How is an AI agent different from a chatbot?

A chatbot answers the question you ask it and stops. An AI agent pursues a goal: it can plan several steps, use tools, check its own work, and move to the next action without being prompted individually for each one. The distinction is autonomy and action, not just conversation.

Is a local AI agent safe for confidential board work?

It removes the main exposure by design: because nothing is transmitted to an outside company, there is no external server to breach, no provider retaining a copy of your board papers, and no query log that could later be discovered. The UK's National Cyber Security Centre has warned that an agent's ability to access data and take actions is exactly why understanding what it can reach matters — running it locally keeps that reach confined to your own device.

Do AI agents work without an internet connection?

A genuinely local AI agent does, because the model and the reasoning both run on your own device. If an 'agent' tool stops working offline, some part of its task — the model, a tool call, or a data lookup — is happening on an outside server.

What Is a Local AI Agent? A Plain-English Guide for Professionals

Q: What is a local AI agent?

A local AI agent is an AI program that pursues a multi-step task and can use tools or take actions with some autonomy, running entirely on your own device rather than an outside company's servers. It works the way any AI agent does — planning steps, using tools, acting on a goal — but nothing about the process or the documents it touches leaves your machine.

Q: What can a local AI agent actually do today?

A local agent can read and cross-reference your own documents and follow a multi-step instruction — summarise this report, then compare it against last quarter's figures — without an internet connection. It is not yet a fully autonomous assistant that reliably browses the web or acts independently across other systems; even leading agents still fail roughly one in three attempts on structured benchmarks.

An AI agent doesn't just answer a question — it works through a task in steps: reading, cross-referencing, drafting, checking its own output. A local AI agent does all of that on your own device, with nothing sent to an outside company's servers. For a director handling confidential board papers, that is the distinction that decides whether the tool is usable at all.

"Local AI agent" names both halves of this: the "agent" part is the kind of AI — task-doing, not just question-answering; the "local" part is where it runs. This guide explains what that combination genuinely does today, what remains hype, and why it matters more for board work than for most other uses of AI.

Key takeaways

An AI agent pursues a goal and takes actions using tools, with some autonomy — not just answering a single question, according to Wikipedia's definition of the term.¹
A local AI agent runs that entire process on your own device — no external servers, no account to sign into, nothing sent anywhere to complete the task.
Agent capability is improving fast but still fails often: on the OSWorld benchmark, accuracy rose from around 12% to 66.3% during 2025, yet agents still fail roughly one in three attempts on structured tasks, according to Stanford HAI's 2026 AI Index Report.²
55% of UK employees already use AI tools their employer hasn't approved, and confidential company documents are among the material most often shared with them, according to Okta and Apprize360's 2026 survey — exactly the exposure a local agent's architecture removes.³
Reliability drops sharply on longer, less supervised tasks: frontier agents succeed almost 100% of the time on tasks that take a person under four minutes, but under 10% of the time on tasks taking around four hours, according to the AI research group METR.⁴

What is an AI agent, in plain terms?

Wikipedia defines it directly: "An AI agent or agentic AI is an artificial intelligence program that can pursue goals, use software or other tools, and take actions with some level of autonomy."¹ The distinction from an ordinary chatbot is action. A chatbot answers the question you type and stops there. An agent plans a sequence of steps toward a goal, uses tools along the way, and moves on to the next step without being prompted individually for each one.

In practice, that might mean reading a document, pulling out several figures, checking them against a second document, and drafting a summary of any discrepancy — one instruction, several actions. The UK's National Cyber Security Centre describes the same shift in its own guidance: agentic systems "can access data sources, remember context, make decisions, use tools, and take actions in pursuit of a goal," and "can operate without continuous human intervention," according to the NCSC, 2026.⁵

"Local" answers a separate question: where does that process run? A local AI agent carries out every one of those steps — the reasoning, the tool use, the reading of your documents — on your own computer, rather than sending the task to an outside company's servers.

How is a local AI agent different from one that runs on external servers?

Most well-known AI assistants — the kind built into a search engine or a general-purpose chat app — process your request on the provider's own infrastructure. You send your documents and instructions out; the agent works on them there; the results come back. A local AI agent inverts that: the agent itself, including the underlying model, is installed on your device, and the documents it works with never have to leave.

Factor	Agent on external servers	Local AI agent
Where the reasoning happens	Provider's own servers	Your own device
Documents leave your device	Yes	No
Works with no internet connection	No	Yes
Provider can log or retain your documents	Depends on provider and settings	No
Needs an account with an outside company	Usually	No
Best suited to	General, non-confidential tasks	Confidential board and governance material

For the fuller comparison of how the two approaches handle risk, see Local AI vs Cloud AI.

Why does it matter for confidential professional work?

Because an agent, by definition, does more than read — it acts, and that combination raises the stakes on where it runs. The NCSC's own guidance puts the test plainly: "If you cannot understand, monitor or contain an agent's actions, it is not ready for deployment," NCSC, 2026.⁵ A chatbot that only answers questions is a comparatively contained risk. An agent that reads a live document, remembers what it found, and acts on that information is not — a private AI agent, one where nothing about the task leaves your device, is what closes that gap.

The way professionals are actually using AI has already run ahead of the controls around it. In Okta and Apprize360's 2026 survey of executives and knowledge workers across seven countries, 55% of UK employees said they used AI tools their employer had not approved — despite 96% of UK executives expressing confidence in their organisation's visibility into AI use, the widest such gap of any country surveyed.³ Across the full sample, confidential company documents were among the material most commonly shared with those unapproved tools.³

For a director, the exposure is personal as well as organisational. A board pack analysed by an agent running on an outside company's servers means that pack, and whatever the agent did with it, exists — however briefly — on infrastructure you do not control. A local AI agent removes that step entirely: the reasoning, the document, and the output all stay on the device in front of you.

