AZHAR KHAN
INTEGRATE · AUTOMATE · ORCHESTRATE
AGENTIC AI DEPLOYMENT · ENTERPRISE OPERATIONS · SAN FRANCISCO

I let bots agents handle the boring details of the business.

The interesting ones still need judgment.
HOME CARE · 300,000+ HOURS SAVED A YEAR
I find where agentic AI changes the work and agree its value with the people who run it. I lead the redesign and delivery with engineering, architecture and platform teams. I stay accountable for adoption and results in production.
See the work ↓ Curriculum vitae →
MICROSOFT-PUBLISHED · THE FORRESTER WAVE Q1 2021 · UIPATH AI10 PRESS . TO OPEN HOME CARE
SELECTED WORK

5 operations. One discipline.

The work: redesign the operation, lead the deployment, keep it running.

Home Care

One operation, across the business.

In a national home-care company, I connect the functions through reconciled data, with a visible stop before bad counts travel onward.

I redesigned the processes with the people who run them and named an owner for every workflow. Cutover waits for the frontline team’s sign-off.

THE WORK IS READY08:00

Each function kept
its own tools.

Reconciled data
Care

Care logs

Count checkReconciled
Bill items

From the
care logs.

Bill itemsCurated
The same work moves forward
Finance

Invoices
staged.

CountsMatch
The morning’s work is ready
Inspect the count holdOpen the route ↓Close the route ↑
01 · TRIGGERCare logs arrive

The source load starts the downstream billing check.

02 · COUNT CHECKSource and target counts compared

I keep both counts visible before bill items move forward.

03 · FAIL CLOSEDMismatch holds the run

No unreconciled bill items pass to finance.

04 · LOG EVIDENCEException recorded

The record carries the failed count check and the source load.

05 · NAMED OWNEROperations owner receives it

I route the held run and its evidence to the named operations owner.

06 · RECONCILECorrected counts checked

The hold stays until the corrected source and target counts agree.

07 · CONTINUEWork resumes on reconciled data

Bill items and invoices move forward after the check passes.

A failed write follows a separate bounded retry. Neither a timer nor retry clears the count hold. Inspect the illustrative operations replay →

26Mjobs across the program

Program measures and technology

99% straight through · 11 in 10,000 reach a person · $0.03 per outcome. An agent routes ~90% of 32K IT tickets. 9 acquisitions absorbed as configuration.

Workato · Celonis · Power Platform · Datadog

Fintech

129 signals,
one decision.

I led the redesign from twelve source screens to one case record and directed the deployment in a payments company’s fraud operation.

Agents gather the evidence, and the investigator approves, corrects or escalates the regulated decision. The deployment runs in production.

Core bankingAccount •••• 4471 · CurrentKYC recordsPassport · verifiedDevice intelligencePhone · unseen / SIM changedTransaction historyVelocity · 10× the 30-day meanSanctions screeningWatchlist · no matchFraud detectionAuthority · recommend onlyCase managementReviewed by · investigatorNetwork graphShared device · 2 accounts
CASE 4471Investigator decides

Account takeover

DevicePhone · unseen
SIM statusChanged · 3 d
WatchlistNo match
Model authorityRecommend only
Evidence assembled129 signals

9:00 → 4:00per account-takeover case

Explore Fintech ↗Inspect the engine map ↗

The sources remain visible beside the assembled record.

Program measures and technology

660,000 cases a month · Every regulated decision reviewed · Overall handling time 60–65% less.

LangGraph · Vertex AI · Glean

Utilities

A back office became a buying desk.

In a utility bill-pay operation, I redesigned the work so every invoice is reconciled against its account first.

A previous balance that matches a payment in transit closes on the record.

The $2.5B payments operation has $1B running on its process-intelligence digital twin. It handles ~700,000 invoices a month; 1 in 100 reaches a person.

UTILITY BILLAccount 4471-0382

Statement · March

Previous balance$1,284.10
Payments received$0.00
Energy charge$1,116.00
Delivery$212.40
Taxes and fees$68.90
Amount due$2,681.40

$0.58 → $0.20per invoice

Program measures and technology

$1B of $2.5B payments on the twin · 700,000 invoices a month · capture accuracy 65% to 99% · 1 in 100 reaches a person · onboarding 90 to 30 days.

Power Platform · Azure · Document Intelligence · Published case study

2 EARLIER OPERATIONS
UK BANK · FRAUD
Sixty-two touchpoints became three.
At a UK bank, I led discovery on 8 fraud journeys and 40 systems before any automation.
13 → 5case steps UK BANK →
MANUFACTURER
I cut the rebate cycle across 4 countries.
I directed the work at a global manufacturer from 2020 to 2023, and the practice still runs it in production.
returned$6.5M+
90 → 2days MANUFACTURER →
Inspect the control plane and execution recordIdentity, decision rights, exception routing and one day of execution.
RUNNING NOW · HOME CARE
99% straight through, for 5 years.

Agents take the repeatable work on the control plane, and a person approves, corrects or escalates where the decision needs one.

CONTROL PLANE IDENTITY GUARDRAILS OBSERVABILITY COST PER OUTCOME AUDIT TRAIL AGENT PERSON WHERE decision = regulated → PERSON
99%straight through
$0.03per outcome
11 in 10,000reach a person
RUN · 5 OPERATIONS

5 operations run in production.

Every run is traced, and every exception goes to a named owner.

SUM runs
GROUP BY value_stream
AVG duration_sec
WHERE reached_person = 1 → PERSON
WHERE status = failed
ONE DAY ON THE HOME-CARE PLANE

One day on the home-care plane, by hour: runs started, runs that reached a person and runs that failed, each line scaled to its own peak. Since the ticket queue began, 11 in 10,000 jobs reached a person.

