Building Your AI Agent Practice

Chapter Two - Building Your AI Agent Practice

Tackling Real and Difficult Challenges

When Palantir launched itself, they didn't go after easy problems. They went after tough challenges across complex industries: aerospace, manufacturing, healthcare, cybersecurity. That focus on hard problems shaped their entire business model.

Reimagining your business as a system integrator means you need to look for complex use cases unlikely to be provided off-the-shelf by vendors or built by companies on their own. Get specific. Here are the kinds of complex, agent-proof engagements you should be hunting for:

Multi-party supply chain orchestration. According to MIT research, this involves building agents that negotiate in real-time across procurement, logistics, and finance while maintaining regulatory compliance across jurisdictions. MIT's Machine Intelligence for Manufacturing and Operations program shows these systems need to understand nuanced regulatory frameworks that vary by geography. EY's work on agentic supply chains shows how autonomous systems can analyse real-time data from multiple sources, adjust procurement schedules, reallocate resources, and communicate with suppliers to achieve timely deliveries, all while providing human managers with oversight insights. One manufacturing use case involved processing 15 years of maintenance logs using retrieval-augmented generation (RAG) to reduce equipment downtime by 23%. That's the level of impact we're talking about.

Clinical trial optimization with regulatory navigation. Harvard Medical School research highlights how healthcare agents must navigate FDA compliance, patient privacy laws, HIPAA requirements, and cross-institutional data governance. Recent Harvard studies show the AI-enabled clinical trials market is growing at nearly 19% CAGR, reaching $21.79 billion by 2030. These aren't simple chatbots. They're systems that understand the nuances of regulatory frameworks that vary by geography. In healthcare supply chains, AI is being used to answer complex 'what-if' questions like 'Among all the scheduled surgical cases for the next two weeks, highlight all affected cases if item number 12345 is on back order' and then automatically identify potential replacements and procurement sources. That kind of contextual reasoning is what separates real value from hype.

Cross-border financial compliance agents. Gartner predicts that by 2027, 40% of AI data breaches will arise from cross-border GenAI misuse, highlighting the complexity of systems that need to understand the intersection of multiple regulatory regimes, tax jurisdictions, and compliance frameworks. Building something that works in the UK doesn't mean it will work in the EU or US. The complexity multiplies.

Manufacturing predictive maintenance with enterprise integration. MIT and McKinsey joint research shows that manufacturing AI leaders deliver 4x results in half the time compared to laggards. This isn't just predicting when a machine will fail, but coordinating maintenance schedules across ERP, HR (staffing), supply chain (parts availability), and production planning systems.

The key insight: these problems require deep context, judgment, and the ability to bridge multiple systems and stakeholders. They can't be solved by a generic agent from a SaaS vendor. Enterprise software partners have been good at building accelerators and industry add-ons on top of ERPs and CRMs. But building for the AI-first world requires a fundamentally different approach.

Three Pillars: Culture, Context, Code

I've identified three areas where technology consulting leaders must focus. These aren't nice-to-haves. They're existential.

1. Culture: Building Teams That Can Execute

The first is culture. I've often seen firms pretending to be know-it-all's. Don't try this with AI projects. Customers can see through it. You and your client-facing teams must come across as trusted advisors, not as experts who already have all the answers.

Traditional consulting is dead. This sounds rude and scary, but it's the brutal reality. Clients want people who've actually been in the trenches. Who've led transformations. Who've failed and fixed it. Who've built something that actually works in the real world.

What you need are blended teams with technical depth and real experience. Your organization needs a bottom-up, meritocratic culture where there's strong bias for action. Let's explore what this looks like in practice.

OpenAI's launch of Codex, a coding agent that can write, edit, and understand code, provides an instructive example. OpenAI launched Codex as a research preview in May 2025 and achieved general availability in October 2025.[1] By early October 2025, daily usage had grown more than 10x since launch, and GPT-5-Codex is one of OpenAI's fastest-growing models, having served over 40 trillion tokens in the three weeks following launch. The key insight isn't the metrics, it's the team structure that enabled this speed. There must have been multiple Codex prototypes floating around OpenAI before they decided to push for launch. Those efforts were typically undertaken by small groups of individuals without asking permission first. That's the culture you need to cultivate.

