Omni-Channel Personalized Concierge · Regulated Industries
A governed AI concierge that meets every customer where they are
I led the vision and the coalition behind AssistIQ — an omni-channel concierge that uses a semantic layer and a context engine to turn every interaction into proactive, personalized, and actionable help. It serves wildly different personas across three of the most regulated industries, on a governed, multi-tenant platform built for rapid experimentation.
One customer · any channel
Semantic layer & context engine
One customer. One context. Everywhere.
Customer context & memory
Governed orchestration
Enterprise systems of record
“Meet customers where they are.”
The real problem
A support problem on the surface — a strategy problem underneath
Support at WEX spanned wildly different customers — benefit administrators, HSA consumers, fleet managers, truck drivers — across healthcare, transportation, and financial services. It was treated as a cost center and an afterthought: fragmented across IVRs, apps, and contact centers, reactive by design, and identical for everyone despite serving no two customers the same way.
The obvious read was a CSAT problem. The real risk was strategic. Generative AI was about to reset customer expectations across every one of those industries, and the company had no shared, governed way to ship AI experiences. Every team that wanted to use AI was starting from zero — re-solving identity, integration, evaluation, and compliance — in regulated environments where a single ungoverned model response could create a real problem at the scale of tens of millions of customers.
So I reframed the question for leadership. Not “how do we build a better chatbot,” but “how does this company earn the ability to ship trustworthy, personalized AI — repeatedly, across regulated industries — without re-litigating governance every time?” Support was the urgent, fundable problem. The platform was the durable one.
The strategic bet
From a chatbot to a capability
- Reframed support into a personalized concierge — proactive where it can be, actionable always — with the support experience as the first tenant of a governed AI platform.
- Anchored on a semantic layer and a context engine that turn each channel’s signals into bi-directional insight: the platform learns from every interaction and feeds personalized, proactive help back out across channels.
- Designed for action, not static information — every persona can complete the task (file, approve, resolve) in place, instead of just reading about how to.
- Built for many tenants and many personas from day one, so one governed foundation could serve totally different journeys without being rebuilt each time.
- Treated governance and security as first-class product surfaces — designed in, not bolted on — because the experience lives in three of the most regulated industries there are.
Who it serves
One platform, radically different people
One platform, radically different people. AssistIQ serves personas whose needs barely overlap — and gives each a journey that feels built just for them.
Benefit Administrator
Manages plans and members; needs fast, accurate answers and bulk actions across complex benefit rules.
HSA Consumer
An everyday member; needs plain-language help and to actually complete a task — check a balance, submit a claim.
Fleet Manager
Oversees vehicles, cards, and drivers; needs operational control and exceptions handled without a phone queue.
Truck Driver
On the road; needs voice-first, hands-free help that resolves the issue in the moment.
Different challenges, journeys, and channels — all customized on one foundation. That is what “meet customers where they are” actually requires.
Semantic layer & context engine
What makes it a concierge, not a chatbot
The difference between a chatbot and a concierge is meaning and memory. AssistIQ’s semantic layer and context engine make every interaction smarter than the last.
Bi-directional insight
Each channel feeds the context engine, and the engine feeds personalized, proactive guidance back to every channel.
Proactive, not just reactive
It surfaces the next best action before the customer has to ask for it.
Actionable, not static
Insights arrive attached to a workflow the customer can complete in place.
Multi-intent & context switching
It follows a customer who changes topics mid-conversation, routing intelligently across skills and channels.
Executive alignment & stakeholder management
Eight functions, one shared definition of done
A platform bet only works if the organization agrees to share — and a concierge that touches regulated data only ships if the whole company trusts it. The hardest work was never the model; it was aligning a coalition of 15+ stakeholders, each with a legitimate and conflicting definition of “done,” around a single vision.
Executive Leadership
Cared about: ROI, strategic fit, and risk exposure of GenAI.
How I aligned them: Reframed the spend as a capability investment with a staged business case, and tied each phase to a visible outcome so the platform had to keep earning its funding.
Legal & Compliance
Cared about: Regulatory exposure across healthcare, transportation & financial services.
How I aligned them: Brought them in at the design stage and turned their constraints into product requirements — policy enforcement, auditability, and disclosure as platform features.
