Who is accountable when an AI agent works the wrong account?
"The AI did it" is not an answer. Give the model, the rules, the agent, and the review each an owner, and a bad email leads to a gap someone can close.
Here is a made-up Tuesday. An AI agent sends a friendly first-touch email, with an offer of an introductory call, to an account that has been a customer for three years and is in the middle of a tense renewal. The customer forwards it to their account manager with one line: “Do you people talk to each other?”
By ten o’clock there is a thread asking who is responsible. The honest first answers are “the agent did it” and “the score was high”. Neither is a person. This piece is about how to make sure there is one.
Accountability attaches to things a person wrote
An agent cannot be accountable. It has no job to lose and nothing to learn from being told off. The people around it can be, and only for what they decided. So the practical question is: which decisions that led to this email were written down by someone, and which were left for the agent to make up?
Every decision the agent made up is one nobody owns. That is the gap to close, and it closes the same way each time: take the decision out of the agent and put it somewhere with a name on it.
Four things that need an owner
| What | What it decides | Who usually owns it | What they answer for |
|---|---|---|---|
| The model | Which records matter, and why. The factors and weights behind the score | Revenue operations, with sales leadership | A record that scored well and should not have |
| The rules | What the agent may do at each score, and who is off limits | Sales leadership | A record the agent was allowed to contact and should not have been |
| The agent | How the tool is set up: its instructions, what it can read, what it can send | Whoever runs the tool, often operations or marketing | An agent that had the right score and the right rules and did something else |
| The review | Whether anyone looked at what went out | A named manager | A problem that ran for weeks before it was caught |
Four owners is not four people. In a small team, one person may hold three of these. What matters is that each has a name next to it, written where the team can see.
Tracing the Tuesday email
With the four owners in place, the question “who is responsible?” becomes four smaller questions, asked in order.
1. Was the score wrong? Open the record and read the score and its factors.
- Industry 30% 8.9
- Company size 25% 7.6
- Contacts in buying roles 25% 8.4
- Region 20% 7.2
The math: 8.9 at 30% is 2.67, 7.6 at 25% is 1.9, 8.4 at 25% is 2.1, and 7.2 at 20% is 1.44. That adds to 8.11, shown as 8.1.
The score is correct for the model as written. A three-year customer is, unsurprisingly, a very good match for your ideal customer. The model has no disqualifying factor for current customers, so nothing took the score down. That is a gap in the model, and it belongs to the model’s owner.
2. Was there a rule that should have stopped it? Read the agent’s rules. If none says “do not contact current customers”, that is a second gap, and it belongs to the owner of the rules. A good setup has the stop in both places, so one failing does not send the email.
3. Did the agent break a rule it was given? If the rule existed and the agent sent anyway, the problem is in how the agent is set up. That belongs to whoever runs the tool.
4. Did anyone look? If this was the fortieth customer to get that email, the review failed too.
In this example the answer is the first two. Nobody wrote down that customers are off limits, in the model or in the rules, so the agent decided for itself. The fix is one disqualifying factor and one line of instruction. The accountability is shared between two named owners, and neither of them is “the AI”.
The incident checklist
Keep this on one page. It is short because the first hour is not the time to think.
In the first hour
- Pause the agent, or the one sequence involved. Do not wait to understand the cause.
- List every record it contacted since the last review.
- Have a person, not the agent, reply to anyone who should not have been contacted.
In the first day
- For the record that raised the problem, copy the score, every factor, and the time it was calculated. Do this before anything recalculates, so you have what the agent saw.
- Ask the four questions in order: score, rule, agent, review.
- Name the owner of the gap. Tell them plainly, without blame.
In the first week
- Fix it where it belongs. A gap in the model gets a factor. A gap in the rules gets a rule. Do not patch a model problem with a line in the agent’s prompt.
- Check the fix on the records from step 2. They should now be stopped.
- Write down what happened, what changed, and the date. Keep it with the rules.
- Turn the agent back on.
Three ways teams get this wrong
Blaming the tool. “The AI did it” ends the conversation without changing anything. The tool will do the same thing tomorrow.
Patching the prompt. Adding “never email customers” to the agent’s instructions fixes this agent. The model still scores customers as top prospects, and the next agent or workflow you connect will read the same score with no such line.
Turning it off for good. One bad email is a reason to fix a gap. Scrapping the agent means the gap is still there for the next thing you build.
Where BoostIdeal fits
BoostIdeal covers the first owner’s job. The model is a list of factors and weights that a person sets, with AI helping to analyze your data and draft. Every score is a weighted sum out of 10, and BoostIdeal can share the reasons behind it wherever your team works: in a HubSpot field on the record, in Slack, or by email. When someone asks why an account scored 8.1, the answer is four lines long and has an owner.
Two limits matter for incidents:
- It does not run the agent or enforce the rules. Those stay in your own tools, with their own owners.
- It does not keep a score history. A record shows its current score and when it was last calculated. That is why step 4 says to copy the breakdown straight away. If you need a record of what an agent saw at the time, have the agent log the score and top factors with each action.
If the model has not been written down yet, that is the first gap to close. Start with your ideal customer profile, the five-step framework for building an ICP scoring model, and the list of common ICP scoring mistakes.
Frequently asked questions
Who is accountable when an AI agent contacts the wrong account?
The person who owns the decision that failed. If the score was wrong, the owner of the model. If no rule stopped it, the owner of the rules. If the agent ignored a rule, whoever runs the tool. If nobody noticed, the reviewer.
Can the AI vendor be held responsible?
For the tool working as described, yes. For your priorities, no. Which accounts matter and who is off limits are your decisions, and they need to be written down on your side.
How do we stop an agent from emailing current customers?
Put it in two places. Add a disqualifying factor for current customers to the scoring model, and add a rule to the agent’s instructions. Either one should stop the email on its own.
What should we record when something goes wrong?
The score, every factor, and the calculation time for the record involved, copied before anything recalculates. Then the cause, the owner, the fix, and the date.
How often should someone review what an agent sends?
Weekly is a reasonable start. Read the list of records it skipped and a sample of what it sent. Review more often after any change to the model or the rules.
The takeaway
When an agent works the wrong account, the useful question is which decision nobody wrote down. Give the model, the rules, the agent, and the review each a named owner. Then a bad email leads to a specific gap and a specific person who can close it, instead of a thread that ends with “the AI did it”.