What can a local AI agent actually do today?

Genuinely useful work, within real limits. On your own documents — a board pack, a set of minutes, a strategy paper — a local agent can search across them, cross-reference a figure in one against a claim in another, and follow a multi-step instruction ("summarise the risk section, then flag anything that contradicts last quarter's numbers") without needing to be walked through each step by hand.

This is the territory meetinginsight.ai works in: it reads and cross-references your board papers on your own device, surfacing risks and sharper questions, without an account or a document ever being sent anywhere. It is a working example of what a document-focused local agent is for — not a general-purpose assistant that browses the open web or acts across unrelated systems, but one built specifically to reason across the material you already have.

That is a meaningful, current capability, not a promise. What remains genuinely early is full autonomy over long, loosely supervised tasks, which the next section covers honestly.

What are the honest limits of AI agents today?

The marketing around "AI agents" has outrun what most of them reliably do, and it is worth saying so plainly. Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls."⁶ Its analyst Anushree Verma put the underlying reason bluntly: "Most agentic AI projects right now are early-stage experiments or proof of concepts that are mostly driven by hype and are often misapplied," Verma, Gartner, 2025.⁶

Reliability, not intelligence, is the sticking point. METR's research on how long a task an AI agent can complete found that success rates stay near 100% on tasks a human expert would finish in under four minutes, but fall to under 10% on tasks taking roughly four hours — with the length of task an agent can reliably handle having doubled every four to seven months over the past two years.⁴ As the Princeton researchers Stephan Rabanser, Sayash Kapoor, Arvind Narayanan and their co-authors put it in their 2026 study of agent reliability: "an agent that succeeds on 90% of tasks but fails unpredictably on the remaining 10% may be a useful assistant yet an unacceptable autonomous system."⁷

None of this is a reason to dismiss agents; it is a reason to be precise about what "agent" means in any given product, and to keep a person reviewing the output — the same discipline you would apply to a very capable but unsupervised junior colleague.

How can you tell if an "AI agent" tool is genuinely local?

The label is used loosely, so it is worth checking directly. Four questions settle it:

Does it complete its task with the internet switched off? A genuinely local agent does. If it stops working offline, some part of the task is happening on an outside server.
Must you create an account before it will act on your documents? A required sign-in usually means the work is routed through an external service.
Does the provider's privacy policy describe retaining or reviewing what the agent reads? With true local processing, there is nothing of yours for them to retain.
Can you get an unambiguous answer to "does anything my documents contain ever leave this device?" A qualified answer — "encrypted before sending," "processed securely" — still means the documents leave.

A tool that passes all four keeps both your documents and the agent's reasoning about them on your own machine.

In summary

An AI agent is AI that works through a task in steps rather than simply answering a question, and a local AI agent does that entire process on your own device, with nothing sent to an outside company's servers. The capability is real and improving quickly, but so are its limits — even the best agents still fail roughly one in three attempts on structured benchmarks, and reliability drops sharply on longer, less supervised tasks. For confidential board work, that combination of genuine capability and honest limitation is exactly why where the agent runs matters as much as what it can do.

If you want the fuller technology picture, read What Is Local AI? and What Is a Local LLM?; for the head-to-head on risk, read Local AI vs Cloud AI. And if you would rather simply use one, meetinginsight.ai works through your board papers entirely on your own device — nothing sent, nothing stored elsewhere. Try a free 30-day trial at meetinginsight.ai/download.

Notes

meetinginsight.ai works through your board papers entirely on your device. Nothing sent. Nothing stored elsewhere. Download a free 30-day trial at meetinginsight.ai/download.

Footnotes

Wikipedia, "AI agent." https://en.wikipedia.org/wiki/AI_agent ↩ ↩²
Stanford Institute for Human-Centered AI (HAI), "The 2026 AI Index Report," Technical Performance chapter, 13 April 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance ↩
Okta and Apprize360, "AI Agents at Work 2026: Securing the Agentic Enterprise" (292 executives and 492 knowledge workers across seven countries, fielded March 2026). https://www.okta.com/newsroom/articles/ai-agents-at-work-2026-agentic-enterprise-security/ — corroborated by The Register, 27 May 2026: https://www.theregister.com/ai-ml/2026/05/27/bosses-blinded-by-confidence-about-shadow-ai-use-by-workers/5247275 ↩ ↩² ↩³
METR, "Measuring AI Ability to Complete Long Software Tasks," 19 March 2025, and METR's live time-horizon tracker, updated May 2026. https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ ; https://metr.org/time-horizons/ ↩ ↩²
National Cyber Security Centre (UK), "Thinking carefully before adopting agentic AI," by Martin R and Dr Kate S, 15 May 2026. https://www.ncsc.gov.uk/blogs/thinking-carefully-before-adopting-agentic-ai ↩ ↩²
Gartner, Inc., press release, 25 June 2025; quote from Anushree Verma, Senior Director Analyst, Gartner, as reported by MarTech. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027 ; https://martech.org/gartner-40-of-agentic-ai-projects-will-fail-making-humans-indispensable/ ↩ ↩²
Stephan Rabanser, Sayash Kapoor, Peter Kirgis, Kangheng Liu, Saiteja Utpala and Arvind Narayanan, "Towards a Science of AI Agent Reliability," ICML 2026, as reported in Fortune, 24 March 2026. https://fortune.com/2026/03/24/ai-agents-are-getting-more-capable-but-reliability-is-lagging-narayanan-kapoor/ ↩