AI agents for business: where they earn their keep

Most of what gets called an AI agent today is a language model that can take a few actions on its own: read a message, look something up, fill in a record, draft a reply, and hand the result to a person or another system. That is genuinely useful. It is also a long way from the autonomous digital employee many vendor demos imply, and the gap between the two is where most disappointing projects end up.

This is how we think about where agents earn their keep in a business, where they do not, and how to start without betting a department on it.

What an AI agent actually is

An agent is a loop. It receives a goal or an input, decides which tool to use, uses it, looks at the result and decides what to do next, until it finishes or hits a limit. The tools are ordinary software: a CRM, a search index over your documents, an email inbox, a spreadsheet, a ticketing system.

That framing matters because it puts the weight where it belongs. An agent is only as good as the tools it can reach, the data behind them and the rules about what it is allowed to do. The model is the easy part to buy.

Use cases that usually work

• Support triage. Reading incoming tickets or WhatsApp messages, classifying them, pulling up the customer's order or account, and either drafting a reply for a person to approve or routing it to the right queue.
• Document processing. Extracting fields from invoices, purchase orders, claim forms or CVs into a structured record, and flagging anything it is not confident about for someone to check.
• Internal knowledge search. Answering staff questions from policies, product manuals and past tickets, with a link to the source so the answer can be verified.
• Sales and CRM housekeeping. Summarising calls and email threads, updating deal stages and drafting follow-ups — the admin sales teams skip when they are busy.
• Routine reporting. Pulling numbers from several systems each week and writing the first draft of the commentary, so an analyst edits instead of assembling.

What these have in common: the task is repetitive, the inputs are digital, the right answer can be checked, and a mistake is caught before it reaches a customer or a ledger.

Where agents disappoint

Open-ended judgement. Pricing a custom deal, approving credit or deciding which supplier to drop all depend on context that is rarely written down. An agent will produce a confident answer anyway, which is worse than no answer.

Messy or missing data. If the CRM is half-empty or the documents contradict each other, the agent inherits the mess and repeats it faster.

No human in the loop. Anything customer-facing or financial needs a review step, at least until you have months of evidence that the error rate is acceptable. Removing the review to save time is how a small error becomes a public one.

Costs nobody modelled. Each step an agent takes is a paid model call. A workflow that costs a fraction of a cent in a demo can become a real monthly line item at thousands of requests a day. Estimate it before building, rather than discovering it on the first invoice.

How to start without betting the business

1. Pick one workflow that is repetitive, high volume and currently done by hand. Write down how long it takes today and how often it goes wrong. That is your baseline.
2. Keep a person in the loop for the first version. The agent drafts or prepares; a human approves. You learn where it fails without your customers learning it for you.
3. Build it into existing tools. An agent that works inside the helpdesk or CRM your team already uses gets adopted. A separate portal gets forgotten.
4. Set guardrails and fallbacks. Define what it may never do, what happens when it is unsure, and who is told when it fails.
5. Model the running cost at realistic volume, then measure against the baseline after a month. Keep it, change it or switch it off based on the numbers.

A note on data

Before any company data goes to a model provider, check the provider's terms on training and retention, and prefer API tiers that do not train on your inputs. Document exactly which fields are sent where. In regulated sectors such as banking, insurance and health, involve compliance at the design stage, not after launch.

If you have a workflow in mind and want a straight answer on whether an agent is the right tool, our AI development team starts with a short readiness review. Sometimes the honest recommendation is a better form or a fixed report, and we will say so.