Data & AI
Cared about: Model quality, data access, and avoiding ten teams re-solving the same thing.
How I aligned them: Gave them a shared semantic layer and evaluation standards, so AI work compounded on one foundation instead of fragmenting.
Enterprise Technology & Architecture
Cared about: Maintainability, multi-tenant integrity, and not creating another silo.
How I aligned them: Designed the platform as reference architecture with explicit extension points, and let architecture co-own the standards.
Partner Services
Cared about: White-label partners’ needs, brand, and trust.
How I aligned them: Built a multi-tenant design so each partner inherits a governed, customizable baseline rather than a one-size-fits-all bot.
Marketing & Sales
Cared about: Brand voice, the customer promise, and a growth story.
How I aligned them: Aligned the concierge experience to the brand and made personalization a differentiator they could take to market.
Security
Cared about: PII, identity, and the prompt/abuse surface.
How I aligned them: Made security the platform’s first feature — PII masking and identity-aware access reviewed once, at the platform, not per app.
I built a coalition across 15+ stakeholders — Legal, Compliance, Partner Services, Marketing, Sales, Data, AI, and Enterprise Technology — and got them aligned on one shared vision. That shared ownership is what turned a risky AI bet into an initiative the whole company carried.
Build vs. buy
Earning the decision, not defaulting to it
With credible options in the market, “build” was not a default — it was a decision I had to earn. I ran a structured evaluation, with multiple proofs of concept against a shared scorecard, so the company chose with its eyes open. (Specific vendors are kept anonymous; the methodology is the point.)
Decision criteria
- Scalability to tens of millions of customers
- Governance & auditability across three regulated industries
- Security & identity model
- Extensibility to new personas, partners & domains
- Total cost of ownership at scale
- Time to market
Options evaluated & what each was tested for
- Turnkey support-agent platformsSpeed-to-value and answer quality — weighed against depth of governance control and extensibility beyond support.
- Incumbent CRM / service suitesDeep integration and fit with the existing estate — weighed against lock-in and cost at our scale.
- Cloud-native building blocksComposability and control — weighed against build effort and time-to-market.
- Emerging enterprise AI providersScalability, security, and extensibility — pressure-tested against the regulated, multi-tenant, multi-persona bar.
The decision
The answer was a hybrid, not a binary: buy and borrow at the edges — best-of-breed components and cloud primitives — and build the connective tissue we could not outsource: the governance layer, the evaluation harness, the semantic/context engine, identity-aware orchestration, and a library of reusable skills and tools. Every turnkey vendor optimized for a great support bot; none gave us a governed, multi-tenant foundation we could point at the next persona, partner, or domain without re-buying. Owning the foundation — not the whole stack — was the lowest-TCO path to repeatable, compliant, personalized AI.
Platform foundations
Built as infrastructure, not a one-off
AssistIQ was designed, from the first spec, as infrastructure other teams could stand on — not a standalone support app. The support experience was the platform’s first tenant, deliberately built to make the next persona, partner, and domain dramatically cheaper to serve.
Semantic layer & context engine
Shared meaning and memory across channels, so personalization compounds with every interaction.
Skills
Reusable, composable capabilities (look up a claim, explain a benefit) that any future assistant can call.
Tools
Governed connectors to enterprise systems — identity, card, claims, benefits, knowledge — with permissioning built in.
Agent registry
A catalog of governed agents and skills teams can discover and reuse instead of rebuilding.
AI Gateway
A single governed entry point to models, with guardrails, PII masking, and observability applied centrally.
Specifications
A shared contract for an assistant’s scope, behavior, and guardrails — new experiences are configured, not re-engineered.
Evaluation frameworks
Automated quality and safety gates — every assistant is graded before it ships and monitored after.
Governance & policy layer
Declarative, auditable policy on what an agent may say, see, and do — decoupled from application code, with guardrails enforced centrally.
Agent orchestration
Multi-intent routing and context switching across skills, channels, and confidence-aware human handoff.
Multi-tenant architecture
One foundation serving many partners and personas — each isolated, governed, and customized.
The real test of a platform is the second use case. These primitives are why a new persona or partner at WEX starts from a governed baseline instead of a blank page.
