How to assemble a lean, AI-ready tech stack: PwC

Matt Hobbs

US and Global Head of PwC’s Cloud, Engineering, Data, and AI, PwC US

Email

Daniel Priest

US Chief AI Officer, PwC US

Email

Jennifer Colapietro

Digital Core Modernization Leader, PwC US

Email

Key takeaways

Own what differentiates your business
Rethink the traditional build-or-buy decision. Own the data, expertise, workflows, and capabilities that set you apart, and rent or assemble more standardized capabilities.
Modernize around outcomes, not systems
Use AI-enabled approaches to connect data, models, agents, and applications around specific business needs—without making large-scale migration or replacement the default.
Build governance into the stack
As agents work across systems, models, and data, strengthen security, permissions, policies, and oversight so your technology stack can evolve while maintaining control and traceability.

If you’ve got a stake in using or shaping your company’s tech stack, you’ve likely been consolidating data, moving to cloud, adopting software-as-a-service (SaaS), integrating all your different pieces, and adding intelligence throughout. Now, AI is creating new demands while innovation moves at a dizzying pace.

But developing your tech stack for AI isn’t about buying or enabling apps that deliver a little incremental productivity. It’s about growth. You can’t create an agentic enterprise in which AI gives you a competitive edge and a workforce dividend without the right stack running underneath.

What does the right tech stack for AI look like? It’s lean, because you own only what defines you, things like your proprietary data and the ontology that gives it meaning, your people’s expertise, and the custom agents and workflows that you build on top. The rest, all the capabilities that AI is commoditizing, you rent, assemble, or shrink.

It adds up to a new set of decision criteria—not whether to build or buy but how to assemble a tech stack that can give you a significant competitive edge. Based on what we’ve seen here at PwC and in our work with clients, here’s how it works.

Assemble capabilities that set you apart: How AI-enabled solution architecture works

Until recently, you likely looked to an app when you had a challenge. Your main decision was whether to buy or build it. But building is usually expensive. Buying is too, especially since you have to customize anything vendors provide. And customization’s usually far from perfect.

Now there’s another choice. AI agents can deliver many of these solutions. So, the question becomes what to “switch on” in your apps or platforms, what to buy or rent from vendors, and what to build and own. Whether for customer engagement, invoice processing, IT support, procurement intake and orchestration, tax, or marketing, determine what can set you apart and own that. For the rest, the tasks that just need to get done, buy or rent to save time and keep your costs down.

Go piece by piece. For each tech stack component of a process, ask what it’s for. System of record? Workflow engine? Service layer? Then decide, piece by piece (as opposed to app by app or process by process) what you need to own.
Own what can differentiate you. Is it a task that you and your competitors will do in similar ways? Buy or rent. If it’s mission critical, then you should build, govern, and own it. You can’t buy differentiation from a vendor.
Keep moving. AI’s advances are making new value possible and commoditizing what used to be a moat. Continually revisit both what you need to own (including new capabilities) and what no longer sets you apart.

New architecture solution pathways

Use capabilities in your current platforms and applications

Choose this when:

An existing platform or application offers a tested, configurable, cost-effective solution.
The capability is commodity or table stakes
You need speed, reliability, ongoing upgrades, and lower costs more than IP ownership.

Syracuse University

Turned on Microsoft Copilot and Fabric on existing Microsoft stack; occupancy accuracy 80% → 95%, reporting from weeks to seconds.

Pay for tested, reliable solutions

Choose this when:

There's no "switch-on capability" that meets your needs, but a trusted vendor offers a solution.
The capability is commodity or table stakes
Your vendor has a track record of innovation.
You want to focus your build effort on higher-value differentiators

Chipotle + Workday + Conversational AI assistant

Used Workday as backbone and rented commercially available conversational AI assistant; 20,000 hires in ~2 months, 75% faster time-to-fill.

Create differentiation

Choose this when:

The workflow is core to your competitive edge
Your proprietary data, judgment, or process logic can create meaningful advantage
You want to fully own the IP
No vendor can deliver what you need
Workflow orchestration and logic or end-user

Formula 1^®

Built custom AI knowledge tools for race operations; 4–8 weeks of documentation time eliminated per project.

Keep it evergreen: Fast, focused data modernization

Data modernization generally has meant years of consolidation, moving everything to a platform and standardizing it. That was expensive and slow. And, once you had data in the right place, you still had to integrate it through hard, pre-defined, and generally 1:1 connections.

You now can build an AI data flywheel. Use agents to access and reason across what you have, where you have it. You’ll likely still need some data modernization, including some consolidation and integration work. But with agents, you can do it quickly, focusing on the specific data necessary for a chosen business outcome. And you can make this work ongoing and “evergreen,” rather than resorting to periodic consolidation efforts.

