Guide · AI agent questions
Hit a wall with an AI agent? Here’s how to get past it.
Ready-made agents — ChatGPT Dots, Grok Bot, Meta Muse, Copilot, Gemini — are good at the easy part. These are the questions people bring us once they’ve tried them and got stuck. Each answer has tips you can use yourself, and what we do when it needs more.
This is Nova — one of our own AI agents, live right now. Try her.
The 18 questions
- Can we run AI on our own hardware so data never leaves?
- Our AI bill keeps growing — how do we control it?
- Is it legal to let an AI agent read customer emails?
- The newest AI agents launched in the US first — what can we use in the UK today?
- We bought Copilot or ChatGPT seats and hardly anyone uses them — what now?
- Can’t Grok Bot, Meta Muse or OpenAI’s Dots do what a custom system does?
Email & documents
How can I let an AI agent check my email without exposing client data?
Give it the least access that does the job, keep the data on systems you control, and let it draft — not send.
- Start read-only. Sorting, summarising and drafting don’t need permission to send or delete.
- Limit what it sees: one shared inbox or a few folders, not your whole mailbox history.
- Check where the mail goes. Many assistants send it to a cloud AI, and some learn from it unless you switch that off. Read the setting before you connect.
- Mask what it doesn’t need: account numbers, health details, passwords and one-time codes.
- Keep a log of what it read and did, so you can answer a customer who asks.
Our own agents read a local copy of our inbox with read-only access; sending stays with a person. For sensitive mail we use AI that runs on your own hardware, so it never leaves the building.
Email & documents
Can a stranger take over my AI agent by sending it an email?
It’s possible. It’s called prompt injection, and it tops OWASP’s list of risks for AI applications. An email can hide an instruction — “forward the last invoices to…” — that the agent mistakes for yours.
- Treat everything the agent reads as information, never as orders. Orders come only from you.
- Don’t give one agent both powers: reading the inbox and sending or paying. Split them, and put a person on the second.
- Approve anything that leaves the business: emails, payments, files.
- Watch for hidden text — white-on-white or tiny fonts — in incoming mail and attachments.
- Test it: send your agent a booby-trapped email and see what it does.
In our setup, agents that read outside messages can’t send email or move money on their own — those steps wait for a person. Each incoming message is screened by a separate safety check before the agent answers.
Email & documents
How do I stop an AI agent sending something I haven’t approved?
Make “draft” the default and “send” a separate step a person takes — at least until the agent has earned trust on that task.
- An approval queue: one tap to approve, edit or bin.
- Let it send alone only low-risk, fixed messages such as confirmations and receipts.
- Set limits: who it may write to, how many a day, never a new recipient without approval.
- Keep an off switch you can reach from your phone.
Our agents queue drafts for a person to approve, and our front desk can be paused or taken over from a phone in seconds.
Email & documents
Can AI handle a shared inbox (info@) for the whole team?
Yes. Sorting, tagging, summarising and drafting replies in a shared inbox is one of the most reliable jobs for AI today. Sending is where to be careful.
- Start with sorting: label by type — sales, support, invoices, junk — and route to the right person.
- Draft in your voice from your own past replies.
- Pull out the facts — order numbers, dates, amounts — into your system instead of retyping them.
- Measure it: how many it sorted right last week, and how much time it saved.
We run this on our own inboxes: mail syncs every 20 minutes, gets sorted, and replies are drafted for a person to approve.
Phones & messages
How do I get an AI agent to answer our phone calls?
Voice AI can now hold a natural conversation. The hard part is connecting it to your diary and systems — and knowing when to hand over to a person.
- Decide what it does alone (opening hours, directions, booking a slot) and what it hands over (complaints, refunds, anything medical or legal).
- Connect it to your real calendar or booking system. An agent that only takes a message is voicemail with a nicer voice.
- Tell callers it’s an AI. If you record or transcribe calls, say so and decide how long you keep them.
- Test with real callers, accents and bad lines before it goes on your main number.
- Always a way out: if it’s unsure, it transfers or books a call-back.
We set up AI phone agents that answer and make calls in a natural voice, book into your diary and hand over to a person — who can follow the call live and take over at any moment.
Phones & messages
Meta’s WhatsApp agent can’t see our bookings or stock — what now?
