AI Agents vs Chatbots: Most Tasks Don't Need an Agent

The opening claim

An agent is a chatbot with a while loop and a credit card.

That is the whole difference. Everything else in the marketing is decoration.

You ask a chatbot something. It answers. It stops. You give an agent a goal. It plans, calls a tool, reads the result, notices it was wrong, calls another tool, and keeps going until it decides it's finished or until you kill it. The loop is the product. The loop is also the bill.

Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear value and weak risk controls. In the same release they estimate only about 130 of the thousands of vendors selling "agentic AI" are actually doing anything agentic. They have a word for the rest: agent washing.

The mechanism 🧠

Three levels get mashed into one word. They are not the same thing.

Level 1, the chatbot. One call in, one answer out. It touches nothing.

Level 2, tool calling. You hand the model a list of functions. It picks one, your code runs it, you feed the result back. Your code still decides what happens next. Anthropic calls this a workflow: "systems where LLMs and tools are orchestrated through predefined code paths."

Level 3, the agent. The model decides what happens next. Same source: "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." No fixed path. It loops.

Most things sold to you as agents are Level 2 with a nice dashboard.

Now the part nobody puts on the sales deck. A May 2025 paper found that agents fail at a roughly constant rate per minute of human work. Constant failure rate means success decays exponentially with task length. Every agent has a half life.

Do the arithmetic yourself. An agent that is right 95% of the time on a single step is right 0.95^20, about 36%, across twenty steps. That is not a bug in the model. That's multiplication.

What the benchmarks actually say

Read that table top to bottom. There's a gradient in it.

The closer a benchmark gets to a real job with real consequences, the worse agents do. Curated coding problems the model has probably seen: 95%. The same coding problems in repos it hasn't seen: 46%. Actual office work with finance and admin tasks: 30%. And the number that matters most for anything unattended, the same task eight times in a row without a single miss, collapses.

The tau-bench figure is old and I'm flagging it. But the 2026 reliability work reproduces the shape: pass@1 flatters agents, repeated trials expose them.

The cost problem

Every step of the loop resends the whole conversation so far. Context grows with the number of steps, and you pay for the whole context on every step. So your input bill grows roughly with the square of the step count. Double the steps, quadruple the input tokens.

Anthropic's own engineering team put a number on it: agents use about 4x the tokens of a chat, and multiple agents working together use about 15x.

At current Claude API rates, Sonnet 5 is $2 per million input tokens and $10 per million output. Sounds like nothing. Then it runs for six hours.

Real bills, all from TechCrunch, 5 June 2026:

Here's the part people miss. The expensive failure isn't a wrong answer. It's an agent that gets stuck in a loop, retries the same broken call two hundred times, and bills you full price for producing nothing. A chatbot that fails costs you a fraction of a cent. An agent that fails can run all night.

The Algerian reality

An agent needs three things before it can do anything: an API to call, structured data to read, and a way to check its own work.

Count how many of those an Algerian merchant has.

Payment rails. GIE Monétique's 2025 report (published Feb 2026): 644 web merchants are integrated into the national online payment platform. In the entire country. Meanwhile the CNRC counted 2,419,913 registered operators at 10 December 2025, of which 1,092,176 are retail merchants.

Six retail merchants in every ten thousand can take a card payment online. That's the digital economy an agent is supposed to be acting inside.

Data. Stock lives in a notebook or a WhatsApp thread. Orders arrive as Instagram DMs. Returns are settled verbally with a delivery driver. There's no database to query, no order table, no state to reason about. An agent with nothing to read is a very expensive chatbot.

Language. On DialectalArabicMMLU, 19 Arabic capable open weight models averaged 51.9% on Modern Standard Arabic and 45.0% on Maghrebi. Those are small open models, not frontier ones, so don't over read it. But your customers write in Darija, in Latin letters, with French words dropped in, and nothing about that makes the loop more reliable.

The bill is in dollars. Your margin is in dinars.

One rail does exist, and it's worth saying: delivery. Yalidine, ZR Express, Maystro and Noest all expose APIs, and Yalidine's is genuinely well documented. If any automation pays for itself here, it starts there. Not with an agent. With a script.

Where agents actually work

Three profiles, and they're narrow.

Coding with tests. The output is machine verifiable, retries are cheap, and the task is expensive to do by hand. This is the one place agents are unambiguously earning money in 2026.

Deep research. Anything where a human would spend four hours reading sixty pages. Parallel, tolerant of noise, and you check the result.

Support, but only on top of real documentation. Intercom reports Fin averaging 76% automatic resolution across 12,000 customers. Vendor number, treat it accordingly. But read their own caveat: deployments that neglect the knowledge base stall at 30% to 45%. The agent's ceiling is set by the quality of the documentation you feed it. If you have no documentation, you have no agent.

And the honest counterweight: Klarna replaced 700 agents' worth of support with AI, then started rehiring humans in 2025. Their CEO: "We focused too much on efficiency and cost. The result was lower quality, and that's not sustainable."

What to actually do 🛠️

Use an agent only if all four are true.

1. The task takes a human more than 30 minutes.
2. The output is verifiable, either by a machine or by you in under a minute.
3. A wrong answer is cheap to throw away.
4. An API already exists. Not "we'll build one." One that exists today.

If a single one of those is false, ask the model the question directly and move on with your day. That covers most of what you need.

If you do run one: hard step cap. Hard token budget with a kill switch. Log every tool call. And never let it write to your stock or trigger a shipment without a human pressing the button.

Anthropic's own guidance, from the people selling the models: "find the simplest solution possible, and only increase complexity when needed. This might mean not building agentic systems at all."

TL;DR 🧾

An agent is a chatbot with a while loop and a credit card. The loop is the value, and the loop is the bill.

The best agents in the world complete about 30% of real office tasks unsupervised. Agents burn roughly 4x the tokens of a chat, and the cost of a loop grows with the square of its length, so a stuck agent bills you all night for nothing. In Algeria, 644 merchants out of 1.09 million can take a card payment online, which means the APIs and the structured data an agent needs mostly don't exist yet.

Ask the question. Get the answer. Move on. The marketing pushes agents because agents sell.