Most enterprises have approved AI budgets and run Claude pilots, but few have built Claude use cases that generate return on investment a CFO would recognize.
phData is an Anthropic Preferred Partner with over 100 Claude certifications that runs on Claude across its own data engineering, machine learning, business intelligence, and operations teams.
Turning Claude use cases into production systems follows four stages: identifying high-value opportunities, proving economic value quickly, building governance and infrastructure in parallel, and scaling on what is already working. Skipping any stage is why most pilots stall before production.
Every enterprise has an AI mandate.
Budgets are approved. Tools are procured. Pilots are forming. Somewhere, a team is building a roadmap.
Does it make sense? What are the Claude best practices? Do Claude best practices exist? We may need some help.
And then comes the familiar parade: frameworks, maturity models, five-phase roadmaps, and opinions from people who are well-read on the topic. “The path forward” presented with total confidence by people who haven’t actually walked it.
There is nothing wrong with a plan. But there is a difference between explaining the frontier and operating at it.
That is the distinction phData is building around.
Anthropic is moving AI capability at an exponential rate. It can be a challenge to keep up and access alone won’t offer the same exponential advantage. We believe the future will be won by the organizations that learn fastest at the edge of what models can do, turn that learning into repeatable systems, and keep moving when the next capability changes the map again.
That is why phData runs on Claude.
“Running on Claude” is a working discipline that goes beyond a collection of skills or prompts. We put the technology into our own consequential work, fail fast, find the constraint it exposes, build the capability required to clear it, and do it again. We do this internally under real load, real deadlines, real money, and deliver real value so we can do the same for our customers.
The model is not the transformation
Claude can make an individual dramatically more effective. A developer can move faster. An analyst can produce a better first pass. A project manager can turn a whiteboard into structured work. A team can turn a week of scattered discussion into a useful decision record.
Those gains are real. They are also only the beginning.
When we first began introducing AI into our business, we saw the same phenomenon over and over. When one part of a workflow gets faster, the bottleneck moves. Code generation creates pressure on review, testing, deployment, requirements, data access, and context. Better content creation creates pressure on approval, review, and distribution. Faster analysis creates pressure on the decisions and operating processes downstream. More project backlogging demanded more prioritization.
The prompts got us part of the way there, but the system had its gaps.
This is where many AI programs stall. They measure activity instead of redesign. They count seats, prompts, and pilots, then mistake local productivity for organizational progress.
But personal productivity is not enterprise velocity.
Enterprise velocity happens when the system around the work changes too: shared context, usable data, permissions, integrations, evaluation, observability, cost controls, clear ownership, and human escalation paths. It happens when a great individual workaround becomes a trusted, reusable capability rather than a new pocket of technical debt.
The question has shifted from, “Where can we add AI?” to “What is the next bottleneck in our systems?”
That is the work.
How phData turns Claude use cases into durable capability
We push the frontier on ourselves first
Frontier models change quickly. Their capabilities are not fixed, and neither are the patterns that make them useful. The only credible way to advise customers in that environment is to operate close enough to the frontier to discover where it holds, where it fails, and what it takes to make it durable.
So phData uses Claude across data engineering, machine learning, business intelligence, go-to-market, marketing, and operations.
The practice is deliberate trial and error. We test the edges of a capability in our own work, learn where it breaks, and distinguish the exciting demo from the pattern that can survive real delivery.
An analytics engineer can update an underlying Power BI model from a plain-English request. A BI developer can create a live dashboard wireframe with a client in the room, rather than defer the work to a follow-up. Teams synthesize Slack discussions into weekly digests, organize work, and build better first drafts as part of how work gets done.
One engineer built a near-real-time data pipeline in seven hours. It now runs autonomously every two weeks and has delivered roughly ten times its build cost. Another integrated Claude into a screenshot workflow and eliminated a persistent manual step.
Not every experiment survives. That is intentional.
Some approaches create more review work than they remove. Some need cleaner context, stronger evaluation, a human escalation path, or simply a better-defined business process before they can be trusted. Those failures are not an inconvenience to hide; they are how we learn what production actually requires.
We have a strict standard: does the capability work under real constraints? Is it reliable enough to trust? Does it improve an actual process? Can another team use it without inheriting the original builder’s tribal knowledge? Does it create value that can be measured?
The experiments that pass become reusable skills, patterns, guardrails, and accelerators. The lessons from those that do not pass become constraints we design around. Together, they become part of phData Forge™, our AI-native delivery system for scoping, building, and executing engagements.
That is how our work compounds. Curiosity becomes a working pattern and a phData internal capability that enables our clients to start well past zero.
Forge turns what our teams learn with Claude into reusable delivery assets, combining specialized AI capabilities with the expert judgment, validation, and accountability clients expect.
The gap is where the work is
Frontier labs are giving enterprises powerful new raw material. Like any raw material, it needs to be refined to produce something of value.