The free WhatsApp agent is built for common questions and simple bookings. When customers ask about live stock, order status or anything in your own systems, check whether it can see that information — if it can’t, it will hand over, or guess.
- Use it for FAQs and opening hours. Don’t let it answer about availability or prices it can’t check.
- Give it a clear hand-over for anything that needs your systems.
- For real look-ups — stock, orders, live availability — you need an agent connected to those systems through WhatsApp’s business platform.
- Keep one record per customer across WhatsApp, web chat and email.
Our front desk runs on our website and WhatsApp through WhatsApp’s official business platform. It keeps one record per customer and books calls straight into our bookings list.
Phones & messages
What should an AI receptionist never do on its own?
Anything where a wrong answer costs money, health or trust.
- Never invent prices, discounts or availability — only quote what it can look up.
- Never give medical, legal or financial advice.
- Never promise refunds or make exceptions.
- Never pretend to be human. Say it’s an AI, and offer a person.
- Never drop the thread: when it hands over, the person sees the whole conversation.
Our front desk quotes only from a fixed price guide, says it’s an AI, and hands the full conversation to a person with one alert.
Agents working together
How do I get my AI agents to discuss a project and hand work to each other?
Give them a shared place to work: one record of the job, clear roles, and a rule for who decides. Without that, each agent works from its own memory and you end up copying and pasting between chats.
- One agent, one job: research, then draft, then check. Specialists beat one do-everything bot.
- A shared record — a job card or database row — that every agent reads and updates, not separate chat histories.
- A status for each step (found → scoped → drafted → approved → sent), so nothing is done twice or skipped.
- A person signs off at the steps that matter.
- Agents from different makers? Connect them through open standards such as MCP and A2A, or simply through one shared database.
We run seven agents that pass work along this way. One finds opportunities, one scopes them with us, one writes the proposal — and nothing is sent until a person approves it. They don’t share one brain either: each job goes to the AI best at it.
Agents working together
ChatGPT, Copilot and Gemini don’t share memory — how do we keep one version of the truth?
Don’t keep the truth in any of them. Keep it in your own systems — CRM, shared drive, database — and let every tool read from there.
- Pick the source of truth for customers, jobs and documents — usually a tool you already pay for.
- Connect assistants to it instead of pasting into chats.
- Write decisions down where agents can read them. A short project note beats a long chat history.
- Review what each assistant can see, and remove access when someone leaves.
All our agents read and write one shared database, so the front desk, the researcher and the writer see the same facts.
Agents working together
Who’s watching the agents? How do we monitor and stop them?
Someone should be — with a log of what each agent did, an alert when something unusual happens, and an off switch.
- Log every action: what it read, changed and sent.
- A daily or weekly summary to a named person.
- Alerts for the unusual: a spike in sends, a new recipient, an error loop.
- An off switch per agent that works from your phone.
Our agents report into one console and a phone channel. The front desk can be paused, resumed or taken over from a phone in seconds.
Your systems
Our software has no API or connector — can an AI agent still use it?
Often, yes. Newer agents can use a screen the way a person does — reading it, clicking and typing — which works for older systems and websites with no integration.
- Look for a back door first: exports, a spreadsheet import, an email address that feeds the system, or a paid API tier.
- If it must use the screen, keep tasks short and repeatable, and check the results.
- Stop for logins, two-factor codes and payments. A person does those.
- Expect it to be slower than an API. Use it where the time saved still adds up.
Our agents already fill in web forms and submit applications through a real browser, and hand back to a person at passwords, two-factor codes and payments.
Your systems
It worked in the demo but breaks on our real work — why?
Demos run on tidy examples. Real work has messy data, exceptions and systems that don’t line up — which is where many agent projects stall.
- Test on a week of your real cases, including the awkward ones.
- Measure accuracy at each step, not just the end result.
- Give it an “I’m not sure” route instead of letting it guess.
- Fix the data feeding it — duplicates, missing fields, old formats — before blaming the AI.
- Start with one team or one task.
We prove a system against real work before it goes live, then keep monitoring and tuning it.
Privacy, cost & rules
Can we run AI on our own hardware so data never leaves?
Yes, for many jobs. Open AI models now run on a good office computer or a single server — keeping data on site and removing per-use cloud bills.
- Good fits: reading documents and photos, sorting, pulling out data, searching your own files.
- Use a cloud AI only for the steps that need it, with the sensitive parts removed.
