AI and automation

How to brief an AI agent that sounds like your business

Most AI experiments disappoint because the output sounds generic, invents facts or can't answer specific questions. All three are briefing problems, and all three can be fixed.

Most businesses that try AI have the same three complaints. The writing sounds like everyone else's. It makes things up. And it can't answer the specific questions customers actually ask.

It's tempting to blame the technology, but in our experience all three are briefing problems. An AI agent is a bit like a capable new starter on their first morning: bright, fast and keen, with no idea how you do things. Give it a vague brief and you get vague work.

Here's how we brief agents so they sound like the business and know what they're talking about.

Start with one job, not "AI for everything"

The most successful projects start small. Pick one task that is:

  • repetitive, done every day or every week
  • text-heavy, mostly reading and writing
  • checkable, so a person can quickly tell whether the output is right

Good first candidates include first drafts of proposals, triaging a shared inbox, turning meeting notes into actions, or answering common questions from your own policies. Write down how long the task takes today. That number is your baseline, and it keeps everyone honest about whether the agent is helping.

Capture your voice from real writing

"Professional but friendly" describes almost every business in Britain. It isn't a brief. Your real voice lives in the writing you're proudest of, so start there.

Gather 10 to 20 examples: emails that landed well, proposals that won work, posts people shared. Then describe what they have in common, using pairs that show where the line sits:

  • warm, not gushing
  • confident, not salesy
  • plain English, with jargon only if the client uses it first
  • British spelling, always

Add a short list of words and phrases you never use, and the ones you always do. Note the practical details too: how you sign off, whether you use first names, how long a typical email runs.

This is exactly what our LinkedIn content engine, PillarFlow, does before it writes anything, building a brand voice profile across twelve dimensions. The same principle works for any agent.

Ground it in your facts

An agent that guesses will eventually guess wrong in front of a customer. The fix is to give it your own documents to work from, an approach often called retrieval-augmented generation, or RAG. When someone asks a question, the agent looks up the relevant passages first and answers from those.

A good knowledge base is:

  • curated: your current services, prices, policies, FAQs and case studies, not every file on the shared drive
  • current: old price lists and superseded policies removed, with someone responsible for keeping it up to date
  • honest about gaps: the agent is told to say "I don't know, let me pass you to the team" rather than fill a gap with something plausible

Set guardrails before you need them

Decide in advance what the agent must never do. Typical rules include:

  • never promise discounts, refunds or delivery dates
  • never give legal, medical or financial advice
  • never share one customer's details with another
  • hand over to a person when someone is upset, vulnerable or making a complaint

Then decide where a person stays in the loop. For most businesses, the right starting point is that the agent drafts and a person approves. You can loosen that later, once you trust the output.

Connect it to where the work happens

An agent that lives in a separate tab gets forgotten. Connect it to the tools your team already uses: email, calendar, CRM and documents. Start with read-only access where you can, and add the ability to send or update things once the drafts are consistently good.

Test it like a sceptic

Before launch, write 20 real questions and tasks, including awkward ones: the customer who's cross, the request that's out of scope, the question your knowledge base doesn't cover. Score each answer for accuracy, tone and usefulness. Fix the brief or the documents, then test again.

It also helps to ask a colleague who wasn't involved to try to break it. They'll find the gaps faster than you will.

Mind the data

If an agent handles personal data, UK GDPR applies just as it does to any other system. Know where the data is processed, check the terms of the AI platform you use, and don't paste customer details into consumer chat tools. The Information Commissioner's Office publishes guidance on AI and data protection that is worth reading before you start.

Keep training it

Voice and knowledge drift. Prices change, services evolve and new questions come in. Set aside half an hour a month to review a sample of the agent's work, update the knowledge base and refine the brief. Small, regular corrections beat a big rebuild every year.

A one-page briefing template

The job: what the agent does, for whom, and how long it takes today.

The voice: five do and don't pairs, plus the words you always and never use.

The facts: the documents it may use, and who keeps them current.

The limits: what it must never do, and when it hands over to a person.

The checks: who approves its work, and how you'll measure whether it's helping.

Fill that in and you're most of the way to an agent people actually want to use. If you'd like help, our AI Agents & Automation service builds brand-safe agents trained on your documents and your voice, and the scoping call is free.

Written by Liberty Web

Liberty Web Ltd is a British web agency building technology for good: websites, AI agents and automation, ecommerce and website support.

Ready to Put AI to Work?

Book a free scoping call and we'll identify the agents that would save your business the most time.