Sep 23, 2026

How to Correct an AI Assistant That Got It Wrong

Your assistant just replied on the wrong thread, or booked next month instead of next week, or sent a polite reminder to a client who paid on Friday.

You can apologize and fix it by hand in two minutes. The real job is making sure it never happens a second time.

The quick answer: figure out which of four causes produced the mistake (a vague instruction, missing context, a stale tool connection, or a plain model slip), fix that root cause, and write the rule down somewhere the assistant will read it every time, not in a chat it will forget. In Carly, that means the rule goes into the agent’s standing instructions, and you add a “require approval” gate for that one action type until the agent earns your trust back.

The four reasons an assistant gets it wrong

Name the cause before you fix anything; each one has a different fix.

Cause	What it looks like	The fix
Ambiguous instruction	”Follow up with the client” and it followed up with the wrong contact.	Rewrite the instruction so a new hire could not misread it: name the person, the date format, the thread.
Missing context	It pulled last quarter’s revenue into the board summary. It filed the signed contract in the general folder because nobody told it each client has its own.	Give it the source of truth: the exact sheet, folder or CRM field, and say which one wins when two disagree.
Stale or wrong tool connection	It chased an invoice that was paid, because the accounting connection expired and it read old data.	Reconnect the integration, remove the account it should not use, and confirm it can read the live record.
Plain model error	Everything was right and it still mixed up two names or misread a date once.	Add a check on that step (an approval gate or a verification rule) and re-run. If it repeats, it is not a plain error; go back up this table.

The first three are permanent fixes. Only the last is the model’s fault, and it is the rarest.

Fix this one, then fix the next hundred

Run the same loop every time, whichever assistant you use.

Point at the exact error. Not “that was wrong,” but “you replied on the Q3 renewal thread; the question was on the new-project thread from Tuesday.” Specific corrections produce specific rules.
Ask it to show its source. “Which email, file or record did you use?” The answer often names the cause: an old thread, the wrong account, a cached number.
Decide the root cause. Match it to one of the four rows above. Skip this and you patch symptoms.
Rewrite the instruction or reconnect the tool. Edit the standing instructions for ambiguity and missing context. Reconnect or remove the integration for a stale connection.
Add a check. For anything that leaves your company (an email, an invoice reminder, a calendar invite to a client), put an approval step in front of it for now.
Re-run on the same input. Hand it the exact task that failed. If it still gets it wrong, you picked the wrong cause.

What this looks like in Carly

Every Carly agent has its own real email address, a name and a written set of instructions. You correct it the way you would correct a colleague.

Reply to the agent’s email with the correction. When an agent gets something wrong on a thread, reply to it: “This was the wrong thread. From now on, when a client has more than one open thread, reply on the most recent one from them unless I say otherwise.” The agent fixes the instance and remembers the rule. For anything you want in writing, put it in the agent’s instructions on the agents page as well. It works the same in Gmail and Outlook.

Edit the agent directly when the fix is bigger. Open the agents page, pick the agent and rewrite its instructions in plain English. This is the right place for context it was missing: “Client contracts go in Google Drive under Clients / [client name] / Contracts. Never file a contract in the shared root folder.” Giving each agent its own name and email makes it obvious which one slipped.

Add a “require approval” rule for that action type. This is where Carly differs from assistants built to draft and wait. Approval-first tools push every action back to you, so one invoice mistake means approving calendar holds too. In Carly you choose per action type: hold external emails about payments for your OK, and let scheduling keep running on its own. Once the agent has been right on that action for a couple of weeks, remove the gate. The first 30 days with an AI agent guide walks through that trust curve.

Check the workflow run log and rerun. If the mistake came from an automation rather than an agent conversation, open the workflow on the workflows page and read the run history. Each run shows which step fired and what data it used, so you can edit the broken step and rerun it on the same trigger.

Reconnect a stale integration. If the agent read old data, open the integrations page, find the app, and reconnect it. Carly connects to thousands of apps, including QuickBooks, Xero, Stripe, HubSpot and Google Drive, and an expired login on any one of them looks exactly like a reasoning error from the outside.

