Deterministic, not probabilistic
Every KPI and every route is formula-driven and reproducible. The same inputs against the same published configuration produce the same result.
Trust and transparency
Fragmented data in. Consistent, replayable decisions out.
AgentIQ is engineered for explainability, governance and deterministic decisioning — so the calls your Customer Success team makes can be explained, reproduced and audited long after they are made.
No black boxes. Just governed intelligence.
Customer Success AI touches revenue
Renewals, escalations and expansion conversations are decisions with money and relationships attached. Ungoverned, an AI system can state things the underlying data does not support, score two similar accounts differently in the same week, and put an action in front of a customer that nobody reviewed.
AgentIQ answers that structurally rather than with reassurance. Authoritative calculation sits outside the AI agent entirely: validated data goes in, versioned deterministic logic produces the score and the route, and a person approves anything that reaches a customer. The agent investigates and explains what the engine already decided.
Four properties, and each one is checkable
Governed context
Validated inputs, versioned Terms and persisted relationship history — so a decision knows what came before it.
Explainable health and risk
A score comes apart into the components that produced it, including the one pulling it down.
Assistive AI, never autonomous
A person approves, modifies or rejects every customer-impacting action, and the choice is recorded.
Deterministic scoring
Every KPI is calculated from an agreed formula on a normalised 0–100 scale, with a full audit trail behind the number.
Explainable by design
Formulas are agreed with you and visible to you in the product. A score traces back to the inputs and the formula that produced it, not to a summary written afterwards.
Versioned and reproducible
Formula changes are version-controlled, so a decision stays tied to the configuration that was in force when it was made.
Structured inputs and versioned Terms
Product usage, support, commercial and engagement data are normalised into governed Terms — defined once, versioned, and meaning the same thing every week.
Consistent definitions
ARR, renewal date, active users and sponsor engagement are defined once in the Terms catalogue rather than re-derived per report.
Schema governance
Versioning is what stops silent drift: a definition cannot change underneath a decision without the version changing with it.
The Data Quality Gate, and context that persists
Before anything is scored or routed, the week's inputs are graded — and the account's history is carried alongside them.
Input validation
Completeness, consistency and freshness are checked first, and every week is graded Ready, Limited or Blocked. Incomplete inputs produce a stated answer rather than a confident-looking wrong one.
Context persistence
Contracts, sponsor relationships and the outcomes of previous decisions are retained, so this week's recommendation is informed by what happened after the last one.
Health and risk you can take apart
A Health score is not useful as a single number. It comes apart into the components that produced it, so the reason an account moved is visible rather than inferred.
Component breakdown
Each component, its weighting and its contribution are shown, so it is clear which measures are driving risk and which are holding steady.
Root cause, not just a rating
The weakest component and the reason it moved are surfaced next to the score, which is what turns a number into something a CSM can act on.
Routing decisions you can replay
Routing is deterministic and threshold-based, and every route is stamped with the rule and configuration version that produced it.
Threshold-based logic
KPI thresholds map to playbooks by rule, so two accounts in the same state this quarter get the same response as one did last quarter.
Full routing history
Every route is recorded and replayable: the state an account was in and the factors that produced the recommendation can both be reconstructed.
Governed agentic execution
The Risk agent is the one agent running in full today. It works from evidence the engine has already validated, and it approves nothing.
Assistive, not autonomous
CSMs keep the judgement and the decision. The agent cannot alter a score, change a threshold, select a different playbook or act on a customer by itself.
Context-grounded outputs
Every brief and plan is written from validated data and recorded history, and names the evidence it rests on.
What it produces to brief you
- Risk brief
- Evidence summary
- The questions it could not answer from the evidence
What it drafts for you to act on
- Selected action plan
- Outreach draft
- Escalation note
- Call agenda
- Preparation checklist
Trust through architecture
Four layers engineered for trust — transparent, replayable, consistent and auditable from raw inputs to agentic execution.
Layer one · Structured inputs
Governed, normalised, versioned Terms.
Layer two · Deterministic KPI engine
Agreed formulas, 0–100 scoring, audit trails.
Layer three · Health and routing
Explainable scores, weakest component, deterministic routing.
Layer four · Risk agent
Governed briefs, plans, drafts and escalations.
One investigation, in full
Investigation record overviewSource evidence, derived signals with confidence levels, the risk hypothesis, the evidence arguing against it, governance stamps and the approval trail — as one record.
Source evidence
Weekly active users down 34% across the last six weeks
Product usage data · 10 Mar 2026Three unresolved P2 tickets on the reporting module, oldest open 19 days
Support data · 09 Mar 2026Renewal date 62 days out; no expansion discussion logged
Commercial data · 01 Mar 2026Executive sponsor has not attended the last two QBRs
Account notes · 24 Feb 2026AI-derived signals
Adoption decline concentrated in one team, not account-wide
High confidence
Support friction is blocking a core reporting workflow
Medium confidence
Sponsor disengagement may indicate an internal ownership change
Low confidence
Risk hypothesis
Renewal is at risk because the reporting workflow that justified the original purchase has degraded, and the sponsor who championed it is no longer engaged.
Medium confidence
Why this hypothesis →Contradictory evidence
Seat count is unchanged and two new users were provisioned last month.
Invoices are paid on time, with no billing disputes on record.
A second team increased usage 18% over the same period.
Governance metadata
Action history
- Investigation created 12 Mar 2026, 08:14
- Evidence collected from 4 sources 12 Mar 2026, 08:14
- Policy applied — routed to renewal-risk playbook 12 Mar 2026, 08:15
- Recommendation generated 12 Mar 2026, 08:15
- Awaiting human approval No action taken until approved
Illustrative. Representative of the shape of a real record rather than a record from a live tenant — the production viewer is not published yet, and the values above are examples.
Evidence first. Approval last.
Within a single investigation, the agent works forward from records it did not choose and stops before the action.
Evidence
Source records are gathered with their system and date attached.
Signals
The agent derives signals from that evidence, each carrying a confidence level.
Hypothesis
A stated hypothesis, together with the evidence that argues against it.
Human approval
Nothing reaches the customer until a person approves it, and the choice is recorded.
The AI agent investigates and explains. Governed logic decides.
Bring us a real Customer Success decision.
The fastest way to judge any of this is to point it at an account you already have an opinion about, in a controlled demonstration, and compare what comes back.