There is somebody in your firm who has been using ChatGPT or Copilot every day since about February. Emails, phrasing something awkward, the first draft of anything. They are quietly quite good at it.
They have also never once been able to ask it anything about a client. The assistant is not short of ability. It simply cannot see the file. Everything that matters at a year end sits inside the accounts system and the assistant sits outside it, so the gap gets closed by a person copying things across, which is slower than doing the job by hand.
That gap is what has closed. Here are five jobs that now run from the assistant your firm already pays for.
What it looks like: five jobs, done from the tool you already have
Nobody types tool names. You ask in plain English and the assistant works out what it needs. Every prompt below runs on what is already live.
Round up the client queries and send them
A client has eleven open queries against them, raised over five weeks by three people and sitting in three places. Somebody has to gather them, work out which have already been answered, write one clear email, then record what was asked.
Ask for the outstanding queries on that client to be compiled into a letter. Each is reconciled against its own evidence first, so what goes out is current rather than half a list of things the client sent last week. You get a letter by default, a table if that is all you wanted, and it is marked as sent only once you confirm.
This does not stop the firm chasing clients. What it stops is the job being put down and picked back up every time somebody notices another gap.
Find and file
A client replies to your queries with fourteen attachments in one go. Someone opens each, works out what it is and puts it where it belongs. It is nobody's job in particular and it takes an afternoon.
Instead, you can ask for the documents the firm already holds to be filed automatically by Active. The inbox is swept and things put on the right job. Done and dusted quickly and efficiently.
Filing then happens as documents arrive rather than when somebody has a clear hour, and the trail sits with the client queries and review points it already holds. This one writes into the file, so it runs only where the firm has turned write access on.
Find things without clicking
Somebody needs to know who sits inside a client group, which entity has a job open this year, or where the signed engagement letter went. None of it is difficult. All of it is four screens deep.
Ask the question instead. Who is in the Hendersons' group, and which of those entities has a job open this year. Pull up the signed engagement letter for Marsden Holdings.
Ten minutes of navigating becomes one question. Get the answers you need without searching around for them.
Brief me on the year end before the call
A partner is about to speak to a client and wants to know where the year end has got to: what is done, what is outstanding, what the firm is waiting on the client for. Assembling that properly is most of somebody's afternoon.
Ask for the brief. It is built from the binder, its workpaper records and the query history rather than from anybody's recollection. Get as detailed a pre-meeting brief as you want; driven automatically by the work you’ve already done.
What the partner walks in with is all the right questions to ask, and all the context they need to get the most from the client meeting.
Raise a query or an internal note
You realise something about a client your team manages; it’s a simple query.
You could go into the file and raise the internal note as usual. Or you can just ask your AI assistant to raise the query against the client and assign it to whoever should own it, with the evidence attached.
What an MCP server actually is, in two paragraphs
An MCP server is a connector that lets an AI assistant safely work with another system. It's built to the Model Context Protocol (MCP), an open standard for this kind of integration. The assistant asks the system for something, the system checks whether that is allowed, and the answer comes back in a form the assistant can use.
Active runs one, and it is live rather than promised. Forty-seven tools today, and more with each release, cover clients, documents, and accounts and workpapers, each a single thing an assistant can do: find a client, open a binder, file a document, read a trial balance. It sits alongside the integrations Active already maintains with Xero, QuickBooks and the rest, and works on the same principle: a connection Active keeps up so the firm does not have to. Nothing is installed inside Active and nobody learns a new interface, because the interface is the assistant they already have open.
Why it is open rather than an Active chatbot
A software company in this market can build its own assistant easily enough. Active has not, because an assistant is a place you have to go and firms already have one they are used to. Adding a second is not a favour to anybody.
The Model Context Protocol is an open specification, belonging to neither Active nor any single AI company. The Active plugin for Claude is deployed and working today, Microsoft 365 Copilot connects through Copilot Studio or the Admin Centre, and across Active's firms Copilot is about as common as Claude. A firm that moves to something else in eighteen months keeps the connection and changes the assistant. You are not buying a language model from your accounts software provider.
Permissions, and what the AI can and cannot see
Connecting AI to client data is safe when the connection cannot do anything the person using it could not already do. Access to Active authenticates as the individual user and is checked on every call, so an assistant can never see or do more than the person signed in.
A firm sets this twice. Once for the firm, choosing whether access is read-only or read and write, and read-only is where most start. Then person by person, ticking who gets it at all. Anybody not ticked has no access, and turning somebody off takes effect straight away rather than at the next renewal.
Nothing is copied or synced in the background and no default quietly opens later. The only thing reaching an assistant is what that person asks for, at the moment they ask, one record at a time.
The line that matters most is drawn in the software rather than a policy. Approving and completing a workpaper is refused to the assistant outright, whatever rights the person signed in holds, and handed back with a link. You can read how client data is protected and certified in full.
Pre-built skills, and building your own
The guided jobs are pre-built skills, so the instruction is an ordinary sentence and the work behind it is already defined. Just ask the question as you would normally and Active will work it out.
The skills Active publishes are free, in a public repository. Anyone can read them, fork them or write their own. What a firm builds from there is specific to how it works: its review conventions, the way it grades a job, the questions its partners always ask and nobody has written down.
Most of that lives in the heads of people who have been there long enough. Built as a skill, a firm's way of working stops being something it explains and becomes something the system does. Work anybody does the same way is a commodity. Work shaped by how your firm does it is not, and enforceable standards without removing professional judgement is what makes that difference durable rather than personal.
Active AI is one module, licensed per user the same way the rest of Active is. No tiers, no light-versus-full split, and the MCP server sits inside the licence rather than on a separate line of the bill. When you use Active through the MCP server, the thinking happens on your own Claude or Copilot subscription, so no meter runs at Active's end and no Active bill moves with how much your team uses it. For anyone told to budget for AI costs nobody can forecast, that is the useful part.
Active AI is free until 31 December 2026, with charging from 2027. Firms come on by invitation, so get in touch if you want it switched on.
If your firm is not ready for any of this, Active works as it does today, and the AI already built into Active covers the part running inside the product. We are walking through all five on a live file at our webinar. To get on the preview, take a look at Active's accounts production and book a demo.
Frequently asked questions
Can ChatGPT access my accounting software?
Active's MCP server is built to an open standard, so any assistant that speaks it can be pointed at Active. Today that means Anthropic's Claude, where the Active plugin is deployed, and Microsoft 365 Copilot, connected through Copilot Studio or the Admin Centre. Both need an Active AI licence.
Can I use Copilot with my accounts production software?
Yes. Microsoft 365 Copilot is set up through Copilot Studio or the Admin Centre, and it is about as common as Claude across Active's firms. The tools are identical whichever assistant reaches them. The guided query and filing skills ship inside the Active plugin for Claude, so Copilot gets the tools without those.
Do I need to know how to write prompts to use this?
No. Nobody types tool names. You ask in plain English and the assistant works out what it needs, and the guided jobs are pre-built skills with the work already defined behind them. If you do want your own, the skills Active publishes are free in a public repository to read, fork or rewrite.
What is Active AI?
Active AI is the single module carrying every AI capability in Active, including the MCP server and the tools built into the product. It is licensed per user, with no tiers and nothing further to unlock by paying more. A firm chooses who gets it, or declines it entirely.