A real example: the client who already paid

Josh runs a small brand studio, and Warby Parker is one of his retainer clients. His finance agent, “Ava,” chases overdue invoices every Monday. One Monday it emailed Warby Parker’s accounts payable team about an invoice they had paid by card the Friday before.

Josh replied to Ava’s email: “Where did you check payment status?” Ava answered that it used the QuickBooks invoice list. That was the clue. Warby Parker paid through Stripe, and the Stripe payment had not synced into QuickBooks yet. Cause: missing context, plus a sync delay.

Josh’s fix took three minutes:

He replied to Ava with the rule: “Before chasing any invoice, check Stripe for a payment from that client in the last 7 days. If you find one, do not chase; tell me instead.”
He confirmed Stripe was connected alongside QuickBooks.
He added a “require approval” rule on payment-reminder emails for two weeks.
He asked Ava to send Warby Parker a one-line apology, and approved it before it went out.

Then he re-ran the Monday chase on the same invoice list. Ava found the Stripe payment and skipped Warby Parker. Two weeks later Josh removed the approval gate. His colleague Claire set up the same Stripe check on her own agent the same afternoon.

Corrections that stick vs corrections that vanish

Correction lives in	What happens next time
A chat thread (“remember, not that folder”)	Gone the moment a new conversation starts, or buried under forty messages of context.
A one-off reply that only fixes this instance	The instance is fixed. The next instance repeats the mistake.
The agent’s standing instructions	Read at the start of every task, every time. This is the only place a rule reliably survives.
An approval gate on the action type	Catches the mistake even if the rule was misapplied, until you decide it is safe to remove.

General chat assistants rarely have a durable place for the rule, so you correct the same thing every week. With Carly, each agent’s instructions are that durable place, and a reply to the agent is the quickest way to get a rule there.

When to stop correcting and change the setup

If you are correcting the same agent more than once a week, the instructions are not the problem. The setup is.

Split one broad agent into two scoped ones. An agent handling scheduling, invoicing and support has three rule sets competing for attention. Give scheduling and billing to separate agents, each with short instructions. See what Carly can do for how teams usually divide the work.

Move rules-only tasks into a workflow with no AI step. “When a Typeform comes in, add a row to Google Sheets and send the welcome email” has no judgment in it. Build it as a workflow and it runs the same way every time. Carly offers free Zapier-style workflows; AI agents from $35/month, so moving the mechanical parts out of the agent costs you nothing.

Keep a human on the send for money and contracts. Refunds, pricing changes, signed agreements and anything legally binding should stay behind a permanent approval rule. If you want help deciding which actions to gate for your team, book a call with the Carly team.

Frequently asked questions

What should I do right after my AI assistant makes a mistake?

Fix the instance by hand if urgent, then ask which source it used. In Carly you can reply to the agent’s email with both the question and the correction, then add the rule to its instructions so it holds. Get started to try it on a real thread.

Why does my AI assistant keep making the same mistake?

Because the correction lives in a chat thread instead of its standing instructions. Put the rule where the assistant reads it at the start of every task. In Carly that is the agent’s instructions on the agents page.

Should I make my AI assistant ask for approval on everything?

No. Gate the specific action type that went wrong, such as payment reminders or external emails to one client, and let everything else keep running. Carly lets you set “require approval” per action type, so one mistake does not slow down the rest of the agent’s work.

How do I know if the mistake was the AI or a broken connection?

Ask it what data it used and check that record yourself. If the record in the app is newer than what the assistant saw, the connection is stale. Reconnect it on the Carly integrations page and re-run the same task.

Can an AI agent handle money or contract emails safely?

Yes, with a human on the send. Keep a permanent approval rule on refunds, pricing and contracts, and let the agent draft, file and track everything around them. If you want help setting that up for a team, book a call.

Ready to automate your busywork?

Carly schedules, researches, and briefs you—so you can focus on what matters.

See what people say

"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.

Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.

On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."

Gus Ibrahim, Founder & Director, IHR