From Diagnosis to Prescription

In Part 1 of this series, the diagnosis was laid bare: nine interlocking traps that cause in-house enterprise AI platforms to consume budgets, produce polished demos, and deliver zero measurable business impact. The data was unambiguous: internal AI builds succeed at one-third the rate of purchased solutions (MIT NANDA, 2025), 95% of enterprise AI pilots fail to move the P&L, and not one of 598 published enterprise AI case studies contained rigorous evidence of business impact (Applied AI, 2025).

But a diagnosis without a prescription is just complaining.

This second part is the practitioner’s playbook: six concrete strategies for enterprises that want AI to deliver outcomes rather than architecture diagrams. These are not theoretical frameworks. They are drawn from direct experience leading automation at scale in a regulated, global enterprise, and are validated against the research that exposed the traps in Part 1.

The unifying principle is simple: own the domain knowledge, rent the AI infrastructure.

Strategy 1: Invest in Process Intelligence, Not Model Infrastructure

The competitive moat for any enterprise is not the ability to run an LLM. That capability is fully commoditised. OpenAI, Anthropic, Google, and a dozen open-source alternatives provide model access at commodity prices. Wrapping one of these models in a corporate UI adds no defensible value.
The moat is understanding your own processes deeply enough to know exactly where AI creates measurable impact: which handoffs introduce delay, which decisions could be automated, where rework occurs, and what the cost of each inefficiency is.

Process mining, workflow analysis, and operational data quality are the foundations that determine whether any AI initiative, bought or built, will succeed. This is precisely the gap that current tools leave unaddressed. Descriptive process mining tells you what happened. The real value lies in prescriptive capabilities, recommending specific process redesigns and empirically validating their effectiveness.

MIT’s research confirmed that the biggest ROI from generative AI comes from back-office automation, eliminating outsourcing, cutting external agency costs, and streamlining operations. These are process problems. Solving them requires process expertise, not another model wrapper.

The moat is not the model. The moat is knowing where the model creates value.

Strategy 2: Solve Specific Business Problems, Not Technology Problems

Every successful enterprise AI deployment in the MIT research started with a well-defined business problem: reduce customer support cost per ticket by 40%, eliminate manual data entry in invoice processing, and cut regulatory submission review time by half. They did not start with “build an AI platform.”
The discipline required is to resist the platform instinct and instead work backward from the P&L. Where does the organisation spend money on low-value cognitive work? Where do processes break down at handoff points? Where is rework consuming capacity? Answer those questions first. The AI tooling decision follows naturally.

BCG’s 2024 global AI survey found that organisations achieving real AI impact tied initiatives directly to revenue-generating workflows and cost reduction, not peripheral experiments. They asked, “How does this change how we work?” before asking, “What can this tool do?” The framing matters. When the starting point is a business problem, success is measurable. When the starting point is a technology ambition, success is whatever the platform team decides to call it.

Strategy 3: Buy the Plumbing, Own the Business Logic

Enterprise AI infrastructure, model hosting, vector search, authentication, observability, and governance are table stakes that multiple vendors deliver at scale. There is no competitive advantage in rebuilding it internally. The MIT NANDA data makes this point with brutal clarity: purchased solutions from specialised vendors succeed 67% of the time versus 22% for internal builds. That is a 3-to-1 advantage in success rate.

What deserves internal investment is the business logic layer: the domain-specific rules, compliance requirements, process expertise, and institutional knowledge that no vendor can replicate. In life sciences, that means understanding the regulatory submission lifecycle, the clinical trial milestone dependencies, the vendor governance workflows, and the country-specific compliance requirements that govern how work actually gets done.

Encode that knowledge into well-structured data, clear process maps, and measurable KPIs. Then plug it into the best available AI tools. The tools will improve every six months. The domain expertise took years to build. Invest accordingly.

Purchased AI solutions succeed 67% of the time versus 22% for internal builds. A 3:1 success rate advantage.
– MIT NANDA, 2025

Strategy 4: Measure Outcomes, Not Outputs

Stop measuring AI success by the number of features deployed, models fine-tuned, or agents built. These are output metrics that tell you the team is busy. They do not tell you the enterprise is better off.

Start measuring by cycle time reduced, cost per transaction lowered, rework eliminated, and compliance risk mitigated. These are outcome metrics that tie directly to the P&L. When the board asks “what did AI deliver this quarter?” the answer should be in currency, not in model counts.

