Agentic AI in 2026: what to hand over to an AI agent — and what to keep

Quick answer: An AI agent is a step up from a chat assistant: instead of only answering, it carries out multi-step tasks using tools — browsing websites, reading and writing files, updating records — while you supervise. The major platforms now ship agent features with human-approval gates, and the vendors themselves state that the biggest open security problem, hidden instructions inside web content (known as prompt injection), is not fully solved. So the practical question is not whether agents can work, but which of your tasks are safe to delegate. Hand over work that is reversible, has a small blast radius, is quick for you to check, and does not require your credentials. Keep payments, contracts, public posts and anything regulated in human hands. The rest of this article turns that into a four-part test you can apply this week.


What "agentic AI" actually means in late 2026

A chat assistant answers. An agent does: you give it a goal, and it plans steps, uses tools (a browser, a terminal, your files, connected apps), and returns finished work. Three things distinguish the 2026 generation from the demos of 2024:

  • Tool access is real. The agent operates software, not just text — opening pages, running commands, editing documents.
  • Tasks are multi-step. You describe an outcome ("collect the last 20 supplier invoices into a spreadsheet"), not a single prompt.
  • Control features are part of the product. Vendors now document permission settings, approval gates and activity logs, because an agent that can act needs supervision in a way a chat window does not.

Between July 2025 (when OpenAI launched ChatGPT agent mode) and November 2026, this went from one product to a category. The table below summarises what each major vendor documents as of 19 November 2026 — vendor-reported capabilities, checked against official pages. We have not independently benchmarked any of them, and no vendor mentioned here has any relationship with CortexLeap.

What the vendors themselves say about control

Vendor Agent surface (status Nov 2026) Control features they document
OpenAI ChatGPT agent mode (launched July 2025; that launch page is now flagged as outdated, pointing to the newer products below) Requests permission before "actions of consequence"; you can interrupt or take over the browser mid-task
OpenAI Workspace agents (launched April 2026 as a research preview; OpenAI's page now lists it as generally available on Business, Enterprise and Edu plans) Admin-defined permissions; approval gates requiring human sign-off before sending messages or updating records; audit logs
Anthropic Computer use on the Claude platform (generally available since August 2026) Your application executes every action in "an environment you control"; browser use, Skills and Files APIs alongside it
Anthropic Cowork — a desktop agent for non-developers (research preview January 2026, macOS; later made available on paid Claude plans, per launch coverage) Permission-based access to chosen local folders; the agent works in your files rather than chat attachments
Google Gemini agent (Google Cloud, announced 8 October 2026) Sub-agents for multi-step jobs; an optional "coworker agent" with its own Workspace account, governed inside your Workspace/Cloud tenancy
Google Agentic bookings in Search AI Mode (UK, from April 2026) The flow takes you from planning to booking in a few steps; booking completes with the provider

Availability varies by plan and country, and research previews change quickly — treat each vendor's own page as the current truth, not this table.

The delegation test: four questions before you hand anything over

The control features above are worth having, but they are the vendor's half of the arrangement. Your half is deciding what to delegate. Before giving any task to an agent, answer four questions in writing.

1. Reversibility. If the agent does the wrong thing, can it be undone with a phone call or a keyboard shortcut — or not at all? Collecting prices into a draft sheet is reversible. A payment to a supplier is (at best) awkwardly reversible. Untangle this before the first run, not after the first mistake.

2. Blast radius. What is the worst realistic outcome if the agent is confidently wrong? A shortlist that misses one good supplier costs you an hour. An email sent to the wrong client list costs you trust. Rank tasks by the size of the hole a bad day would leave.

3. Checkability. Can you verify the output quickly, with your own eyes and your own knowledge? A comparison table of published prices is easy to spot-check. A "summary of everything discussed across our last ten client calls" is not, because the errors hide until they surface later. Delegate what you can check; keep what you cannot.

4. Credentials and identity. Which account does the agent act through, and whose name is on the actions? Modern platforms connect through scoped permissions rather than shared passwords — keep it that way. Never type your master password into an agent's browser view, and prefer a dedicated or restricted work account over your primary one.

If a task passes all four — reversible, small blast radius, easy to check, no sensitive credentials — it is a candidate. If it fails any one of them, redesign the task (human approval at the risky step) or keep it.

Green, amber and red: sorting your own task list

Applied honestly, most small-business tasks fall into three bands.

Green — delegate the whole task. Research and shortlisting (supplier options, published prices, funding-scheme overviews), drafting documents that you will review before anything leaves the building, organising files in a folder you have explicitly shared, meeting-prep briefs, extracting structured data from a pile of documents into a draft spreadsheet.

Amber — delegate with an approval gate. Sending routine emails (even internal ones), updating records in a CRM, creating calendar invitations. Let the agent prepare; you approve the send. The Workspace agents control described above exists precisely for this band, and other platforms offer equivalents — use them rather than raw autonomy.

