Fish Innovation Group
Ben Fish · AI Strategy, Roadmaps and Builds

The roadmap, the build,
and the number it returned.

AI Consultant and Builder · 20+ Years · Fortune 50 to Owner-Operated

Most companies know AI matters and cannot say where it pays. I come in, interview the business end to end, and come back with a ranked roadmap: what to build, what to buy, what to skip, what each one returns, and in what order to do it. Then I build the ones that matter. Everything below is a system designed, delivered, or still operated by me.

See what I've built How an engagement works
$875M+
Net income before taxes driven
32
AI and automation systems shipped
30+
Client engagements
9
Industries built in
The Engagement

Find the AI worth building. Then build it.

The failure mode is the same everywhere. Everyone starts building, nobody decides what is worth building, and a year later there is a lot of spend and no number anyone can point to. This sequence exists to prevent that.

Step One

Assess

  • Interviews across the C-suite, the operators, and the people actually doing the work
  • Every recurring task inventoried and scored for whether AI genuinely helps
  • What people say and what the data says are usually different. Both get captured
  • Three days to two weeks, depending on size
Step Two

Roadmap

  • A ranked list: what to build, what to buy off the shelf, what to skip entirely
  • The return on each one, and when it lands
  • Sequenced by payback and by integration readiness, not by what sounds impressive
  • Some items rank low purely because the integration is not there yet. Fighting that is wasted budget
Step Three

Build

  • Start with the two or three with the clearest payoff, not all twenty
  • Scoped to a real budget and a real MVP, not a wish list
  • Delivered in weeks, not quarters
  • You own the system outright when it ships
Step Four

Measure

  • Return tracked against a baseline set before the build, not estimated after
  • Reported as KPIs an executive team can actually govern
  • A published kill list for whatever is not working
  • Then the next item on the roadmap

What the assessment is actually for

A large consultancy will hand you sixty-two findings and a template. I have been brought in after that engagement more than once, and most of the list does not survive contact with the business.

The point is not to generate ideas. It is to find the small number of things genuinely worth doing at this company, in this order, at a cost that makes the return obvious. Sometimes the honest answer is that you do not need anything built at all. You need three off-the-shelf tools and someone to teach your team to use them properly.

Cost Governance

Most AI budgets break the same way.

Everyone builds, nobody meters, and the line item outruns the governance. AI spend becomes the fastest-growing item in the IT budget with no oversight attached, because saying yes to AI is politically easy and saying no to it is not. Cost discipline belongs in the design of a system, not in a cleanup exercise afterward.

01

Model tier matched to task

Frontier models only where they earn it. Most production work does not need the most expensive model available, and the difference compounds on every single call.

02

Inference economics tracked per workload

Every workload carries a known cost per unit of work. If you cannot say what a task costs to run, you cannot say whether it is worth running.

03

Metered spend consolidated out

Every call site in a running system gets inventoried and migrated off metered per-call pricing onto a path with no marginal cost. Then the migration gets proven by confirming the old billing path is genuinely dead.

04

Kill switches that fail off

Anything that spends money or touches a customer has a global switch defaulting to off. Runaway spend is a design failure, not an accident.

The cheapest fix is usually not a system

When spend runs away, the instinct is to buy a tool that governs it. Usually the root cause is that nobody taught anyone how to use what they already have. An untrained person burns tokens like nothing else, and no dashboard fixes that.

One engagement took a fifteen-person team spanning sales, operations and marketing, assessed where their gaps actually were, and taught them where AI mattered and where it did not. Their product sales rose more than 24%, adoption of what they were selling rose more than 250%, and their operating spend came down. The build budget was the smallest part of it.

The Build Bank

What I have built, and what it returned.

Grouped by capability rather than by client, because the useful question is not whether I have worked in your industry. It is whether your problem belongs to a pattern I have solved before. Client and employer names are withheld, and nothing here carries a date. What matters is that these systems exist and that I can build them now.

01

Voice AI in production

Inbound Voice
6.1% to 19.4% booking rate

24/7 inbound voice agent

A bilingual inbound agent using LLM-driven language understanding, intent classification and scheduling automation to capture demand around the clock and take after-hours load off staff.

