Human-in-the-Loop: why Cheku will never touch your records without asking

There is a fear running quietly through every accounting firm right now: is the AI going to replace me?

The honest answer, from the people building one of these tools: no. And if you ever meet a vendor who says otherwise, close the tab.

Accounting is not a data-entry job. It is a judgement job wrapped in a paperwork job. AI is excellent at the paperwork. It is genuinely bad — and will remain genuinely bad for the foreseeable future — at the part where a client describes a messy divorce, a failed restructure, or a director loan they “forgot” about, and you have to decide what that means for the return.

What AI can do is stop you spending your Tuesday afternoon drafting the same follow-up email for the eighteenth time. That is the promise. Everything else is marketing.

What “Human-in-the-Loop” actually means

Human-in-the-Loop (HITL) is a design principle, not a feature. It says: the AI can read, research, draft, and suggest — but it cannot do. Every action that changes a record, sends a message, or touches a client file has to pass through a human.

In practice that means three boundaries:

1. The AI proposes. The human disposes. Nothing leaves the building without explicit sign-off.
2. Every suggestion shows its working. You see what the AI read, where it pulled from, and why it concluded what it did.
3. Dismissal is as easy as approval. One click to reject, no guilt, no “are you sure?” modal trying to talk you back into it.

The word we keep coming back to internally is preview. Never auto.

How Cheku does it

Open the Agent Inbox and you will see what the AI has been working on while you were in a meeting. Four kinds of cards, all of them waiting for you:

• Drafted emails — a client has gone quiet on a VAT document request. The agent drafts the follow-up in your firm’s voice, citing the exact documents outstanding and the filing deadline. You read it, edit two words, hit send. Or you delete it.
• Record updates — a client signed a new engagement letter from sarah.jones@jonesco.com instead of the sarah.j@oldmail.com on file. The agent flags it. You approve the contact update or ignore it. The old email is never silently overwritten.
• Research summaries — you asked about Corporate Interest Restriction on a group consolidation. The agent comes back with the HMRC manual reference, the relevant legislation, and how it applies to the specific client’s accounts. You read the citations yourself before you trust the conclusion.
• Follow-up reminders — a query has been open for five days. The agent nudges you, with the context already pulled up. It does not nudge the client.

What the agent explicitly cannot do:

• Send an email on your behalf without approval.
• Edit a client record without approval.
• File anything, anywhere, ever.
• Mark a query as resolved without approval.

That is not a limitation we are working to remove. It is the product.

Why we designed it this way

We spoke to a lot of accountants before we wrote a line of code. The consistent message was not “I want AI to do my job”. It was “I want AI to stop me doing the parts of my job that aren’t my job”.

Chasing clients for documents is not accounting. Typing the same status-update email nine times a week is not accounting. Re-reading the same HMRC page to find a reference you already half-remember is not accounting.

Approving a drafted letter, deciding what a piece of legislation means for a specific client, catching that something in the numbers looks off — that is accounting. And that is yours.

The trust ladder

We think about trust as a ladder, not a switch. Firms climb it at their own pace:

1. Rung one — research only. The AI answers questions. Nothing it says leaves the chat. Good for two weeks of kicking the tyres.
2. Rung two — draft mode. The AI starts drafting emails and record updates. Everything sits in the Agent Inbox. You approve or dismiss. No exceptions.
3. Rung three — trusted patterns. You have approved 200 “client email address change” updates from the agent. Every one was correct. At that point, and only at that point, you might choose to let that specific pattern go through with a lighter-touch review. Your choice, your pattern, your audit trail.

Most firms sit comfortably on rung two forever. That is completely fine. The ladder is not a funnel.

What this means for your firm

If you take one thing away: AI in accounting only works when the accountant stays in the loop. Any product that tries to remove you from the workflow is either dangerous, or about to be.

The goal is not a firm with fewer accountants. It is a firm with the same accountants, doing less admin and more advisory, winning more clients because they have the headspace to actually talk to them.

That is what we are building. And we will never ship a feature that breaks that rule.