That gap between model capability and enterprise reality is where phData works.
Repeating the AI market narrative back to our customers helps nobody. That’s a route up the mountain without a sherpa. We prefer to chase the implications of the frontier with customers: to take a new capability seriously enough to test it. We bring the pragmatism (and hard-earned failure) to operationalize Claude-first systems thinking and the engineering chops to make it survive as frontier capabilities continue to grow.
Sometimes the answer is an AI tutor rebuilt after a public proof of concept hallucinated in front of an executive team. We built a system powered by Claude; it now serves hundreds of thousands of students per day in production.
Sometimes it is browser agents for a procurement technology company that had been manually placing orders across dozens of vendor sites through brittle scripts and human fallbacks. The Claude-powered agents now automate more than $1 billion in annual purchasing volume.
Sometimes it is a financial-services contract assistant that cites the exact language behind every answer. It cut response time from roughly three days to under five seconds and saves more than $400,000 annually.
These Claude use cases become operating systems for important work. Their value is measured in how they reshape a critical part of the business with very specific users and optimize for speed, governance, confidence, and capacity.
Build the flywheel, not a pile of pilots
Procuring Claude and running on high-value Claude use cases are two different projects. Most enterprises have already solved the first. Far fewer have turned access into measurable enterprise AI ROI.
The path that actually works has four stages, and skipping any of them is usually why a promising pilot never makes it to production. Here’s what we’ve learned:
Identify the highest-value Claude use cases first. A short, disciplined pass (we prefer to use Claude, of course) that surfaces the handful of opportunities that positively impact EBITDA. We ranked them by impact, feasibility, data readiness, and operational fit before a single line of code is written.
Prove economic value fast, not adoption. The first Claude use case needs to be delivered in weeks, not quarters, and be measured against the outcome that justified building it in the first place: cost reduced, cycle time cut, revenue protected, not seat utilization or prompt counts. Those don’t prove value.
Build the infrastructure in parallel. A working demo is not a production system. Governance, evaluation frameworks, cost controls, data-platform integration, observability, and human escalation paths cannot be bolted on after usage becomes expensive or a security review stalls the project.
Scale on top of what’s already built. Every subsequent Claude use case should be faster and cheaper than the last, because the governance, context, and reusable components from the first one already exist. If your tenth use case costs the same as the first, nothing was actually built; it was just repeated.
Scaling also means operating AI capabilities after launch. Teams need clear ownership, enablement, cost management, and continuous improvement as models, data, and business requirements change. Production value compounds only when the organization can sustain it.
The frontier will keep moving
Every advance in AI expands the set of problems an enterprise can meaningfully take on. The opportunity goes beyond using the most capable model to building a more capable organization around it.
That takes technical fluency, systems thinking, and a willingness to continually reimagine how work moves. It means treating each breakthrough as an invitation to learn: to discover the new bottleneck, redesign the surrounding workflow, and turn the result into a capability the organization can use again and again.
The enterprises that lead will build this muscle at the frontier. They will learn quickly, prove value, harden what works, and carry that momentum into the next meaningful challenge. Over time, this is how experimentation becomes durable advantage.
As Claude practitioners, we hold ourselves to that standard.
phData is an Anthropic Preferred Services Partner. We help organizations turn Claude from access into enterprise velocity by building what matters today and preparing for what the frontier makes possible tomorrow.
Ready to work with real Claude practitioners?
phData is an Anthropic Preferred Services Partner. See how we take Claude from proof of concept to production.
FAQs
What does it mean to run on Claude?
Running on Claude means moving beyond prompts or licensed access to build AI systems that create measurable business value. It requires identifying high-impact Claude use cases, proving economic outcomes quickly, building governance and infrastructure in parallel, and scaling on top of what is already built. Most enterprises have access to Claude; far fewer have turned that access into something a CFO would recognize as return on investment.
How is phData qualified to help customers run on Claude?
phData has over 100 Claude certifications. We run Claude internally across data engineering, machine learning, business intelligence, presales, marketing, and operations. That internal practice generates a compounding body of experience and Claude use cases that inform every customer transformation. phData is also a Preferred Partner in the Claude Partner Network, with a certified delivery team and a record of production transformations.
Why do most Claude use cases fail to reach production?
Most Claude use cases stall because teams treat governance, evaluation frameworks, data-platform integration, and observability as post-launch work. By the time a pilot succeeds technically, there is no infrastructure to support it in production. Organizations that make it to production build the trust conditions and the use case simultaneously.
What is the right first Claude use case?
The right first Claude use case is high-impact, technically feasible with your current data, operationally ready for AI, and measurable against a financial outcome. phData uses a structured prioritization pass, typically conducted with Claude itself, to rank opportunities by EBITDA impact, feasibility, data readiness, and operational fit before any code is written. Speed to economic proof matters more than technical ambition in the first cycle.