860 calls across a five-month deployment
Outbound Voice
$100K+ per month, per site

Outbound reactivation agent

An outbound agent that autonomously works dormant backlogs and recovers revenue the business had already written off, running voice and SMS against a deterministic booking spine.

40% conversion on revenue already written off
Home Services
 

Booking and estimates agent

Voice AI deployed well outside healthcare, handling appointment booking and delivering quotes and estimates to homeowners at the moment of interest instead of making them wait for a callback.

The same voice stack, in an entirely different sales and regulatory context
Fleet Architecture
 

Multi-tenant agent factory

An agent fleet built from version-controlled templates rather than configured by hand one at a time. Every new build auto-provisions its own agents in two languages, and a fleet-wide fix propagates through a single parameterized script.

1,100+ live calls handled, one fix propagating across every template and agent
Voice Engineering
 

Deterministic booking spine

A two-lane conversation architecture separating referring callers from end customers over a deterministic booking core, so the model handles language and the system handles anything that must never be improvised.

Restructured from 22 nodes to 17, every structural guard passing
Platform Depth
 

Vendor platform defect, diagnosed and solved

Root-caused a genuine bug in a commercial voice platform where interrupted end-of-call speech silently dropped the hangup signal, then engineered a reliable termination pattern now reused across the whole fleet.

The kind of problem that has no documentation and no support ticket
02

AI quality, evaluation and monitoring

Portable System
 

Autonomous voice QA

A quality system that watches a voice agent fleet on its own, combining deterministic checks, fleet-wide audits, LLM-judge evaluation and connection-layer monitoring. It is not coupled to any one industry and runs against any voice agent fleet.

16 detection rules, surfacing real production failures nobody had reported
Testing
 

AI that tests AI

A simulation harness driving a live conversational flow scenario by scenario with a synthetic customer, and a second model judging the outcome. Regression testing for systems whose output is different every time.

15-scenario automated suite plus a 28-case human validation pass
Pre-Deploy Review
 

Adversarial multi-agent review

Every significant deployment gets an independent red-team pass run by multiple agents across different review lenses, each finding verified separately.

Panels of 5 to 60 reviewers, catching blockers before they reach production
03

Prediction, pricing and forecasting

Churn and Winback
$80M a month in saves

Deposit churn prediction

Machine learning on behavioral product interaction that flagged customers months before they closed their accounts. The strongest single signal: when roughly 92% of a given product went unused or changed drastically, the account closed about four months later. That indicator drove designed winback campaigns.

2% save rate, churn flagged three to four months ahead
Pricing Models
$400K in repairs avoided

Pricing and walk-away models

Regression models telling buyers and agents what to offer and, more valuably, which deals to walk away from. The model priced in inventory pressure and how people actually behave under it, not just comparable sales.

A home listed $200K over asking. The model said no, and the inspection proved it right
Merchandising
Sales up 17%

Assortment and floor-layout optimization

Analytics on sales data and customer buying behavior determining which items to push, feeding a redesigned store layout built for faster purchasing and better service on the floor.

At national retail scale
Computer Vision
 

Expert-grade image scoring

Framed and baselined a supervised vision problem, deliberately proving a workable baseline with no GPU before anyone spent money on infrastructure.

Within one expert grade 91% of the time, on a 4,313-item dataset
Decision Support
 

Grounded data-to-model pipeline

Databases wired directly into models so the model answers from live business data rather than from whatever it happens to remember. Built well before this was a common pattern, and the direct ancestor of the operating system below.

An early ancestor of the operating system below
04

Enterprise platforms and customer engagement

Enterprise Platform
$200M+ per year

AI customer engagement platform

Designed and delivered an enterprise AI engagement platform, leading the 75-person cross-functional team that built and scaled it, along with the data governance models and AI decision frameworks adopted across Digital, Operations and Technology.

Top-box CSAT from 54% to 82%, built by a 75-person team
Contact Center
$20M spend reduction

Contact-center intelligence and IVR

Machine learning driving the predictions behind IVR routing and program design, plus the analytics engines running quality monitoring and customer satisfaction measurement across the operation.

Prediction-driven routing at enterprise call volume
Sales Enablement
Sales up 38%

AI-enhanced sales and service workflows

Enterprise product roadmaps with embedded machine learning on governed data, putting better next actions in front of the people actually closing business.

