Build your data and technology based on where the value is. Your strategy should set the objective, but leave room for technical experts to determine how to get there. Size each investment, including people's time, to the opportunity and how likely it is to pay off. Then prove it works before you build the rest. Keep that first version thin enough to deliver quickly but broad enough to cut across the business, as we describe in Strategy.

If your data is scattered across systems that don’t talk to each other, trying to scale AI will likely end in failure. Instead of a large upfront data program, start from the use case and build the training data it needs as part of the work. If you’re building a platform, build reusable foundations, since they can cut the cost of each additional use case by 30 to 50%. Agree on how you will judge whether the output is good enough before you start building, and keep that same test running once the platform is live.

Making data discoverable across silos matters more than simply holding it, because teams cannot ask for what they don't know exists. A well-described dataset that a team can find and trust beats a richer one sitting unlabeled in a system nobody outside its owners knows about.

Start with the simplest AI approach that can deliver the desired business outcome, then build on it as you prove feasibility, validate value, and identify where additional capabilities are needed. Sometimes that means a predictive machine learning model, sometimes generative AI, and increasingly, agents that can reason, use tools, and execute work across systems. The goal is not the most sophisticated technology, but the right combination of capabilities to solve the problem reliably and economically.

Test solutions against the realities of your business, not just technical benchmarks. For machine learning, understand the data, the reasons behind its patterns, and how well predictions hold up as conditions change. For generative and agentic AI, evaluate the quality of outputs, the reliability of decisions and actions, and how systems perform when faced with unexpected situations. A solution that works consistently in the real world is more valuable than one that excels in controlled testing but fails in practice.

Design for deployment from the start. Integrate AI into the systems and workflows where decisions are made and work gets done, with clear accountability for outcomes. Build in ongoing evaluation, monitoring, and improvement so solutions remain effective as business conditions, data, and AI capabilities evolve. Make successful solutions reusable across the organization rather than rebuilding them for every new use case.

Scaling means embedding capability into how the business runs, not running progressively bigger pilots. Agentic AI raises the bar here.

Agents need to coordinate with each other and humans, remember what they have already done, and connect to your key systems. Your infrastructure must monitor, audit, and intervene in autonomous actions, not just track performance metrics.

When it comes to partnerships, manage them actively and make sure the integration layer is owned by your organization rather than a vendor. Make knowledge transfer a contractual commitment with partners, since the true test often becomes whether you can operate what they built after they've gone. Be clear about what you are buying: a capability you will eventually bring in-house or a partnership you plan to invest in long term.

Treat AI infrastructure cost as a capital investment rather than a project expense. Token consumption that looks trivial in a pilot can become material at production volume. Know what each team, product, and use case is spending before that happens, not after.

Finally, decide where AI should act on its own and where it should support a person, based on the stakes of the workflow rather than the capability of the model. For low-complexity, low-consequence agents, a kill switch and clear scope boundaries are enough. Once agents touch core operations, payments, or regulated decisions, the oversight question becomes considerably harder, and we cover how to approach it in our Governance section.

These two examples sit at opposite ends of the same journey, because how you prepare for agents is as important as how you deploy them.

Positioning agents for maximum success and minimum risk

The new CEO of a private equity-backed software company faced slowing growth just as competitors shipped agentic features, and the company believed an agentic product would win customers back. The risk was building it on a decade-old platform carrying significant technical debt, maintained by engineers proud of what they had built. Before committing significant engineering effort and the existing customer base to a path, the CEO needed an approach that the board, the PE owners, and the engineers could all get behind.

AlixPartners traced priority agentic use cases through the architecture, combining interviews with a review of all 12 million lines of code. Our review surfaced weaknesses the company had not seen. Parts of the platform believed to be separate were so entangled that small changes rippled everywhere, making every agentic feature slower and riskier to build. We also found that an AI feature could reach sensitive data that the human interface protects, a serious risk once agents act on users' behalf. The analysis also showed that rebuilding would take roughly the same effort as extending the old platform, without leaving the technical debt in place.

We then framed three paths in the board's terms: how fast customer needs would change, and how much risk to place on existing customers. In 4.5 weeks, the board had a recommendation it went on to endorse, with management, the engineers, and the PE owners aligned behind it. The company now had a clear and credible route to the agentic product it was counting on to win customers back.

Putting agents to work

A US healthcare insurer had thousands of unstructured contracts, no standard procurement process, and so many overlapping vendors that leadership could not see what the company bought or from whom. AlixPartners built AI pipelines to read and structure around 40,000 documents across roughly 2,300 suppliers, then added workflows that reviewed contracts clause by clause. With that foundation in place, we pointed agents at the work where manual review had been missing discrepancies at scale: reconciling statements of work against purchase orders and escalating high-risk mismatches to people. The work delivered annual run-rate savings of 7 to 12% of addressable spend, worth $123 million to $228 million.