AI agents for business: why it looks simple in the demo and becomes complex and expensive in reality

-

Over the past year it’s been hard to open social media without seeing a new promise:

An AI agent that will answer customers, write content, manage the calendar, send out quotes, follow up on leads, update systems and maybe even run the business while we sleep.

In a one-minute video, everything looks perfect.

You press a button, give the agent one instruction and it starts working. It reads messages, connects information, sends replies and takes actions without anyone needing to step in.

But then you try to bring the same system into a real business.

Suddenly you discover that the information is spread across WhatsApp, emails, a CRM, Excel files, a website nobody has updated in two years and different people in the business who each work in their own way.

And that’s exactly where the simple promise starts to get complicated.

First of all: what is an AI agent?

An AI agent isn’t just a chat that answers questions.

It’s a mechanism that receives a goal, decides which actions to take, uses external tools and systems, and checks whether it has completed the task.

For example, an agent could in theory receive a new lead, check which campaign it came from, read the enquiry, classify it, send a reply, suggest a meeting time and update the customer management system.

For all this to work, the agent needs access to all the systems, clear instructions, the right permissions and the ability to deal with situations where something doesn’t work as planned.

OpenAI’s official guides also stress that a real agent requires a combination of a model, tools, instructions, safeguards and the option to hand control back to a person when the system fails or reaches a sensitive action.

The first problem: the business itself doesn’t always follow a clear process

Before you teach an AI agent to carry out a process, you need to define the process.

Who receives a new lead?
How quickly do you get back to them?
What counts as a quality lead?
When do you send a price?
Who is allowed to give a discount?
What happens when there’s no answer?
What do you do if the customer writes something that doesn’t fit any of the defined options?

In many businesses, the answer is: “It depends.”

An experienced employee understands the context, knows the customer and knows when to step outside the procedure. An AI agent needs to receive that knowledge in an organised way, or know when to stop and ask a person.

Sometimes the biggest investment isn’t in technology, but in mapping and organising processes the business never defined.

The second problem: the information isn’t organised

An AI agent can only work well when it receives correct, relevant information.

In reality, business information usually includes:

  • Old and new prices in the same file.
  • Services the business no longer provides.
  • Several versions of the same quote.
  • Missing customer details.
  • Documents without clear names.
  • Important information that exists only in the business owner’s head.

When the agent receives partial, contradictory or outdated information, it can give an answer that sounds great but isn’t correct.

So serious AI implementation first requires cleaning up information, setting a source of truth and defining who is responsible for keeping it up to date.

The third problem: every system has to connect to something else

The agent doesn’t really “do everything” on its own.

It needs to connect to email, the calendar, the CRM, the website, WhatsApp, the invoicing system, forms and every other system the business uses.

Each such connection requires:

  • Permissions.
  • An API or automation tool.
  • Field mapping.
  • Testing.
  • Error handling.
  • Security.
  • Maintenance when one of the systems changes.

Even a process that sounds simple, like “send a message to a new lead”, can get stuck if the phone number arrives in the wrong format, a field is missing, the connection has expired or the customer sent a voice message instead of text.

The more tools and freedom you give the agent, the more possible points of failure there are.

Anthropic recommends starting with the simplest solution, because complex agent systems add costs, response time and the risk of errors that build up along the chain of actions. In many cases, a simple, fixed automated process can be more effective than a fully autonomous agent.

The fourth problem: the real world is full of exceptions

In the demo, a tidy enquiry always arrives:

> Hello, I’m interested in the service. I’d be happy to get a quote.

In reality, a message like this arrives:

> mate I talked to you once about that thing is it still relevant?

Or a two-minute voice message, a screenshot with no explanation, a wrong phone number or a customer asking for something the business doesn’t offer exactly – but might be able to solve another way.

This is where the real work begins.

You need to define when the agent continues, when it asks for clarification and when it hands the conversation to a person.

An agent that stops at every question doesn’t really save work. An agent that never stops can misread the customer and take an action that shouldn’t have happened.

Even the AI companies themselves describe the balance between the agent’s independence and human control as one of the main problems in implementing these systems.

The fifth problem: you need to test, monitor and maintain

An AI system isn’t a project you set up once and forget about.

You need to track:

  • Mistakes in answers.
  • Actions that weren’t completed.
  • Usage costs.
  • Response times.
  • Changes in connected systems.
  • New cases that weren’t in the original plan.
  • Changes in behaviour after a model or prompt update.

In regular systems you can check whether an action was carried out or not. With AI agents there’s also a layer of judgement, so you need to check not only whether the agent took an action – but whether it made the right decision.

Anthropic notes that agent systems are harder to evaluate because of the many steps, the use of tools and the changes in state along the way. That’s why you need structured testing, success metrics and tracking of cost, time and errors.

This gap shows up in research too: in a benchmark published in 2026 that tested long, realistic browsing tasks across several websites, the strongest agents successfully completed only 44.5% of the tasks tested. This isn’t proof that agents “don’t work”, but an illustration that long, multi-step tasks are still very different from short demos.

So why does it get expensive?

Many people look only at the subscription price of an AI model.

But that’s usually just one line in the total cost.

A real system can include:

  • Process design.
  • Development and integrations.
  • Automation systems.
  • AI model usage.
  • Data storage.
  • Security and permissions.
  • Testing and quality control.
  • Maintenance.
  • Fixing problems.
  • Time from the team members who oversee the system.

The more critical the process, the more safeguards and controls are needed.

An agent that drafts a post can make a mistake without causing significant damage. An agent that sends quotes, changes ad budgets, issues a credit or makes a commitment to a customer needs to work under completely different rules.

NIST’s risk management frameworks also stress that responsible use of AI requires attention to reliability, security, privacy, transparency, testing and risk management throughout the system’s life cycle.

Where does AI already work really well today?

After all this, it’s important not to fall to the other extreme.

AI isn’t a bluff. It’s a very significant tool when it’s used in the right place.

It can be excellent for:

  • Summarising conversations and documents.
  • Preparing drafts.
  • Sorting and classifying enquiries.
  • Extracting information from forms and emails.
  • Suggesting replies to a service representative.
  • Preparing initial reports.
  • Creating variations of content and ads.
  • Spotting anomalies and flagging tasks to a person.
  • Carrying out defined processes under supervision.

The mistake isn’t using AI.

The mistake is starting from the question: how do I get the agent to do everything?

Instead of asking: which one process in the business eats our time, repeats itself and has success that can be measured?

How to implement an AI agent in a business the right way

The right way isn’t to start with a huge system.

Start with one narrow, clear process.

For example:

  • Receiving a lead and summarising it.
  • Preparing a draft reply.
  • Sorting enquiries by service.
  • Summarising a sales call.
  • Creating follow-up tasks.
  • Preparing a weekly report.

Define what counts as success, which actions require human approval and what happens when there’s doubt.

Only after the process works reliably and proves real savings do you add another stage.

That’s how you build a system that serves the business – not a system the business has to serve.

The bottom line

AI agents can save time, improve processes and let businesses do more with less manual work.

But they aren’t magic employees that come ready out of the box.

Behind an agent that works well there’s an organised process, clean information, integrations, permissions, testing, oversight and maintenance.

Everything can look easy on paper or in a demo video. The real challenge is bringing the technology down to earth and connecting it to a real business — with real people, real systems and problems that don’t always fit the script.

At Barak Digital we believe the goal isn’t to put AI everywhere just because you can.

The goal is to identify where it really saves time, improves work and creates value – and to keep people at the points where experience, responsibility and judgement are still needed.

Read more about an experiment in which an AI agent tried to build an independent business.