Portfolio performance up 18%, with up to 29% KPI lift
05

Industrial and scientific AI

Manufacturing QA
8 vials recovered per run

Non-destructive batch quality analysis

An infrared AI system detecting molecules and impurities during the batch process without destroying the sample. Quality assurance on biologics had meant sacrificing vials worth thousands of dollars each, several runs a day, every working day.

Biologics worth thousands of dollars a vial, on every run, every working day
Process Optimization
2 hours a day recovered

Purification process redesign

A better system built inside the molecule handling process, good enough that research and manufacturing operations bought it for standard use.

Adopted as standard across research and manufacturing
R&D Automation
10x research efficiency

Global research automation platform

Automated research workflows, data pipelines and an enterprise ordering platform used across global scientific operations.

5,000+ users across global scientific operations
06

Responsible AI, safety and compliance

Operating Model
 

Responsible-AI operating model

Not a policy document. Human-in-the-loop encoded in code on every irreversible action, adversarial review before anything ships, model-tier governance, and money paths kill-switched and gate-tested before they are ever armed.

Human-in-the-loop written into code, not into policy
Safety Engineering
 

Safety engine on a live revenue path

A precision-calibrated safety system sitting on a revenue path, built to a specific federal regulation rather than papered over afterward, then generalized across nine industry verticals plus a safe default without weakening its protections.

Zero false positives on 50 real cases, 6 of 6 seeded violations caught
Regulated Data
 

Protected-data architecture

Layered scrubbing of protected information at multiple independent points, API surfaces that never carry it at all, mandatory human review before any of it reaches a document, and HIPAA-compliant, BAA-backed flows with SOC 2-aligned practices around them.

Compliance designed into the architecture, not bolted on afterward
07

Data engineering and integration

Payload-Level EDI
 

Clearinghouse integration

Claims and remittance integration built at the raw transaction-set level rather than through a convenience wrapper, covering the full claim, remittance, eligibility and status message set, with automated monitoring on payer enrollments.

Full claim, remittance, eligibility and status message set
Migration
 

Records synchronization at scale

Onboarded a live multi-provider institution onto a new platform: tens of thousands of claims, thousands of patients, years of historical remittance and dozens of providers, verified against real production data.

40,200 claims and 11,151 patients, verified against real production data
Reliability
 

Zero-downtime deploys to live money systems

Back up, migrate off the serving path, verify behind a gate, retag to roll back. Repeatedly deployed to databases handling live payments with no downtime and no incident.

Real-money gate tests fired, verified and refunded before arming
08

Operations automation and AI-native transformation

Whole-Business
 

An AI operating system for a real business

A layered stack priming an AI assistant on live business metrics, running a fleet of operational automations, and reporting against KPIs an executive team can govern. Every recurring task was audited and scored for automatability, then the majority automated from a zero baseline.

60 tasks across nine functional areas, tracked from a zero baseline
Enablement
Product adoption up 250%

AI enablement for an operations team

A fifteen-person team spanning sales, operations and marketing that was losing digital product sales and overspending to do it. Assessed the real gaps, built the strategy, and taught them where AI gave genuine leverage and where it did not.

Sales up 24%+ and operating spend down, from a fifteen-person team
Demand Generation
 

Automated outreach and AI presenter video

Social engagement and posting run by bot to widen reach and generate inbound interest, replacing the manual work of monitoring and responding. Separately, AI twin presenter video produced at volume for agents who needed constant social presence and had no time to film.

Manual monitoring and posting replaced entirely
Revenue Operations
 

Contract-to-cash automation

An automated chain running discovery through contract, agreement, payment and provisioning with no manual handoff, plus a multi-tier channel program carrying automated agreements, onboarding, commission logic, attribution windows and payouts.

Discovery through contract, payment and provisioning with no manual handoff
Incident Response
 

Forensic recovery of failed transactions

Recovered paying customers whose payments had been captured but whose fulfillment silently failed, by re-signing and replaying the original events with fresh signatures. Also root-caused multi-day silent outages traced to a misconfigured provider fallback and an unpinned dependency.

