, or you cannot write it at all, that is agent territory.

• Build a workflow when the steps are known, the volume is high and a wrong answer is cheap to catch. Invoice coding, order intake, support triage, lead enrichment. This is where most AI automation in small businesses actually lives, and it is where the money is.
• Build an agent when the task is open-ended, each case is different, and the value of a correct outcome is high enough to justify a system that costs more per run and needs more watching.
• Build neither when the volume is low. A task that happens twice a month will not repay either design, however tedious it is for the person doing it.

We hold ourselves to this. Boro, the bilingual property assistant we built for Houzez, is a workflow with retrieval and a handful of fixed tool calls. It could have been an agent. It would have been slower, more expensive and harder to trust, and the buyers using it would not have noticed any benefit.

Where agents earn their keep in a Bulgarian business

Bulgaria starts from a low base. In 2025, 8.5% of Bulgarian companies with ten or more employees used AI for anything, against an EU average of 20%. That means almost every company reading this has not yet built its first workflow, let alone its first agent. The honest advice is to build the workflow first. Still, there are tasks inside ordinary Bulgarian companies where an agent is the right design from the start.

• Supplier reconciliation. A distributor in Plovdiv with two hundred suppliers gets invoices that do not match purchase orders for reasons nobody can predict: partial shipments, substituted products, a price change agreed by phone. An agent that can read the invoice, query the ERP, read the email thread and draft a query to the supplier replaces hours of detective work a week. The steps differ every time, which is exactly why a scripted workflow struggles.
• Pre-sales research. Before a meeting, an agent pulls the prospect's company registration, recent news, website, job listings and your own CRM history, then writes a one-page brief. The sources it needs differ for a Sofia software firm and a Varna hotel group, and it has to decide which ones matter.
• Multi-system customer cases. An order is late. The answer lives in the shop platform, the courier's tracking page and a payment provider's dashboard, and which one explains the delay is unknown until you look. An agent can look in all three and tell the customer, or a human, what actually happened.
• Internal operations with messy inputs. Chasing missing documents for a tender, assembling a monthly report from five spreadsheets that change shape, or triaging a shared inbox where a quarter of the messages need three systems checked before anyone can reply.

The common thread is investigation. Agents are good at tasks where a person would otherwise open several tabs, poke around and work out what is going on. They are a poor fit for tasks that are simply high volume. High volume wants a workflow, because a workflow does the same thing a million times for almost nothing.

Where agents go wrong

Errors compound. A workflow with one model step that is right 95% of the time is wrong one case in twenty, and you can catch that at review. An agent that takes ten steps, each right 95% of the time, finishes the whole task correctly about 60% of the time, because each step inherits the mistakes of the one before. The maths is unforgiving, and it is the single biggest reason agent projects stall after the demo.

Costs run away. A workflow makes a known number of model calls. An agent makes as many as it decides it needs, and a confused agent can loop, retrying the same search or re-reading the same document until somebody notices the bill. Every production agent needs a hard cap on steps and spend per run, set before the first real case, not after the first surprising invoice.

Permissions are the quiet problem. An agent is only useful if it can act: send the email, update the record, issue the refund. Those are exactly the actions you do not want a probabilistic system taking unsupervised. The usual answer is to let the agent read anything and draft anything, but require a person to approve anything that writes to a customer or moves money. That approval step is where most of the design effort should go.

Auditability catches companies late. When an agent reaches a wrong conclusion, somebody has to work out why, and that means every step, every tool call and every intermediate result has to be logged in a form a non-engineer can read. Gartner's phrase for the third cause of cancelled projects is "inadequate risk controls". In practice that means nobody could explain what the agent did on Tuesday.

There is a regulatory layer too. The EU AI Act became applicable on 2 August 2026, and an agent that corresponds with customers has to make clear it is a machine. Most business agents sit in the minimal-risk band, but anything touching hiring, credit or access to essential services is high-risk and needs documentation from day one. Treat those as a separate class of project.

What AI agents cost to build and run

Model calls are cheap. That has been true for a while and it stays true for agents, with one caveat: an agent makes many more of them per task than a workflow does. A single workflow run might cost a fraction of a cent. An agent run that reads a dozen documents and makes fifteen tool calls can cost tens of cents, and a few hundred such runs a day is real money. Price the running cost per completed case, not per model call.