How do you actually build this culture?

Start with your incentive structure. If you're still bonusing people purely on utilization rates, you're incentivizing the wrong behavior. Create 'AI pods', small teams of 3-5 people with protected time, say 20% of their hours, for experimentation. Don't charge this time to clients. Yes, it hurts short-term margins, but it's your R&D investment. That's how winning firms operate.

Change your hiring profile. You need to attract and retain people who know tech cold and who've sat where the client sits. People who understand that execution is messy and political. Hire for technical depth and real-world scars. The future belongs to people who can bridge analytical depth with operational experience from actually doing the work. Everyone else gets ignored by clients who've seen it all before.

What about your existing workforce? This is the question nobody wants to ask: What do you do with a 15-year SAP consultant who's brilliant at understanding business requirements but has never written a line of code?

Create a reskilling path. Three months intensive on Python basics and AI fundamentals. Pair them with technical consultants on real projects. Teach them prompt engineering, their domain knowledge is gold when combined with the ability to direct AI. Move them into 'AI collaboration specialist' roles where they become the bridge between technical delivery and business outcomes.

2. Context: Your Secret Weapon

The second pillar is context. If you've been in the business for a long time and have built solutions for specific industries, you already have enormous amounts of context available. By context, I mean domain know-how, operational scars, bottlenecks specific to industries. In many cases, this is tribal knowledge sitting across your functional consultants, technical consultants, solution architects, onsite PMs, they've all learned a great deal about customer businesses and pain areas. It's time you harvested that knowledge and leveraged it to build for the AI-first world.

Few years ago, Palantir was dismissed by many as a glorified staff augmentation company. Now it has a market cap over 400 billion dollars. There are lessons for tech consulting leaders from Palantir's playbook.

Palantir divided its engineers into two types:[2]

• Forward Deployed Engineers (FDE): Engineers who work with customers on-site

• Product Development (PD): Engineers who work on the core product team

An ex-Palantir employee provided insights into this model: "The key idea is that you gain intricate knowledge of business processes in difficult industries, manufacturing, healthcare, intel, aerospace and then use that knowledge to design software that actually solves the problem. PD engineers then 'productize' what FDEs build and build software that provides leverage for FDEs to do their work better and faster."

"This is how much of the Foundry product took initial shape: FDEs went to customer sites, had to do a bunch of cruft work manually, and PD engineers built tools that automated it. Need to bring in data from SAP or AWS? Here's a data ingestion tool. Need to visualize data? Here's a visualization tool. Need to spin up a quick web app? Here's a low-code UI builder. Eventually, you had a comprehensive set of tools around the loose theme of 'integrate data and make it useful somehow."

This model essentially allowed Palantir to pull off a rare pivot from service company to product company. Palantir maintains 80% gross margin compared to Accenture's 32%.[3]

Here's how you can adapt this model. Create your own version of the FDE model, call them 'Embedded Innovation Engineers' or 'Client Labs Teams.' Here's the structure:

• Deploy 2-3 person teams on-site with your largest clients for 6 - 12 months rotations

• Give them a dual mandate: deliver immediate value AND identify productization opportunities

• Create feedback loops to your central product/engineering team

• Every quarter, have embedded teams present 'pain harvests' documented patterns they've seen that could be automated or turned into reusable IP

Direct immersion in the client environment via approaches like the FDE model permits accelerated domain learning, builds trust, and results in highly relevant solutions. Professional services leaders should consider secondments, embedded consulting, or client co-location as mechanisms to deepen impact and understanding. Real enterprise value, especially in complex sectors, emerges from mutual trust and firsthand context, not distant analysis.

Now here's something critical that's happening right now: the FDE role is exploding. In November 2025, the Financial Times reported that monthly job listings for forward-deployed engineers increased more than 800% between January and September 2025.[4] This isn't a future trend, it's here. OpenAI set up its FDE team at the start of 2025 and expects to grow it to about 50 engineers by year-end. Anthropic said it would grow its applied AI team, which includes FDEs and product engineers, fivefold in 2025. Cohere's co-founder Aidan Gomez noted that deploying engineers at the beginning of a customer's contract helps build long, durable relationships. They embed engineers at contract start to ensure customers get exactly what they need, then scale back once companies are up and running.