Multi-day silent outages root-caused and hardened
Delivery Speed
 

Systems shipped in a single session

Complete working systems designed, built, deployed and validated in one day when the situation called for it, including a lead-capture application deployed live to a conference booth and a multi-phase quality platform.

An 88-plan engineering program sustained over four months
Range

Nine industries, one method.

I am not a domain specialist, and that is deliberate. AI capability transfers between industries. What does not transfer is knowing which problems in a given business are actually worth solving, and that judgment only comes from having done this in businesses that look nothing like each other.

Banking and financial services
Healthcare
Pharmaceutical and biologics manufacturing
Retail
Real estate
Home services
SaaS
E-commerce
Local services and reputation

Fortune 50 enterprises

Global banking, global pharmaceutical, national retail. Regulated, matrixed and slow to move, which is exactly where governance and sequencing decide whether anything ships at all.

Mid-market and community

Regional and community banks, multi-site operators, professional practices. Enough scale for AI to pay back quickly, not enough internal capacity to build it alone.

Startups and owner-operated

Venture-backed startups and businesses run by the person who owns them, where the roadmap has to fit a real budget and return inside the same fiscal year.

How I Work

The part that survives contact with your IT organization.

Most AI consultants arrive with a wish list and leave behind a parallel stack, an unbudgeted spend curve, and a security review nobody scheduled. I run production systems on my own signature, so I hold the same constraints your technology leadership does.

01

Borrow before you build

Roughly 80% proven platform, 20% custom where it creates real advantage. Integration first, no shadow stack, and a documented rationale for every build-versus-buy call. If you can buy it for fifty dollars a month, buy it.

02

Spend governed from day one

AI cost is a design constraint, not a quarterly surprise. Model tier matched to task, economics tracked per workload, metered cost consolidated out wherever a cheaper path does the same job.

03

Nothing irreversible without a human

Human-in-the-loop lives in code, not in policy. AI drafts and prepares. People hold the trigger on anything customer-facing or money-moving, and kill switches default to off.

04

Adversarial review before anything ships

Every significant deployment gets an independent multi-agent red-team pass across multiple review lenses. It has caught real blockers before release more than once.

05

Security and compliance by architecture

Regulated data flows designed to the rule rather than papered over afterward: layered scrubbing, API surfaces that never carry protected data, SOC 2-aligned practices, and compliance built into the design.

06

Change management with a rollback path

Dry-run by default, migrations off the serving path, snapshot rollbacks, gated verification. Deployed into live money databases with zero downtime and no drama.

07

Vendor decisions as findings, not verdicts

Years spent running technology partnerships and vendor selection means knowing what is real and what is a demo. Evaluations arrive as evidence and open questions, because the people holding the non-public context should make the call.

08

You own what gets built

The system is yours outright when it ships. Return is tracked against a baseline set beforehand, so you can see whether it worked instead of taking my word for it.

Background

Twenty years in, and still building.

A hard-science researcher who moved into banking product and strategy, then into AI as a founder. The through-line is the same in all three: find the mechanism, prove it with data, then build the thing.

Enterprise AI strategy
Generative AI and agents
Machine learning and forecasting
Responsible-AI governance
AI cost governance
Product and platform
Vendor and partnership strategy
Change management and adoption
Executive and board communication
  • Banking executiveEnterprise AI, analytics and product strategy inside large regulated financial institutions, including platform delivery at scale and governance frameworks adopted across Digital, Operations and Technology.
  • Founder and operatorBuilding and running AI systems in production today, on my own signature, under real regulatory constraint. Not advising on AI from a distance.
  • Research scientistSynthetic chemistry and drug development. Collapsed a 25-step synthetic route to 3 steps, yielding a candidate that reached Phase 2 trials for Alzheimer's disease.
  • EducationMIT Artificial Intelligence Program (AI/ML Roadmap and Strategy) · MBA, University of Connecticut School of Business · MS, Organic Chemistry, UCLA · BS, Chemistry (Honors, Cum Laude), University of Arizona

Start with the assessment.

The fastest way to find out whether there is anything here worth doing is a short conversation about where your business actually loses time and money. If the answer is that you need three off-the-shelf tools and some training, I will tell you that.

ben@fishinnovationgroup.com LinkedIn 310.780.3185