AI agents for business: working systems, not demos

What is an AI agent for a business?

An AI agent is a program that uses a language model to decide which action to take next, then calls tools to do it: query a database, create a ticket, draft a reply, look up an order in Shopify, post a journal entry to QuickBooks Online. The model chooses the step. Your code controls which tools exist and what each one is allowed to do.

That split matters. A chatbot answers questions. An agent changes state in your systems, so its mistakes cost more and its design needs more care than a prompt in a chat window.

Common builds include a back-office agent that reads supplier emails and updates purchase orders, a front-desk agent that answers booking questions from a property management system, and a sales-support agent that assembles account history from a CRM before a call. Each one is a small set of tools plus a policy, which is why the first version can ship quickly.

Where agents earn their keep

The tasks that suit agents have three properties: the input is messy text or documents, the correct action comes from a small set of options, and a wrong answer is cheap to catch. Examples include sorting inbound requests, pulling fields out of supplier paperwork, preparing a first draft of a quote, and reconciling two lists that never quite match.

Tasks that suit agents poorly include anything with no feedback signal, such as a vague goal like improving customer satisfaction. Give the agent a specific job with a measurable result: tickets routed correctly, fields extracted correctly, replies accepted without edits.

Hospitality, real estate, aviation and manufacturing each have their own versions. A hotel agent can check a reservation in the PMS and draft a response to a guest. A property agent can match a maintenance request to a unit and a vendor. A parts planner can compare a delay notice against open work orders. The pattern stays the same while the tools change.

• Inbound email and web form routing to the right queue with a suggested reply
• Supplier or guest documents turned into structured records
• Status checks that span three systems and today require three logins
• Draft-then-approve work such as proposals, dispute responses and collection notices

How much of the market is using AI already?

The Census Bureau survey puts AI use at 19.8% of US businesses as of May 3, 2026. The Federal Reserve notes a much larger figure when firms are weighted by headcount: 78% of the labor force works at a firm that has adopted AI. In practice, large employers adopted first and many small and mid-size firms are still deciding where to begin.

Where agents break

Agents fail in predictable ways. A tool call returns an error and the model retries it five times. An API rate limit is hit during a batch. The model reads an ambiguous instruction and picks the wrong customer record. A prompt that worked on ten test emails meets a hundred real ones with attachments, forwarded chains and signatures in three languages.

Each of these has an engineering answer: idempotency keys on write calls, a hard cap on steps per run, a lookup that requires an exact ID match before any write, and a review queue for anything below a confidence threshold. Vascoh builds those controls first and the clever part second.

Cost is a failure mode too. An agent in a loop can spend a large number of model calls on one stuck case. Set a budget per run and per day, alert when either is reached, and log the cost beside the outcome so you can see what each completed task actually costs.

Keeping a person in the loop

Start every agent in suggest mode. It writes its proposed action to a queue, and a staff member approves or edits it. The edits are the most useful data you will get, because they show exactly where the agent and your team disagree. Once approval rates are stable on a narrow task, that task can run without review while the broader work stays supervised.

Plan who reviews the queue. If nobody owns it, approvals pile up and the agent becomes an extra inbox. Name a person, set a response window, and report weekly on volume, approvals and edits.