Trust, security & governance
Security is the first feature, not the last
AssistIQ operates in three of the most regulated industries there are. So I led a massive undertaking to build the governance standards this initiative runs on — and made security the platform’s first feature, not its last. Security is the number-one requirement, by design.
PII masking
Sensitive data is masked before it ever reaches a model.
Bias detection
Outputs are monitored for bias, not assumed to be fair.
LLM-as-a-judge
Automated evaluation grades responses for quality and safety at scale.
Observability (Datadog)
Real-time dashboards on behavior, drift, and platform health.
Automated fraud checks
Fraud signals are caught inside the workflow, not after the fact.
Guardrails by default
Every agent inherits the right guardrails via shared policy — compliant by construction.
Multi-intent, context switching, and routing all happen inside these guardrails — so the experience stays fluid for the customer without ever stepping outside what’s compliant.
Key product decisions
The tradeoffs behind the build
| Decision | Alternative considered | Tradeoff | Outcome |
|---|---|---|---|
| Build a platform, not a point solution | Ship a standalone support bot fast | Slower first release; a harder executive sell | Reusable, governed foundations many personas now share |
| Governance standards as a company-wide undertaking | Let each team self-certify compliance | Heavier up-front coordination | Compliant-by-construction across 3 regulated industries |
| Multi-tenant from day one | Single-tenant now, replicate later | More architectural complexity early | New partners & personas onboard without a rebuild |
| Semantic layer / context engine | Per-channel context, kept siloed | A shared abstraction to build & maintain | Personalization compounds across every channel |
| Hybrid build-vs-buy | Adopt a single turnkey vendor agent | More build effort on the core | Owned roadmap & TCO; no per-domain re-buy |
| Security as the first feature | Add masking & monitoring after launch | Upfront design cost | PII masking, fraud checks & bias detection built in |
| Action over static answers | Surface information and link out | Deeper system integration required | Customers complete the task, not just read about it |
Impact
Separating company scale from platform impact
Two different things get called “impact.” Below I separate the scale of the environment the platform operates in from the outcomes the platform is actually accountable for.
Company scale · context
Product · platform-attributable
- A personalized, context-aware concierge across mobile, web, voice, and partner portals — one customer, one context, consistent assistance everywhere.
- Proactive, actionable help: customers complete transactions in place, with a human one confidence-threshold away.
Operational · platform-attributable
- ~30% reduction in routine support call volume in the targeted flows.
- Live agents refocused from repetitive requests onto complex, high-value cases.
Organizational · platform-attributable
- Reusable, multi-tenant foundations — semantic layer, skills, tools, evaluation, policy — that let new personas and partners launch from a governed baseline.
- Authored the governance standards now relied on across the company’s regulated AI work.
Strategic · platform-attributable
- Established the company’s shared answer for shipping trustworthy, personalized AI in regulated industries.
- Enabled a culture of rapid experimentation, learning, and iteration — safely, inside the guardrails.
Lessons learned
What I carry into the next AI platform
Get them aligned on one vision, then let them run
With 15+ stakeholders from Legal to Sales, the real deliverable before any feature was a shared vision. Once the coalition owned the same picture, I could enable a culture of rapid experimentation instead of policing every decision — alignment is what lets you move fast in a regulated environment.
Govern first, and governance becomes a feature
Teams expect compliance to slow them down. Designing PII masking, policy, bias detection, and evaluation as platform primitives turned governance into the thing that let us move faster and launch safely — precisely because we were under regulatory scrutiny in three industries at once.
In the enterprise, distribution beats the demo
The model was the easy 20%. The value was unlocked by identity, integrations, the semantic layer, and the right to act — and by an organization willing to adopt it. I now scope AI products from the adoption surface inward, not from the model outward.
Personalization is an architecture decision, not a prompt
Serving a truck driver and a benefit administrator from one platform only works if meaning and memory are shared infrastructure. The context engine — not clever wording — is what made the concierge feel built for each person.
Build-vs-buy is a portfolio decision, not a binary
The right answer was buy at the edges and build the connective tissue. Owning the governed, multi-tenant core — rather than the whole stack — was both cheaper and more defensible than either extreme.
Stack
Happy to walk through the real specifics, the tradeoffs, and my specific role in a conversation.
Get in touch