Agent Intelligence
See what your AI keeps getting wrong
PRISM brings user intent, sentiment, sessions, failure patterns, knowledge gaps, and priorities into one view.
- Balance inquiry24%94%
- Payment change18%88%
- Dispute escalation11%68%
- Fee exception request9%63%
- Identity verification7%72%
- Statement request6%91%
- Card replacement5%86%
- Wire transfer help4%79%
What it brings together
Three answers from every day of production
Know what people came to do
Group similar requests into the jobs and questions that matter most, and see how confidence and frustration move across each journey.
User intent · sentiment
Find the problems that keep coming back
Group runs that share the same cause, and surface the repeated questions your AI could not answer from its knowledge.
Failure patterns · knowledge gaps
Know what deserves attention next
Rank work by user impact, repeat rate, severity, and how many journeys the change may improve.
Priorities
The problem
Your agent has a score, and the score is lying to you
Not deliberately. An average over everything an agent does is arithmetic, and arithmetic hides the two things it does badly behind the six it does well. Teams ship on 87% and find out what the other 13% was from a customer.
87%
is the number a dashboard shows you
3 of 8
things it does are actually broken
27%
of all traffic hits one of those three
Holding
5 of 8
- Balance inquiry94%11,570
- Statement request91%2,893
- Payment change88%8,678
- Card replacement86%2,411
- Wire transfer help79%1,928
Below the 75% floor
3 of 8
- Identity verification72%3,375
- Dispute escalation68%5,303
- Fee exception request63%4,339
27% of the month’s traffic — 13,017 sessions — landed in one of these three.
Fee exception request · one session of 4,339
4.1 turns avg · 63% reliable
- 1userDo international fee waivers apply on my account tier?
- 2agentRetrieved: general fee policyretrieval ok
- 3agentkb.search → fee_waiver_policy200 OK
- 3systemSame query repeated · 3× in one sessionloop
- 4userSession abandonedfail
What was missing · the tier exception clause. Every step succeeded, the tool returned 200, the model answered fluently — and the answer was not in the knowledge base. This is why reading logs never finds it: no single layer looks wrong.
Fee-exception answers missing tier exception clause
- Evidence
- 214 failed sessions · 4-turn avg loop
- Fix
- Update KB article, add retrieval tags, revise answer template
- Owner
- Knowledge Eng
- Expected
- Fee-exception reliability 63% → ~80%
37% of severe failures trace to missing or stale content rather than model quality — which is why this ends with a content owner and a date, not a retraining run. The same clustering that found the pattern grouped 214 sessions behind one missing article.
1,605
sessions ended without an answer
985
handed to a person instead
4.1
turns before they gave up, on average
One intent, one month. Derived from 4,339 sessions at 22.7% escalation.
- Human escalations avoided~154 support hours · ~$5,400
Escalation rate fell 9.2% → 7.6% of 48,210 sessions = 771 avoided escalations
- Engineering triage eliminated~48 engineering hours · ~$4,300
Failure clustering + root-cause classification replaces manual log review
- Completions recoveredRevenue-linked
Refund KB fix recovered ~380 sessions; refusal tuning adds ~120/month
Who this is for
You already know something is wrong. You cannot name it.
The team running the agent
Today · Support keeps escalating the same thing and nobody can say which flow it is.
With PRISM · A named intent, a session count, and a fix with an owner.
The person accountable for it
Today · Asked monthly whether reliability is improving, answering from vibes.
With PRISM · A score with six dimensions under it and a delta that survives scrutiny.
Whoever gets the audit
Today · Evidence lives in logs nobody can hand to a regulator.
With PRISM · Every claim drills to trace level and exports.
All three are the same complaint at different altitudes: something is wrong and nobody can name it. Naming it is the entire product.
How it works
One session, kept whole
Intelligence is not a separate dataset. It is what falls out when the intent, the state, the tool call and the outcome stay attached to each other instead of landing in four systems.
One session from the fee-exception cluster. Nothing in any single layer looks wrong — the failure is the disagreement between the first layer and the last.
Classified on every session, not a sample
“214 sessions with this exact signature” is a sentence you can only write if you classified all 48,210 of them.
The failure is between layers
The tool returned 200. The retrieval worked. The model answered fluently. Nothing in any single layer looks wrong, which is exactly why log review never finds this.
It ends in an owner and a date
A pattern that stops at “knowledge gap” is a finding. One carrying a fix, a named owner and an expected lift is a backlog item.
What it finds
Why they fail, and what nobody answered
37% is missing or stale content, not model quality — so the fix is a content owner and a knowledge-base update, not a retraining run.
- Knowledge gap / missing content37%KB update
- Workflow / tool failure24%Code PR
- Prompt logic18%Prompt version
- Retrieval quality13%Retrieval config
- Guardrail / policy config8%Guardrail rule
PRISM root-cause classifier
214 people asked the same thing about tier fee waivers in one month and none of them got an answer. It maps to a single missing article.
- 214Whether international fee waivers apply by account tier
- 167How to escalate a merchant dispute past the first rejection
- 88Whether early payoff changes autopay schedules
Figures from a sample FinVault reliability report. Design-partner data, synthetic and illustrative of real deliverables.
What you get in week one
Connect the agent. Get the two things it is worst at.
- 1
Point PRISM at your production traffic. No SDK rewrite, no sampling decisions.
- 2
Sessions classify by intent and cluster by failure signature as they arrive.
- 3
Week one ends with a ranked list, a root cause per cluster, and a named owner.
Move from the pattern to the real run
See what users want, which failures repeat, and what deserves attention next.