What can an AI assistant actually do?

Capability lists usually answer the wrong question. Whether an assistant 'can write emails' matters less than whether you can trust the result without checking it — because checking everything erases the time saved. So this inventory is organised by trust level, based on operating these systems daily inside PFE's own businesses, not on vendor claims.

Work where errors are rare, obvious, or low-cost. This is where assistants pay for themselves immediately:

• Reminders and follow-ups — holding deadlines, renewals, and 'chase this if no reply by Friday' with perfect patience.
• Summarising — long email threads, documents, meeting transcripts, reports. Errors happen but are quickly visible.
• Answering from verified documents — when grounded in facts you supplied, with citations you can check.
• Triage and routing — sorting incoming messages by topic, urgency, or required handler.
• Data extraction — pulling structured fields out of invoices, CVs, enquiry emails; spot-check a sample, not every row.
• Meeting scheduling — finding slots, sending invites, handling reschedules against a live calendar.

Work where the assistant does 80% and your review is the quality gate:

• Email and message drafting — excellent first drafts in your tone; you catch the occasional wrong assumption before it ships.
• Research and comparisons — systematic sweeps with sources; verify the two or three facts your decision actually turns on.
• Quotes and proposals — assembled from your pricing rules; a human approves anything customer-binding.
• Reports and documentation — structure and first prose from raw notes or data.
• Translations and localisation — strong drafts; native review where the stakes are real.

Work assistants can do only inside carefully built guardrails — this is where 'AI employee' deployments live, and where most disappointments happen:

• Autonomous multi-step actions in business systems — updating CRMs, issuing refunds, changing bookings. Needs permissions, audit, rollback.
• Continuous monitoring with judgment — watching inventory, prices, or infrastructure and deciding what deserves escalation.
• Customer-facing conversations without review — safe only with strict grounding and honest escape hatches to a human.

Not a capability gap — a category error: final judgment on legal, medical, and financial questions; anything where being confidently wrong is expensive and hard to detect; negotiations and relationships where the other side deserves a human; and decisions you'd be uncomfortable explaining with 'the AI decided'. We keep a fuller list in the companion piece on what you should never delegate.

The pattern across all four tiers: assistants multiply the value of clear rules and verified facts. The businesses that get the most from them are the ones willing to write down how their work actually happens.