Why your AI sales agent isn't converting (and how to diagnose it in a week)

There is an uncomfortable moment in almost every sales AI project. The agent is live, talking to hundreds of leads a day, it doesn't freeze, it doesn't break the integration, it logs everything in the CRM. The goal was to book 15% of leads. The real number is close to zero.

The first reaction is usually to hunt for a bug. Sometimes there is one. But most of the time, the flow is working exactly as designed, and the problem is the design. Fixing technical errors is necessary, but it isn't strategy. An agent that runs smoothly while carrying out the wrong approach still doesn't sell.

This article presents the method we use to diagnose AI agents that don't deliver: split the problem into three layers, break the metric down from back to front and read the conversations at scale.

The three layers where an AI agent fails

Every result from a sales agent depends on three things working at the same time. If any one fails, there's no result, and each one calls for a different kind of fix.

The trap is spending weeks fixing the bottom layer while the problem sits in the top one. A simple question helps tell them apart: if the agent did exactly what it's programmed to do, with no errors at all, would the goal be hit? If the answer is "I don't know", the problem is probably strategy.

Break the metric down from back to front

Every project has a key indicator: booked meeting, sale, quote sent. The diagnosis starts there and works backward, like a tree. A booking is not an event, it's the end of a sequence of stages. Each stage has a conversion rate. And each stage corresponds to a section of the agent's playbook.

An illustrative example for a booking goal:

With the tree built, the break jumps out. In this example, the agent attracts, engages and qualifies well. The funnel dies at the invitation: people get the meeting offer and don't accept it. There's no point touching the opener. What needs investigating is why the meeting doesn't seem worth it.

This reverse analysis should exist for every project from day one, not only when results disappoint. Once the tree is drawn, checking it becomes routine.

Ask "why" until you reach the leaf

Once you've found the stage that breaks, the next question is why. And the answer is rarely at the first level.

People don't accept the meeting. Why?

Because they don't understand what they'll get out of it. Why?

Because the message talks about "a conversation about a special partnership" without saying what it is. Why?

Because the team can't promise terms before the meeting and the agent was instructed not to go into any detail.

Now there's something actionable: define what can be said to make the meeting concrete without promising what can't be promised. You keep going down until you reach a point where you can no longer answer "why". That's the leaf of the tree, and that's where the action lives.

Read the conversations, all of them

Numbers show where. Conversations show why. The problem is that nobody can read a thousand conversations a week, and a sample of ten usually tells the story you were already expecting.

The solution is to use AI to read the conversations AI ran:

• A score per conversation. Each conversation assessed against the playbook: did the agent ask the right question, present the offer, handle the objection?
• Objection mining. Which objections come up, how often and at what stage.
• Competitors mentioned. Who shows up in the conversation and in what context.
• Loss reasons. Structured, not in a free-text field nobody reads.

That's how Compass was born at Sailer: an internal tool to explain to customers why an agent converted, or didn't. It became our most requested feature.

Diagnoses that come up often

Some patterns repeat across projects:

• The lead distrusts the channel. A message from an unknown number asking for personal details triggers fear of a scam. Some ask to continue by email. The fix involves signals of legitimacy: a clear company name, context on where the contact came from, an official link.
• The purpose of the approach is vague. "We'd like to talk about an opportunity" gives no reason to reply. A concrete value proposition, right in the opener.
• Too many questions before delivering value. The agent qualifies with five questions before saying anything useful. Flip it: a bit of value, then a question.
• Follow-up that repeats the message. "Hi, did you see my message?" three times. Every follow-up needs to bring something new. See our follow-up cadence.
• Handoff too late. The warm lead keeps chatting with the AI when they already wanted to talk to a person.
• An objection that never changes. If everyone gets stuck on price, the conversation isn't the problem. The offer is. That's the kind of finding that needs to go back to marketing and product, and it's at the center of a RevOps operation.

The weekly ritual: what worked, what didn't, how to improve

The diagnosis only turns into results if there's a routine. The format we've adopted with customers is simple, five questions every week:

1. What worked?
2. Why did it work?
3. What didn't work?
4. Why didn't it work?
5. How do we make it work better?

Alongside it, a one-page summary per account: the project goal, today's number, where it's weak and the action plan with an owner and a date. Bringing this to the customer meeting changes the nature of the conversation: instead of "we're making adjustments", you present "this worked because of this, that didn't work because of that, and here is the plan".

A one-week method

If your agent isn't delivering, here is a sequence that fits into five days:

• Day 1. Write down the key indicator and build the stage tree with the real conversion rates.
• Day 2. Sort the problems you find into the three layers. Fix anything on the platform layer immediately.
• Day 3. Read the conversations from the stage that loses the most, with scoring and objection mining.
• Day 4. Ask the whys down to the leaf and write two or three hypotheses for changes to the playbook.
• Day 5. Put the changes live for part of the traffic and decide when you'll measure.

The following week, the ritual starts again with fresh data.

Frequently asked questions

How long should I wait before concluding an agent doesn't work?

More than a few days and less than a few months. With reasonable volume, two to four weeks with the weekly ritual is usually enough to know whether the strategy has a future.

Could the problem be the AI model?

Rarely. Today's models converse well. What's usually missing is the playbook, context and integration with the right systems.

Do I need a data analyst to do this?

Not to get started. The reverse tree fits in a spreadsheet. What takes a tool is reading conversations at scale.

The agent doesn't convert, but the humans do. What does that tell me?

That there's knowledge in your salespeople's heads that isn't in the playbook. Sit down with your best salesperson, map their objections and answers and bring that into the agent.