Resolution, not deflection

Every AI support vendor has a slide with a big deflection number on it. We had one too, back in 2021. It said 80%.

The number was accurate, and it was useless. It counted every conversation that ended without a human touching it. Some of those customers got what they needed. Some of them gave up. Some closed the chat window, opened their email client, and wrote to the support address in capital letters. Deflection counted all three the same way.

We stopped reporting it internally in 2022. Since then we've stopped reporting it to customers, stopped quoting it in sales conversations, and rebuilt our pricing around a different number. This post explains why.

What deflection actually measures

Deflection rate is usually defined as the share of inbound contacts that don't reach a human agent. It's easy to compute and easy to improve. Unfortunately, the easiest ways to improve it make customers worse off:

• Make the handoff harder to find. Bury "talk to a person" behind three menus. Deflection goes up.
• Answer the question instead of solving the problem. "Here's our refund policy" ends the chat. The refund still hasn't happened.
• Time out quietly. A customer who walks away mid-conversation counts as deflected.
• Push people to a different channel. "Please email us at…" is a deflection in chat and a new ticket in email.

None of these are hypothetical. When we audited conversations from teams migrating to Pazau from other tools, we found that roughly one in three "deflected" chats was followed by a repeat contact on the same issue within a week. The customer hadn't been helped. They'd been delayed.

A deflected customer is often just a customer who's going to contact you again, angrier, through a more expensive channel.

The cost shows up somewhere else, which is why it's so easy to miss. The chat dashboard looks great. The email queue grows. Phone volume creeps up. CSAT drifts down and nobody connects it to the bot.

What we measure instead

We count a conversation as resolved only when all three of these are true:

1. Pazau Agent handled it end-to-end. No human agent read it, replied to it, or approved anything in it.
2. The customer didn't come back. No repeat contact about the same issue, on any channel, within 7 days.
3. The customer didn't bail. They didn't abandon the conversation mid-procedure, and they didn't explicitly ask for a human.

If any condition fails, it isn't a resolution. A handoff is a handoff, even if the AI did 90% of the work first. A conversation where the customer typed "agent" three times and then left isn't resolved, even though no human ever saw it.

This is stricter than any definition we've seen elsewhere, and it's deliberate. The 7-day window catches the "answered the question, didn't fix the problem" failure. The abandonment rule catches the "hid the exit" failure. Requiring no human touch at all catches the quiet cheat where a person approves every AI reply and the vendor still calls it automated.

What this looks like in practice

Here's a real (anonymized) week from a retail customer, measured both ways:

Metric	Count	Rate
Inbound conversations	48,210	—
Ended without a human ("deflected")	39,050	81%
…of which the customer contacted again within 7 days	4,870
…of which the customer abandoned or asked for a human	2,140
Resolved	32,040	66%

The gap between 81% and 66% is where customers were getting failed. Every one of those 7,010 conversations cost the brand something: a repeat contact, a worse review, a customer who now believes the chat widget doesn't work.

Why we bill on it

Measuring resolution is one thing. Billing on it is how you prove you mean it.

Pazau charges per resolution: $0.90 on Growth, $0.75 on Business, and $0.99 for Agent-only deployments. Handoffs are free. Abandoned conversations are free. Repeat contacts un-resolve the original conversation, and we don't bill it.

This aligns our incentives with yours in a way per-seat or per-conversation pricing never can:

• If Pazau Agent hands off too early, we don't get paid.
• If it ends conversations before the job is done, the 7-day window catches it and we don't get paid.
• If it frustrates customers into leaving, we don't get paid.

The only way for us to make more money is for the agent to finish more work. That's the product we want to be building anyway.

It also forces a kind of honesty into sales conversations. We can't promise you 80% because a flattering definition would let us hit it. Across our customer base, Pazau Agent resolves 67% of conversations end-to-end on average. For enterprise deployments, where procedures are more complex and more of the work requires regulated actions, the median is 58%. Those numbers are lower than the deflection figures you'll see on other vendors' homepages. They are also real.

Resolution changes what you build

Once you commit to resolution, the product roadmap reorganizes itself. You stop investing in ways to keep customers in the widget and start investing in ways to actually complete the task.

For us, that meant three things.

Actions over answers. A customer asking where their refund is doesn't want a policy page. They want the refund. So Pazau Agent connects to order management, billing, booking and claims systems, and does the work: issues the refund, rebooks the flight, updates the address. Most of the gap between answering and resolving is closed by being able to take the action.

Procedures over prompts. Your best human agents don't improvise. They follow a process, and they know when the process says to stop and escalate. We model that directly, with SOPs written in plain language that the agent follows step by step. That's the subject of a longer post from our Head of AI.

Honest handoffs. When the agent can't resolve something, it should say so quickly and hand over the full context, so the human doesn't have to start from zero. A fast, clean handoff doesn't count as a resolution for us, but it's much better for the customer than a slow, confused one. Customers can ask for a human at any point, in any words, and they get one.

The objection we hear most

"Resolution is harder to measure. Our BI team can compute deflection from a single table."

That's true. Resolution requires linking conversations across channels and across days, which means you need identity resolution and topic matching. We had to build both. Pazau computes resolution automatically for every conversation, shows you the ones that failed each condition, and lets you drill into the trace to see why.

But "harder to measure" isn't a reason to measure the wrong thing. Support leaders don't get promoted for deflection. They get promoted for lower cost per contact and higher CSAT, and resolution is the metric that actually moves both. Deflection moves neither, reliably.

If you want the full breakdown of how cost per contact shifts when you optimize for resolution, including the hidden costs of repeat contacts, we've since written it up in The real cost per contact.

What we're asking of the industry

We'd like to see three changes across AI support, not just at Pazau:

1. Publish your definition. If you quote a rate, say exactly what counts. Include the repeat-contact window.
2. Report repeat contacts next to automation. A single number without the recontact rate beside it can't be trusted.
3. Price on outcomes. If the vendor gets paid whether or not the customer is helped, the vendor will optimize for something other than the customer.

None of this is complicated. It's just less flattering. We think customers, and the support leaders who answer to them, deserve the less flattering number.

See your own number

The fastest way to understand the gap is to measure it on your own data. Bring us 1,000 real conversations from last month. In a 30-minute call we'll show you what Pazau Agent would have resolved under the definition above, what would have been handed off, and why.

Book a demo and we'll run it with you.