An AI assistant that actually knows your business

Most people try AI on their email once, get a draft that reads like a press release, spend longer fixing it than writing from scratch would have taken, and quietly stop.

That reaction is the right one. The tool was doing its best with nothing to go on.

What separates an AI novelty from a system that gives you back a day a week isn’t the model. It’s what the model was allowed to read before it answered.

Why generic AI drafts are worse than useless

Ask a general chatbot to reply to a customer enquiry and you’ll get something grammatical, polite and completely detached from your business. It doesn’t know your prices, your lead times, what you’ve already told this client, or that you never use the word “solutions”.

So you rewrite it. And rewriting someone else’s near-miss is often slower than starting from a blank page, because first you have to read it, work out what’s wrong, and fix it while preserving the parts that were fine.

Generic drafts are also confidently wrong in ways that are easy to miss. A made-up delivery timeframe reads exactly like a real one. Send it and you’ve committed to it.

That’s why “we tried AI and it wasn’t useful” is such a common verdict. The experiment was real and it failed for a real reason.

What “knows your business” actually means

The fix is undramatic. Give the system your material, and require it to answer from that material rather than from general knowledge.

In practice it reads some mix of:

• Your past correspondence, which shows how you phrase things, how formal you are, how you open and close
• Your reference material: price lists, service descriptions, standard terms, warranty conditions
• Your job or client history, so it knows what this particular customer has bought, been quoted or complained about
• Your policies: what you will and won’t do, payment terms, cancellation rules

Now when an enquiry arrives, the draft isn’t invented. It’s assembled from things that are true about your business, in wording that resembles yours.

The technical name for this is retrieval-augmented generation, and you can forget that phrase immediately. What matters is the behaviour. It looks things up before it answers, and it can show you where an answer came from. That last capability is what makes it trustworthy enough to lean on when you’re busy.

What it looks like in practice

I built one of these for an academic who was drowning in correspondence. A constant flow of similar questions, each needing a considered and individual reply, in a role where tone genuinely matters.

The system reads incoming mail, works out what’s being asked, retrieves the relevant material and past responses, and prepares a draft in their voice. They read it, adjust whatever needs adjusting, and send.

Daily triage went from hours to minutes. Not because the assistant handled everything, because it didn’t, but because it handled the routine majority and left full attention for the part that genuinely needed thought. Consistency improved too, which nobody expected going in. Similar questions started getting similarly good answers regardless of how tired the person was by the time they reached them.

There’s more detail in the email and knowledge system case study.

Where it works, and where it doesn’t

It works well when:

• You answer variations of the same questions over and over
• Good answers already exist somewhere in your records
• Tone and consistency matter
• Slow replies cost you something real, like lost quotes or frustrated clients

It works badly when:

• Every response is genuinely novel and needs judgement no record contains
• Your reference material is contradictory or well out of date, because the assistant will faithfully reproduce the contradictions
• Nobody will review drafts before they go out and a wrong answer is expensive

That middle point deserves more weight than people give it. Grounding an assistant in messy documentation doesn’t tidy it up, it spreads it faster. Occasionally the most valuable thing to come out of one of these projects is discovering that your price list exists in three versions and nobody’s sure which one is current.

What it takes to set up

Less than people expect, and most of the effort sits in preparation rather than building.

Gather the material. Usually a few hundred past replies plus whatever reference documents you rely on. People underestimate this step, not because it’s hard, but because it’s often the first time anyone has had to decide which version of a document is the authoritative one.

Decide the boundaries. What should it never do? Common answers: never quote a price without review, never commit to a date, never respond to complaints. Those are business rules and they need stating rather than assuming.

Decide how much autonomy it gets. Most people start with draft-only and review before sending. Some later let routine categories go out automatically once the thing has proved itself. Starting cautious costs almost nothing and makes the first mistake survivable.

Sort out access and privacy. What it can read, where data is stored, whether anything is retained for training. Every one of those has a defensible answer, but only if somebody actually chooses it.

The honest limitation

An assistant grounded in your business is far better than a generic one. It still isn’t a replacement for you.

It will occasionally retrieve the wrong precedent, or miss that this client is a special case for reasons recorded nowhere. Human review exists for a reason, and the systems that hold up over time are the ones that plan for it rather than designing it out.

A better way to think about it: the assistant does the first eighty per cent of the answering so your attention goes to the part that needs you. That’s a smaller promise than “it answers my email”, and it’s one these systems keep reliably.

If your inbox is what’s eating your week, it’s worth a conversation. The first question is simply whether good answers already exist somewhere in your records, and they usually do.