Treat all data as first class. Agents can often access unstructured data (tickets, documents, call transcripts, emails) as easily as structured, transactional data.
Skip the migration. With embeddings/vectors (numeric representations that let agents search by meaning rather than exact match), you can often find data and connect it where it currently sits.
Build what sets you apart. Anyone can buy the tools for AI to work across data. What can give you an edge is your domain knowledge and ontology, how you define terms and how core entities relate with each other and agree upon data’s meaning and value.

Tiered intelligence: Manage a portfolio of AI models

If you wanted machine intelligence, you used to have to build, buy, or rent (and train, manage, and govern) a specific model for a specific purpose, such as churn prediction or demand forecasting. If you wanted intelligence for another purpose, you had to start from scratch.

With the rise of multipurpose AI models, you can now assemble a tiered portfolio, where a few “engines” can power all your needs—frontier models for hard reasoning, cheaper open-weight or open-source models for high-volume work, and high-security models for sensitive work.

Be dynamic yet disciplined. Don’t default to the highest-performing or cheapest model. Instead, with a disciplined approach that dynamically matches the right model and tier to the right task, you can get more high-performing AI for your money.
Make it easy for your users. Even as you dynamically shift among models, the user experience can be seamless. Large language models (LLMs) can sit within workflows, grounded in your data, translating requests into action.
Create your edge. AI models are becoming commoditized, but you can still achieve differentiation. Train models on your data, embed your people’s expertise in workflows, and customize instructions, tools, and guardrails.

How work gets done: Orchestrating agents and apps

Until recently, if your apps, platforms, and databases talked to each other (which they often didn’t), it was through hard-coded connections, built in advance. If you didn’t have the “wiring” in place for a specific pair of apps or databases, people had to stitch outputs together manually.

Now you can have a coordination layer that decides how different pieces should work together in a specific moment, then makes the needed connections. That can mean greater speed, lower costs, and new capabilities as you dynamically mix and match parts into greater wholes.

Put agents where the work is. Enable agents to meet people in the tools they already use, CRM, ERP, email, chat, and your own apps. The experience then lives in your surfaces, not a vendor's, and the underlying agent is swappable, not locked in.
Engineer connections. Even with an orchestration layer, you’ll still need to roll out standards, including model context protocol (MCP), to connect agents, tools, and data.
Make context persistent and mobile. To enable complex yet reliable agent workflows, have your system (rather than individual agents) track of what each agent has already done (state) and carry needed information (context) into the next action.

Shrink it down to its real job: New ways to use your system of record

Your systems of record have always done more than just hold records. Your ERP ran finance and operations, while your CRM ran sales and services. That often meant that the data and processes all belonged to the vendor, and changing a rule meant a dev-heavy project buried in vendor logic, invisible to the people who owned it.

Systems of record now can shrink and just do what they’re supposed to: Hold authoritative, governed data. Workflow logic, business rules, and routing move up to your intelligence and orchestration layers, where the people who own them can change them.

Don’t rip and replace. Even in the AI age, you need the regulated, compliant, governed, auditable, and authoritative source of truth that an ERP or CRM can deliver.
Narrow your footprint and cut costs. As workflows and business logic move up the stack, you can often stop paying for these platforms’ full suite.
Make it agent-friendly. Acquire or engineer systems so agents can pull and record safely (including clean APIs and events) yet can also create and govern custom agents that you will own and run within your workflows.

Your inheritance won’t protect you: Upgrade governance and security

Traditional ERP and CRM enforced access rules and controls, kept an audit trail, and carried much of the compliance burden. Your cyber teams still had a big job, but it was often predictable. They were defending known, bounded systems doing deterministic work.

But these “inherited” systems can’t keep up with agents who act on their own, creating new solutions to problems that didn’t exist a moment before. Malicious agents are innovating too, creating new attacks. And the business needs you to govern AI without slowing it down.

Buy the tooling. Often, your agents’ gateway or platform can come with security machinery like agent registries, identity and permissions, checkpoints, and logging. If it’s good, wire it together as needed and use it. There’s no need to build your own.
Write the policies. With the help of agentic scaffolding, your risk and security teams should set rules that match your priorities and risk appetite—what each agent is allowed to do, when a person’s approval is needed, and where the hard limits sit.
Make coverage complete. With agents so connected, including across platforms and models, it’s critical to confirm that every agent you run (including ones bought from vendors or spun up outside IT) is covered by your chosen governance and security.

Keep finding new paths to value

AI is creating new demands on your tech stack while also opening faster, more cost-effective ways to deliver value. You don’t need to rip and replace and try to do everything at once. But you do need to start moving and then keep moving. Start with the decisions where better intelligence changes a high value outcome, then own what defines you, assemble the rest, and let each win fund the next.