The 30-second answer

If 100 people started a chat with your AI last week and 60 of them never spoke to a human, your containment rate is 60%. That's the whole calculation. The number is real, and on its own it overstates how well the AI is doing.

Why it overstates

A conversation can end without a human in three ways, and only one of them is good:

  • The customer got a real answer and went on with their day.
  • The customer gave up and closed the window. Still counts as "contained."
  • The customer was sent in circles and silently emailed you about the same problem instead. Also still counts as "contained."

The bot, the dashboard, and the ROI calculator cannot tell these three apart. The customer can.

Why every vendor leads with it

Containment is easy. It comes out of one channel's logs. It rolls up to a single big-sounding percentage that fits on a slide. It moves when you tune the bot, which makes it feel like a real lever. And it maps cleanly to a cost story: each "contained" contact is a human ticket that didn't happen, multiplied by cost-per-ticket.

Every AI-agent vendor's ROI calculator uses containment as the headline: Intercom Fin, Ada, Forethought, HubSpot Breeze, and most of the rest. The math is technically right; the framing is generous.

Containment vs deflection — same thing, different label

Most vendors use "containment" and "deflection" interchangeably for the same underlying number — conversations that ended without a human. A few try to draw a distinction:

  • Containment as the single-channel, single-session number ("the bot closed the chat").
  • Deflection as the cross-channel, multi-day version ("the customer did not come back about the same thing within 7-30 days").

The second one is harder to measure because it requires identity stitching across channels. Most companies don't have that, so most companies measure containment and call it deflection.

What to measure alongside

Two corrections, both essential:

  • Resolution rate. The share where the customer's problem was actually solved, confirmed by a survey or the absence of recontact.
  • Recontact rate. The share of "contained" customers who came back about the same thing through any channel within 72 hours.

A reported 70% containment, paired with a 35% recontact rate, means your real resolution is more like 45%. Same data, very different story.

Honest range vs reported range

Vendor case studies quote 50-70% containment as a normal outcome for a mature AI agent. Independent measurements that subtract recontact land in the 30-50% range for real resolution. The gap is the part the vendor dashboard doesn't show. When someone quotes you a containment number, ask for resolution and recontact alongside; if they can't produce them, you only have half the picture.