Red — keep the decision, delegate only the preparation. Payments and bank transfers, signing anything, public posts under your brand, responses to formal complaints, tax or compliance filings, and anything that sends client personal data somewhere new. An agent can draft the payment instruction or the filing; a named human presses the button, because accountability cannot be delegated even when the typing can.

As a fictional example: imagine a six-person interiors studio, "Studio Fern" (invented for this article). Its agent shortlists 20 upholstery suppliers into a comparison sheet — green. It drafts polite chasers for three overdue invoices, with the owner approving each send — amber. It pays the chosen supplier's deposit — red, and it stays that way regardless of how well the first two bands went.

The failure mode the vendors name themselves

There is one risk worth reading about before your first delegation, because the vendors themselves call it out: prompt injection. Web pages, documents and emails can contain hidden instructions — visible to a model, invisible to you — and OWASP lists prompt injection as the number-one risk in its Top 10 for LLM applications. OpenAI, writing about its own browser agent, describes prompt injection as "one of the most significant risks we actively defend against", while being candid that defences are continuously evolving rather than finished.

The practical consequence for a small business is simple and does not require you to understand the attack: an agent that is reading untrusted web content should not, at the same time, hold broad write access to your important systems. Keep approval gates switched on for anything that sends, pays or deletes. Give agents the narrowest access that gets the job done. This is the same least-privilege discipline you already apply to staff accounts — agents are new colleagues with excellent typing speed and no sense of self-preservation.

Plausible-but-wrong output deserves its own mention: agents summarise, and summaries inherit model error. Spot-check figures against the original source on anything a client or an authority will see. Our earlier piece on how to test an AI tool before you depend on it gives you the method; the same discipline applies here, just with more moving parts.

A first week that stays safe

A seven-day sequence that produces evidence rather than anecdotes:

  • Day 1: Pick one green task from your own list. Write down what "done and correct" looks like.
  • Day 2: Run it on anonymised copies, read-only if the tool allows. No real client data yet.
  • Day 3: Compare the agent's output with how you did the same task last month. Count the time you spent checking, not just the time saved.
  • Day 4: Check plan availability honestly (see the FAQ below) and set a spending ceiling before any trial converts to paid.
  • Day 5: Turn on the vendor's approval gates and connect the minimum set of tools. If your business has written rules about AI use, this is where they earn their keep.
  • Day 6: Run the task under supervision with gates on. Sample-check the outputs.
  • Day 7: Make a written go/no-go decision — keep, adjust, or abandon — and note the rules your business will follow for the next task.

If the whole week proves only that the task was faster by hand, that is a useful result too. Not every task suits an agent, and the costs of finding out on a small task are trivial compared with finding out on a payment run.

Where this leaves a small business

Agent features are becoming standard across the tools you may already pay for, so the sensible posture is neither enthusiasm nor refusal — it is deliberate, narrow delegation with human checkpoints, expanded task by task as evidence accumulates.

If you want a grounded view of which of your own processes are genuinely ready for delegation — and which should stay manual — the Business Process & AI Diagnostic looks at up to three agreed processes and returns written findings with prioritised improvements, for a one-off fee shown on that page. There is no obligation to automate anything afterwards. Prefer to start smaller? The Personal AI Session is a one-to-one hour on one task you prepare, where you practise checking AI output under guidance. And if you would rather talk fit before either, the free 30-minute conversation is optional and commits you to nothing.

FAQ

Do I need a business or premium plan to try agent features? It depends on the vendor. ChatGPT agent mode launched on Pro, Plus and Team plans in July 2025; OpenAI's Workspace agents launched in April 2026 as a research preview and are now generally available on Business, Enterprise and Edu plans; Google's Gemini agent is a Google Cloud/Workspace product; Anthropic's Cowork began as a macOS research preview on a high tier and, per launch coverage, later reached paid plans. Availability changes often — check the vendor's own page for your tier, and see our piece on free versus paid AI plans before upgrading for one feature.

Are AI agents safe for business tasks? For reversible, low-blast-radius, checkable work with approval gates on — reasonable to try, with supervision. For payments, contracts, public communications and regulated filings — not without a named human deciding and pressing the button. The honest caveat: the security problem of hidden instructions in web content is acknowledged by the vendors themselves as significant and unsolved.

Will an AI agent replace needing a person at all? Agent features automate bounded, describable tasks. They do not carry judgement about your clients, accountability for your commitments, or knowledge of the unwritten rules of your trade — and the tasks that depend on those are exactly the ones the four-question test keeps with you. The realistic outcome for most small businesses is a small set of delegated tasks, not a role that disappears.

Last updated: 19 November 2026

Book the free 30-minute call