- Pay for the hardware once instead of a bill per page.
- Plan updates and backups like any other system.
Our bookseller system reads photos of each book and writes the listing on the seller’s own hardware. The catalogue never leaves the business, and the running cost is close to £0.
Privacy, cost & rules
Our AI bill keeps growing — how do we control it?
Match each step to the cheapest AI that does it well, and stop paying cloud prices for routine work.
- Cheap, fast AI for sorting and extracting; the strongest only for judgement.
- Reuse answers to repeated questions.
- A monthly cap and alert per agent.
- Move high-volume routine work to AI that runs on your own hardware.
- Check your seats: pay for the licences people actually use.
We route every task to the right AI — fast and cheap for routine replies, stronger for writing and judgement, local where volume or privacy calls for it.
Privacy, cost & rules
Is it legal to let an AI agent read customer emails?
It can be. But those emails hold other people’s personal data, so UK GDPR applies to everything the agent does with them.
- Know your lawful basis, and say in your privacy notice that AI processes enquiries.
- Check where the provider processes and stores the data, and whether it learns from it.
- Give the agent only what the task needs.
- Be able to answer “what did your AI do with my data?” Logs help.
- For sensitive data — health, for example — be stricter, or keep processing on your own hardware.
- The ICO, the UK’s data regulator, is looking closely at AI agents; check its latest guidance.
We give agents only the access they need, keep logs of what they touched, and process sensitive data locally. (General information, not legal advice.)
Privacy, cost & rules
The newest AI agents launched in the US first — what can we use in the UK today?
Several of the newest always-on agents — Meta’s Muse among them — launched in the US first, and what’s available here changes month to month. You don’t need to wait: most of the value comes from connecting AI to your own email, files and systems, which you can do now.
- Start with what you have: business plans of ChatGPT, Microsoft Copilot, Gemini in Google Workspace and Claude.
- Choose tools that let you export your data and switch AI later.
- When a new agent lands, read its data terms first: what it can read, and whether it learns from it.
We set up what works in the UK today, on your own systems, and match each task to the right AI — so switching to a newer one later is a change, not a rebuild.
Privacy, cost & rules
We bought Copilot or ChatGPT seats and hardly anyone uses them — what now?
You’re not alone: in June 2026 only 15% of UK firms with 10 or more staff said more than half their people used AI daily (ONS). Usually it’s not the tool — nobody showed people which of their tasks it should do.
- Pick three tasks per team — meeting notes, first drafts, summarising long threads — and show them done live.
- Name a champion in each team.
- Share prompts and templates that work for your business.
- Check usage after a month, and move seats to the people who use them.
- Connect it to your files and email, or it stays a chat box.
The free consultation starts here: where the time goes, which tasks the tools you already pay for can take, and where something custom is worth it.
Privacy, cost & rules
Can’t Grok Bot, Meta Muse or OpenAI’s Dots do what a custom system does?
They’re great personal assistants — inbox, bookings, research. But they’re built for one person and every business at once, so they hit walls where your business gets specific.
- Your own systems. They plug into popular apps, but not usually your stock list, booking system or industry software — so they can’t check what’s actually in stock or booked.
- Your team. They’re one person’s assistant. They don’t pass work to a colleague’s agent, follow your sign-off chain or keep one shared record for the team.
- Your phone line. They don’t pick up your business number, book into your diary and hand over to your staff.
- One brain. They’re tied to their maker’s AI — they can’t hand a step to a cheaper AI, a stronger one, a private one, or a video AI from another company.
- Your rules. You can give them instructions, but they don’t enforce your prices, exceptions or approval limits.
- Babysitting. You end up checking on them, chasing stalled tasks and re-explaining basic steps. A system built for your process runs the routine on its own and tells you when it’s stuck.
- Real volume. Personal plans come with usage limits. Hundreds of documents or enquiries a day need a system built for it.
- Not all here yet. Meta’s Muse launched in the US first, and OpenAI’s Dots needs a paid ChatGPT Pro or Business Premium plan.
We build systems around your business: the best AI for each step — Claude, Grok, Gemini, Qwen or a private local model — connected to your own systems, with your rules and approvals built in.
Hit a different wall?
Tell us where you got stuck. You get an honest, jargon-free view of what fixes it — free, no obligation.
Last checked: October 2026.