MOHIT Managing Outcomes, Human Intelligence & Talent
AI Capabilities · v1.0
Capabilities Brief

Six AI capabilities working quietly behind your operations.

MOHIT pairs lightweight statistical models with a frontier language model (Claude Sonnet 4.6) to remove busywork from Makers and Checkers — without ever taking the final decision out of human hands.

6
AI capabilities
100%
Decisions logged
Human
In the loop, always
Claude
Sonnet 4.6 · Vision

The six capabilities, with live demonstrations

Each card shows the business outcome, a working demo you can interact with, and a short technical note for your engineering reviewers.
01

Anomaly Detection on KPA Actuals

When a Maker submits a value that's far outside the department's normal range, MOHIT raises a warning before it ever reaches the Checker. Critical anomalies require an explicit override.

Under the hood
$$$ Premium implementation. Full technical details are part of the licensed deployment.
Live · Anomaly threshold Normal
120
Drag to change the current month's reported actual.
02

Smart Auto-Fill of Predicted Actuals

Instead of starting from a blank form, the Maker sees a recommended value drawn from their department's recent history — with a confidence band so they know how much to trust it.

Under the hood
$$$ Premium implementation. Full technical details are part of the licensed deployment.
Live · Maker form Confidence ≈ 80%
Last 6 months
142 156 168
Predicted range: 138 – 172
03

AI-Generated Return Remarks

When a Checker rejects a submission, MOHIT drafts the review comment for them — clear, specific, and respectful in tone. The Checker can edit any word before it reaches the Maker.

Under the hood
$$$ Premium implementation. Full technical details are part of the licensed deployment.
Live · Checker workflow AI
Reason for return
—
Generated draft — editable before send Show fallback
04

Document AI · Evidence Verification

MOHIT reads the actual numbers out of uploaded proofs — PDFs and images — and silently compares them to what the Maker reported. Mismatches above 5% are flagged for the Checker.

Under the hood
$$$ Premium implementation. Full technical details are part of the licensed deployment.
Live · Proof verification No mismatch
Q3 Proof.pdf · page 1
Total customers onboarded: 312
Region: West · Quarter: Q3
Reported by Maker
312
Extracted from proof
312
Values match.
Reported
05

Predictive Deadline-Risk Nudges

Mid-cycle, MOHIT projects whether each department will hit its monthly target. If the projection falls below 90%, the lagging Makers get a polite in-app nudge — long before the deadline slips.

Under the hood
$$$ Premium implementation. Full technical details are part of the licensed deployment.
Live · Mid-cycle projection On track
Day 18 of 30 · 72% complete
Projected month-end: 120%
A "Critical" projection auto-sends a nudge to the lagging Makers.
06

AI Performance Dashboard

Every AI feature is graded weekly: how often suggestions were accepted, how often Checkers overrode the model, where the model was wrong. You always know whether the AI is earning its keep.

Under the hood
$$$ Premium implementation. Full technical details are part of the licensed deployment.
Live · This week Auto-refreshed
Anomaly override rate
4.2%
↓ 0.8 pp w/w
Remark acceptance
88%
↑ 3 pp w/w
Auto-fill kept
71%
↑ 1 pp w/w
Doc extraction success
94%
↑ 2 pp w/w

How a single submission flows through MOHIT

Every AI touchpoint is bracketed by a human decision. AI never commits — it informs.
STEP 01

Maker drafts

Auto-fill suggests a likely value with a confidence band.

→
STEP 02

Submit

Anomaly check runs inline; outliers are flagged.

→
STEP 03

Proof verified

Document AI extracts figures and compares against the report.

→
STEP 04

Checker reviews

If returning, AI drafts the remark; Checker edits and sends.

→
STEP 05

Cycle close

Deadline-risk nudges fire; weekly eval scores every feature.

Trust, governance and the small print

The non-negotiables that make AI safe to deploy inside performance management.
◉

Human-in-the-loop

AI suggests; people decide. Makers and Checkers retain the final say on every value, remark and acceptance.

⎙

Full audit trail

Every suggestion, flag and override is persisted with model version and confidence in AISuggestion and AIFlag tables.

↻

Graceful fallbacks

If the LLM is unreachable, deterministic templates and statistical baselines take over. The product never stops working.

✦

Privacy-respecting

No personal data is sent to the LLM. Only the structured KPA context and a redacted style guide leave the perimeter.

⎈

Measured, not assumed

The built-in AI evaluation dashboard reports acceptance, override and accuracy weekly — so you always know what's earning its keep.

⏻

Switchable

Any capability can be disabled per-tenant via settings. Models, thresholds and prompts are config-driven, not hard-coded.