The practical impact? OpenAI customized its technology for John Deere, an agricultural machinery manufacturer, to help create more precise farming tools. Farmers reduced chemical spraying by 60 to 70 percent. That's real value. That's what forward deployment looks like in practice.

Agents are only as good as the context they're provided. Context is that which is scarce. That's the foundational insight your agents will need. Of course, with company data, they can reason and perform actions. But to build the right product for the right use case that solves a real problem, you need the context only you and your team have for the industry, processes, and people who work on them.

3. Code: Building and Scaling Engineering Teams

The third pillar is code, how you build and scale your engineering team.

Agents aren't just chatbots. They're goal-oriented systems that can take actions on your behalf. Unlike traditional AI models that simply respond to prompts, agents are designed to autonomously perform tasks. They can plan, reason, and execute sequences of actions to achieve specific objectives.

They're not magic. An agent is just a system that helps AI do useful work. Here's what actually happens: The agent gathers context, what you asked for, previous conversations, relevant information. It sends that to the AI model. The model responds with an answer, a request to use a tool (like searching data or running calculations), or both. The agent runs those tools, feeds result back. This keeps going until the model has a final answer. Then the agent gives you that answer or takes the action per given instructions.

Getting your team AI-ready comes down to one thing: Python.

Not because it's trendy. Because it's what actually works when you're building AI agents and systems at scale. Every major AI framework, OpenAI, Anthropic, LangChain, LangGraph, Microsoft Semantic Kernel, AutoGPT the tools your engineers will use daily are built for Python. The libraries, documentation, and community support are all there.

Your team doesn't need to be Python experts. They need to understand:

• How to work with APIs (calling AI models, connecting systems)

• Data handling (reading files, processing information)

• Basic scripting (automating tasks, building workflows)

That's 80% of what you'll use building AI agents.

But here's what the textbooks won't tell you: you need to understand agent frameworks deeply. Not surface level. I'm talking about:

• LangChain and LangGraph for building stateful, multi-step agents

• Semantic Kernel if you're in the Microsoft ecosystem

• RAG (Retrieval Augmented Generation) patterns for connecting agents to your client's proprietary data

• Testing and evaluation frameworks for non-deterministic systems

Here's the real challenge: How do you QA a system that gives different answers each time? Traditional software testing doesn't work. You need:

• Evaluation datasets with expected behaviors, not exact outputs

• Human-in-the-loop review processes

• A/B testing frameworks that measure business outcomes, not just technical metrics

• Red teaming exercises to find edge cases

Build vs Buy decisions matter more than ever. Not every client needs a custom agent. Sometimes you're better off configuring an off-the-shelf agentic platform (Palantir's AIP, Microsoft's Copilot Studio, or Salesforce's Agentforce) and adding your client-specific context layer.

Your decision tree should be:

• Is this a unique, complex problem requiring deep customization? → Build custom

• Is this a common workflow with specific industry context? → Configure existing platform + add your IP layer

• Is this a simple automation? → Use out-of-the-box agents

Don't waste four weeks reinventing the wheel when you can configure an existing platform in four days.

Notes:

[1] TechCrunch & VentureBeat. (2025, May-October). Coverage of OpenAI Codex launch as research preview (May 2025) and general availability (October 2025), with 10x daily usage growth since launch.

[2] Nabeel Qureshi. (2024). "Reflections on Palantir" essay detailing FDE (Forward Deployed Engineers) and PD (Product Development) model showing how field context drives product development and gross margin expansion.

[3] Palantir Technologies. (2024). Q4 2024 Financial Results showing 80% gross margin (vs 81% in 2023) compared to Accenture's 32% gross margin, demonstrating successful service-to-product pivot.

[4] Financial Times. (2025, November 2). "The new hot job in AI: forward-deployed engineers." Reporting 800%+ growth in FDE job postings between January-September 2025, with OpenAI expanding FDE team to ~50 engineers and Anthropic planning fivefold growth.