AI Agents

Yes, easily, from a very short sample — which is precisely the problem. Voice cloning is cheap, fast and convincing enough for audiobooks, dubbing and accessibility, and equally convincing to a relative receiving a scam call.

Yes, and it is genuinely good. Phones, earbuds and some glasses translate spoken conversation with a second or two of lag across major language pairs. It is excellent for travel and everyday chat, and still risky for legal, medical or contractual conversations.

Yes, mostly. Consumer phones and cheap call-screening services will pick up, transcribe and hold a short conversation with a caller, and they handle spam and simple appointment calls well. Anything nuanced — an emotional caller, a complex booking, a call that needs your judgement — still gets handed back to you.

Close. Assistants triage, summarise, draft in your voice and handle scheduling well enough that many people no longer read every message. Sending important replies unsupervised is still a bad idea, and quality drops on unfamiliar threads.

It can write most of it, but shipping is still a human responsibility. Coding agents build features, fix bugs and open pull requests unaided, and teams genuinely rely on them. Architecture decisions, subtle security issues and long-lived maintenance still need engineers in the loop.

Yes for practice, explanation and patient repetition — with adult oversight. AI tutors adapt to a pupil's level and never tire. They are weaker at motivation, assessment integrity and knowing when a child is quietly disengaging.

It can remember a lot, not everything. Mainstream assistants now keep persistent memory of your preferences, projects and past conversations, and it noticeably improves them. Recall is still selective, occasionally wrong, and confined to whichever assistant you told.

For small, conventional apps, genuinely yes — a non-developer can now describe a tool and get something deployed that works. The gap appears later: security review, awkward edge cases, data modelling and the unglamorous maintenance that decides whether software survives its first hundred users.

For short clips, adverts and b-roll, yes — and it is already in commercial use. Sustained narrative video with consistent characters, dialogue and continuity across minutes remains difficult and needs heavy human editing.

It can run the routine layer. Scheduling, invoicing chase-ups, customer replies, bookkeeping categorisation and marketing drafts are all realistically automatable now. Judgement calls, exceptions and anything with legal weight still need an owner.

It can react intelligently; genuine anticipation is patchy. Modern smart homes handle presence, light, heat and routines well, and LLM-driven assistants finally understand loose natural instructions. Predicting what you want before you ask still produces as many annoyances as wins.

It plans brilliantly and books adequately. Agents will research destinations, build itineraries and complete simple bookings. Multi-leg trips with dependencies, loyalty programmes and awkward change policies still go wrong in expensive ways.

You can send a note-taker; you cannot credibly send a stand-in. AI meeting assistants join, transcribe, summarise and assign actions extremely well. Photorealistic avatars that speak for you exist, but participating convincingly in live discussion is beyond them — and socially fraught.

Partly. An assistant can plan meals, build a basket and reorder your usual items, and some retailers offer their own AI basket builders. Fully hands-off shopping — choosing substitutions the way you would, spotting a bad price, adapting to what is actually in stock — still needs you to check the basket before checkout.

It works often enough to be worth it, but not predictably. Services and agents will call or chat with providers, ask for retention offers and cancel subscriptions. Outcomes depend heavily on the provider, and some actively block automated agents.

You can buy one, but expect a companion rather than an assistant. Pendants, pins and earbuds will listen on demand, transcribe your day and answer questions well enough to be useful. Continuous all-day capture that reliably surfaces the right thing at the right moment, on a battery that lasts, is still beyond them.

AI supports diagnosis; it does not own it. Cleared tools read scans, flag findings and triage in clinical settings, and general models are startlingly good at plausible differentials. Regulation, liability and error modes mean a clinician signs off, and a chatbot is not a doctor.

It can fill the basket; you should still press the button. Agents research products, compare prices and complete simple reorders competently. Letting one spend freely without confirmation invites the failure mode nobody wants: a confidently wrong purchase, at speed, with your card.

No, not reliably. Detectors work reasonably on known generators and degrade badly on new ones, on compressed social media re-uploads and on partially edited media. Content provenance signing is the more promising route, and adoption is still partial.

Barely, and this is becoming the important question. As agents log into services and spend money, sites need to know which agent is acting, for whom, and within what mandate. The standards exist in draft; the deployed reality is mostly API keys and hope.