AI agents in your business: where to start and why projects stall

What is an AI agent?

An AI agent is a program that, given a goal, decides which steps to take to reach it and carries them out in other software. A chat assistant answers a question. An agent does the work the answer calls for: it reads the incoming email, finds the company in the CRM, opens a new quote request and assigns a task to the right person.

An example makes the difference clear. An assistant asked "does this customer have a late payment?" reads the accounting record and tells you. An agent finds late payments every morning, drafts the reminder email to the customer and puts it in front of the person responsible for approval. An agent is a workflow, not a report.

Which job should you start with?

A good first job has four traits: it repeats often, its input is digital, the correct result is clear, and mistakes can be undone. Examples that tick all four (our AI and automation projects usually start with jobs like these):

• Reading requests from web forms or email and logging them in the CRM under the right category. WhatsApp messages count too; we worked out their cost in WhatsApp Business API pricing.
• Reading the amount, date and tax number on supplier invoices and entering them in accounting software as drafts.
• Matching support tickets to similar past tickets and preparing a reply draft for the agent on duty.
• Flagging missing clauses in contracts or quotes.

A bad first job is broad and hard to measure, such as "help the sales team" or "answer any question about the company". With goals like these, success is open to debate and it becomes hard to make a decision at the end of the pilot.

Why do agent projects stall?

In a forecast published in June 2025, Gartner said over 40 percent of agentic AI projects will be cancelled by the end of 2027, pointing to escalating costs, unclear business value and inadequate risk controls. The same release estimates that only about 130 of the thousands of vendors calling themselves agentic offer real agent capabilities. Gartner calls the relabelling of older chatbots and automation tools "agent washing".

What we see in practice matches this. Projects usually get stuck in one of three places:

• There is no measure. The pilot ends with "seems good", but nobody knows how many jobs the agent got right or how many a person had to fix. A pilot without a decision drags on and eventually gets shut down.
• Integration is underestimated. The model is tried out in a few days. The real work is letting the agent write safely to the CRM, ERP or email. Many older systems have no proper API, and this part takes most of the budget.
• Cost is not worked out upfront. Every agent step is a model call. If calls per job are not measured, the bill grows unexpectedly as usage rises.

How do you measure a pilot?

Before the pilot, collect 20 to 30 real cases done by hand and write down their correct results. Run the agent on them and look at three numbers: the share of jobs it got right with no correction, the share a person corrected slightly, and the share that went wrong from the start. Add time per job and model cost per job.

Without these numbers, "does the agent work?" is a matter of opinion. With them, the decision is simple: compare the share of jobs needing no correction and the cost per job with what the job costs by hand.

Where does human approval belong?

At first, the agent does the work and a person approves the last step. Approval must stay on for steps that are hard to undo, such as payments, emails to customers, deleting records or changing prices. As the agent's accuracy passes a set threshold, approval can be lifted for low-risk steps. Every step should be written to a log so that when something goes wrong you can see what was done, when and why.