The orchestration is the job.

26M JOBS · 99% STRAIGHT THROUGH

A home-care company has run on one plane for 5 years. 11 in 10,000 reach a person.

660,000 CASES A MONTH

The engine builds an account-takeover case from 8 systems in 4:00, down from 9:00. A person approves, corrects or escalates every regulated decision.

700,000 INVOICES A MONTH

$1B of a $2.5B payments operation runs on the digital twin. 1 invoice in 100 reaches a person.

HOW I WORK
Inspect how I select a process →
No process, no agents.
My deployments run on these 7 steps, from discovery to production.
In each of the 5 operations, I redesigned the workflow and set its guardrails before any agent took it.
BUILD

Agents start from a redesigned operating model.

RUN

Agents and people run on one plane, with a cost per outcome.

GOVERN

Every action carries an identity, guardrails and an audit trail, with a person approving, correcting or escalating at each gate.

DISCOVER · 1
Discover
I work with stakeholders and the people running the operation to clarify the need, map the process and agree what improvement is worth pursuing. Process intelligence first. Every step is measured and priced before an agent takes it.
70+ systems read in a 5-day discovery sprint.~40% of nominated use cases cut at review.
DECIDE · 2
Redesign the operating model
I work with process owners to redesign the work, data flow and responsibilities. Agents take repeatable work; people keep consequential decisions, with named owners for the workflow and its operation.
Every regulated decision is approved, corrected or escalated by a person.
DECIDE · 3
Decide the tooling
I decide build versus buy with the client, on live volume, tool by tool.
Workato, Power Platform, Appian and Azure AI, each where it won.
BUILD · 4
Stand up the orchestration
With engineers, architects and platform teams, I work through data, integration, evaluation and failure behavior. Engineers implement the systems; I lead delivery and agree readiness and the quality bar with the client and teams.
8 systems, one record per case.
GOVERN · 5
Identity and guardrails
Every agent has an identity and scoped access. Every action has a fail-closed path, and every regulated decision goes to a person to approve, correct or escalate.
At Fintech, only name and card changes close themselves.
GOVERN · 6
Observe and route
The system traces every run, step by step. Every failure has a code and a route to the team that owns it.
1 error topic, 1 consumer, 108 routes.
RUN · 7
Run it on cost per outcome
I lead change and adoption with the client throughout delivery and into production. I direct the production team with a named executive sponsor; cost per outcome and versioned corrections keep the operation accountable.
$0.03 per outcome.
AGENT GOVERNANCE
IDENTITY·GUARDRAILS·PERSON IN THE LOOP·OBSERVABILITY·EVALUATION·FINOPS·AUDIT TRAIL
DECISION RIGHTS

I draw the decision rights before any agent ships.

STRATEGIC INITIATIVE QUICK WIN DEPRIORITIZE BACKLOG 01 · Rate-increase letter release
ASSIST · QUICK WIN
A person reviews
The system builds the case. A person approves, rejects, or sends it back.
~40% OF NOMINATED USE CASES CUT AT REVIEW · 5-DAY SPRINT
A CLOSER LOOK AT THE METHOD

How I govern AI changes.

Illustrative method · Case 4471

Runtime trace
I keep the model's recommendation inside the case boundary.

Process intelligence supplies the context for Case 4471. The model proposes a hold; policy permits a recommendation only. Only name and card changes close themselves. I trace the sources, rule, recommendation and investigator's disposition.

Acceptance cases
I test a candidate before it reaches a case.

For Case 4471, acceptance cases define when the candidate may pass, stop or escalate. A score alone cannot authorize a release or a regulated decision.

Human correction
I keep the investigator's correction in the loop.

The investigator approves, corrects and sends back, or escalates Case 4471. I record the reason and version the correction before it changes a rule, route or model.

Change boundary
I release a tested version with a way back.

The named owner approves the candidate after acceptance cases pass. I record the rollout and restore the prior approved version if its behavior fails the criteria.

THE RECORD

$100M+ measured across all programs since 2016.

I ask where the check goes, what fails closed, who gets the ticket and with what evidence, and what an outcome costs.

US telecom carrier
Stood up automation governance, intake and delivery. Invoicing errors down 85%, processing time down 90%.
2016 – 2018
Fortune 20 energy
70+ automations. 200,000 hours returned, $20M saved.
2018 – 2019
Top US telecom
Set up how the operation runs automation. 250 automations, $18M returned, $988K a year from one redesign.
2019 – 2021
The practice
Grew the Center of Excellence from 4 to 30+ across the US, Mexico and India. Discovery runs as a 5-day sprint reading 70+ systems.
2019 – present
Governed operations
26M jobs at 99% straight through over 5 years, 9 acquisitions absorbed as configuration, no rewrite since 2021.
2021 – present
EXPERIENCE MICROSOFT-PUBLISHED · FORRESTER 2021 · UIPATH AI10
Concentrix Agentic AI Deployment Lead. Built the Workato partnership, with co-selling and joint proofs of concept, and partners with Cursor on AI for software delivery. 2019 – present
UiPath Automation Delivery Lead. 2018 – 2019
ProKarma Automation Program Lead. 2016 – 2018
I hire the team that deploys the agents.
THE PRACTICE
Enterprise readiness is agentic readiness.
Agentic orchestrationProcess miningDocument intelligenceEvalsDecision rightsCost per outcomeEnterprise search · RAGObservabilityIntegration · iPaaSData warehouseRobotic process automation
I lead agentic AI delivery and production operations, with people at consequential decisions.
CONTACT
Let's talk about a deployment leadership role.