The build is where the budget goes. A narrow workflow automation usually lands in the low thousands of euros, which our pricing page sets out in more detail. An agent costs more, because the work is in the parts around the model rather than the model itself: the tools it can call, the permission layer, the step and spend caps, the logging, the review screen a person uses to approve or correct its actions, and the test set that proves it works. Expect a first production agent to cost several times what a first workflow costs, and to take longer.

The sums are the same as for any automation. Count the hours the task eats each month, multiply by what those hours cost you, and take the share an agent can realistically absorb. For investigation-type tasks that share is often lower than for workflows, because a person still reviews the output. If the saving does not clear the build cost inside a year, build something else, or build a workflow for the part of the task that is predictable and leave the rest with a human.

How to build a first agent that survives contact with reality

Pick one task where a person currently investigates, not one where they process. Write down what a good outcome looks like, and collect twenty or thirty real past cases with the answer a human reached. That set is your test. Without it you have a demo, and demos are what Gartner expects to be cancelled.

• Start the agent read-only. Give it tools that look things up and none that change anything. Let it draft the email, the ERP update or the customer reply, and measure how often a person would send the draft unchanged.
• Add write actions one at a time, each behind a human approval step. Refunds and anything sent to a customer stay behind approval for good.
• Cap steps and spend per run from the first day. Ten steps and a fixed budget in cents is a reasonable starting point for most business tasks, and the cap tells you when the agent is lost.
• Log every step in plain language and keep the log where operations staff can read it, not just engineers.
• Re-run the test set on a schedule. Models change behind the same API name, and your suppliers, customers and documents change too. A score that drifts is your early warning.
• Decide in advance what rate of correct outcomes means the project continues. If a two-week prototype does not reach it against real cases, the right answer is often a workflow for the predictable part of the task, and that is a good result, not a failure.

That last point is why our AI automation work begins with a short proof of concept against your own data rather than a platform licence. The deliverable is a working system, a measured score, and a recommendation, which is sometimes "build the workflow instead" and occasionally "do not build this at all".

The verdict

AI agents are real, and they are the right design for a narrow set of tasks: open-ended investigation across several systems, where each case is different and a correct outcome is worth enough to pay for the supervision. For almost everything else a business wants automated, a workflow is cheaper, more predictable and easier to trust, and it should be built first.

Our view is that the companies in Bulgaria who get value from agents in the next two years will be the ones that already run two or three workflows, know their accuracy numbers, and have a human review habit in place. Buying an "agent platform" before any of that exists is how a project joins the 40%. Automate the predictable task first, measure it honestly, and let the agent be the second project, paid for by the first.

Frequently asked questions

What is the difference between an AI agent and AI automation?

AI automation is the broad category: handing a task that needs reading or judgement to a model, inside software you already use. An AI agent is one way to build it, where the model decides its own steps and tool calls. A workflow is the other way, where a developer fixes the steps and the model handles only the interpretation. Most AI automation in business is, and should be, workflows.

Are AI agents the same as chatbots?

No. A chatbot answers questions, usually from a fixed knowledge base, and does not act on anything. An agent is given a goal and tools, and takes actions across systems to reach it. Some chatbots have agent features bolted on, such as the ability to look up an order or issue a return, but the conversation is not what makes something an agent. The autonomy is.

Can an AI agent work in Bulgarian?

Yes. The major models read and write Bulgarian well, and mixed Bulgarian and English input, which is what most Bulgarian inboxes and documents contain, needs no special handling. The harder language question is your own systems: an agent can only act through tools, and if your ERP or courier portal has no API, that part of the task still needs a person or a scripted workaround.

How much does an AI agent cost for a small business?

More than a workflow for the same task, mainly because of the permission layer, the logging and the testing an agent needs before anyone should trust it. A narrow workflow automation lands in the low thousands of euros to build. A first production agent typically costs several times that. Running costs are measured per completed case, and for an agent that makes many tool calls they are tens of cents rather than fractions of a cent.

Do I need a developer to build an AI agent?

For anything that touches your real systems, yes. No-code agent builders can produce a convincing demo in an afternoon, but the parts that make an agent safe in production, which are tool permissions, spend caps, logging and a test set, are engineering work. A business owner can and should define the task, the good outcome and the approval rules. Somebody technical should build the rest.

How long does it take to build an AI agent?

Two to four weeks for a read-only prototype that runs against real past cases and produces a measured score. Production hardening, which means write actions behind approval, caps, logging and monitoring, usually adds one to two months on top. That is longer than a workflow, and it is one more reason to build the workflow first.