Lab · You are here
LAB · RESEARCH & INSIGHTSAI Agents and Our Work
What AI agents are, what they are really doing in workplaces today, the honest limits behind the hype — and why people must stay in charge.
What AI agents are, what they are really doing in workplaces today, the honest limits behind the hype — and why people must stay in charge.
“AI agents” is the phrase of the moment — in job adverts, in sales pitches, and increasingly in the tools our employers use. This article explains, in plain language, what an AI agent actually is, what these systems are really doing in workplaces today, where they fail, and what workers and communities should insist on: people stay in charge.
Anthropic, one of the leading AI labs, draws the line this way: in a workflow, software follows steps a person wrote in advance. In an agent, the AI model itself decides the next steps — it works in a loop: it reads a goal, plans, uses tools such as search, a calendar or a customer database, checks the result, and decides what to do next. The working shorthand is: Agent = AI model + memory + planning + tools.
That is what makes an agent different from the tools most of us already know:
Answers questions in conversation. It talks — it does not act on other systems, and it does not decide a sequence of actions.
Follows fixed rules a developer set in advance: when this happens, do that. Reliable — but it breaks when something unexpected appears.
Chooses its own sequence of actions at run time, using real tools, and adjusts when results surprise it. That flexibility is both its value and its risk.
Anthropic’s own advice is deliberately anti-hype: use the simplest tool that works, and only add an agent when the task truly needs judgement step by step. Many “agents” on sale today are ordinary chatbots or automation with a new label — a practice Gartner calls agent washing.
The deployments that work reliably today share a pattern: the task is routine, the information is structured, and the result can be checked. On that evidence, these are working now:
Answering routine calls and messages around the clock, collecting names, needs and contact details, and logging them for staff.
Checking enquiries against written rules — the right service, the right area, the right timing — and passing the right ones to a person. It works only when the rules are written down; “judge quality like I do” does not work.
Turning a conversation into clean records: contacts created, fields filled, next steps noted. Among the most reliable agent tasks, because a person can verify the record afterwards.
Writing first drafts of replies, quotes and routine documents for a person to check, edit and approve before anything leaves the building.
Checking a real calendar live and booking confirmed slots. The failures come when an agent guesses availability instead of querying the calendar.
Answering frequent questions grounded in an organisation’s own documents — opening hours, policies, status — and handing the rest to a human.
The evidence, labelled honestly. Vendasta reports (vendor-reported) that its AI receptionist handled 1,017 inbound calls for one car-rental business and qualified 778 of them. Intercom reports its support agent Fin resolves about 67% of conversations on average across its customers (vendor-reported); independent readings of published case studies put typical real-world resolution nearer 42–50%. Both halves belong together: the routine majority is real, and so is the gap.
MarketsandMarkets estimates the AI agents market at $7.84 billion in 2025 and forecasts $52.62 billion by 2030 (46.3% yearly growth). A forecast from a commercial research firm — a projection, not measured revenue.
Gartner forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027 — because of escalating costs, unclear value and inadequate risk controls.
Gartner also estimates that only about 130 of the thousands of vendors selling “agentic AI” are real; the rest is agent washing — assistants, automation and chatbots relabelled.
Independent evaluations by METR find today’s frontier agents succeed almost 100% of the time on tasks a person finishes in under 4 minutes — and under 10% on tasks that take a person more than 4 hours. Errors compound, step after step.
MIT’s 2025 “GenAI Divide” study found about 95% of enterprise AI pilots delivered no measurable impact on profit. It was a preliminary study and the figure is about business impact, not technical failure — but it is a standing warning against buying hype.
Klarna’s AI assistant handled 2.3 million chats in its first month — work the company compared to 700 agents (company-reported). In May 2025 Klarna began rehiring people after quality complaints, and its leadership conceded cost had been weighted over quality. The steady state it landed on is hybrid: AI for the routine, humans for the complex and sensitive.
An agent that can act is an agent that can act wrongly — send the wrong message, update the wrong record, or state a policy that does not exist. The proven protections are not technical magic; they are the same disciplines any accountable workplace runs on:
Nothing an agent drafts should reach a customer — and nothing about price, refunds, timelines or eligibility should be promised — until a named person approves it. OpenAI builds this into its agent tools as a formal approval step that pauses the work before any side effect.
Open →When Air Canada’s chatbot invented a bereavement-fare policy, a Canadian tribunal in 2024 rejected the airline’s claim that the chatbot was a separate entity, held the airline responsible for what its chatbot said, and ordered CA$812.02 paid. “The AI said it” is not a defence.
Open →Every agent action should leave an audit trail: what it did, what information it used, who approved it, and when. If there is no record, there is no accountability — for the worker affected or the customer either.
If an automated system affects your work, your application or your pay, ask who oversees it, what data it used, and how to reach a human. Our Worker Rights pages list the official channels that hold real power.
Open →One more risk deserves plain words: prompt injection — instructions hidden in a message, email or web page that try to hijack an agent into ignoring its rules. The OWASP security community ranks it the number-one risk for AI applications, and there is no complete fix yet. That is exactly why the approval gate matters: a draft can be checked; an action already taken cannot be untaken.
AI & Community is one of the Lab’s six study areas, and our position follows the evidence in this article:
For workers, the practical skill of the next few years is not competing with these systems — it is supervising them: writing the rules clearly, checking the drafts, and knowing your rights when a system gets it wrong. That is workforce education we intend to build, and it starts with honest information like this.
Reliability figures for long tasks are from METR’s published evaluations, and the Air Canada decision is Moffatt v. Air Canada, 2024 BCCRT 149 — both as reviewed in the Lab’s research notes. Bulsho Forum provides information and education, not legal advice.