MIT’s research found that the biggest ROI comes from back-office automation, eliminating BPO, cutting external agency costs, and streamlining operations. These are measurable, unglamorous, and high-impact. They will never make a compelling demo at an all-hands meeting. They will, however, make a compelling case at a board meeting.

The organisations that survived the dot-com bubble were not the ones with the most ambitious technology visions. They were the ones who could answer a simple question: how does this make money? The same filter applies to enterprise AI in 2026.

Strategy 5: Establish Baselines Before You Build Anything

If 598 out of 598 published enterprise AI case studies lack rigorous evidence (Applied AI, 2025), the problem is not a lack of AI capability. It is a lack of measurement discipline.

Before any AI initiative launches, document the current state: cost per transaction, average cycle time, error rate, volume of manual handoffs, and rework percentage. Without that baseline, every claim of improvement is unfalsifiable, and unfalsifiable claims are the fuel that keeps innovation theatre running.

This is not optional. It is the single most important step in any AI initiative, and it is the step that nearly every organisation skips. The reason is uncomfortable but obvious: establishing a rigorous baseline creates accountability. If the team knows that outcomes will be measured against a documented starting point, the incentive shifts from building impressive demos to delivering measurable results. That shift in incentives is the difference between the 5% who succeed and the 95% who don’t.

No baseline = no accountability = no outcomes. This is the single step that separates the 5% from the 95%.

The discipline of baselining also forces a harder conversation upstream: Is this problem worth solving with AI at all? If the current cost per transaction is already low or the cycle time is already within target, an AI initiative adds complexity without value. The baseline reveals that. Without it, every process becomes a candidate for AI intervention, and the team optimises for quantity of initiatives rather than quality of impact.

Strategy 6: Fix the Foundation Before Adding Intelligence

Before any AI initiative, audit the systems it will touch. Is the metadata clean? Are integrations documented? Is the process data captured in event logs or scattered across emails and spreadsheets?

If 63% of organisations lack AI-ready data management practices (Gartner, 2024), the highest-ROI investment for most enterprises is not an AI platform. It is getting the data house in order. Sweep’s 2025 post-mortem was definitive: enterprise AI did not fail because of model limitations. It failed because the system’s AI was deployed into environments that were not legible enough.

In practical terms, this means three things. First, ensure that process data is captured in structured event logs rather than trapped in emails, spreadsheets, and tribal knowledge. Second, audit and clean the metadata in core enterprise systems, Salesforce, SAP, and ServiceNow, so that AI tools can read and reason about them accurately. Third, document integration contracts between systems so that any AI layer built on top has a stable, predictable foundation.

This work is not glamorous. It does not produce impressive demos or make for exciting quarterly updates. But it is the work that determines whether any subsequent AI initiative, bought or built, will succeed or join the 95% that fail.

The Bottom Line

The enterprise AI platform trap is not a technology failure. It is a strategy failure compounded by a measurement failure, wrapped in a self-sustaining narrative that nobody has an incentive to challenge, until the board starts asking for numbers.

The data is unambiguous: internal builds succeed at one-third the rate of purchased solutions. Ninety-five percent of AI pilots fail to move the P&L. Eighty-eight percent of pilots never reach production. 40% of agentic AI projects will be cancelled within 2 years. And when independent researchers examined 598 published enterprise AI case studies, not a single one contained rigorous evidence of business impact.

The most damaging version of this failure is not the team that tries and fails visibly. It is the team that builds perpetually, demos impressively, measures nothing rigorously, and convinces the organisation for years that it has an enterprise AI capability when what it actually has is a permanent pilot programme with a growing budget and no exit criteria.

The six strategies in this playbook share a common thread: they redirect attention from the technology to the business problem, from outputs to outcomes, from building infrastructure to understanding processes. The winning strategy is not to build the platform. It is to understand the business deeply enough to know which problems are worth solving, buy the best available tools to solve them, establish measurable baselines before you start, and hold the programme accountable to business outcomes, not technical activity.

The trap is not the technology. The trap is the belief that building the technology is the same as creating value. It is not. And in 2026, the organisations that cannot tell the difference will discover that the board has stopped believing the demos.

The reckoning is here. The question is which side of it we choose to be on.

➤ Read Part 1: “The In-House Enterprise AI Platform Paradox: Nine Ways In-House AI Platforms Burn Budgets and Deliver Demos, Not Outcomes”

References

• Applied AI. (2025). Meta-Analysis of 598 Enterprise AI Case Studies. Applied AI Newsletter, Issue 01.
• BCG. (2024). Where’s the Value in AI? BCG Global AI Survey.
• Dataiku / Harris Poll. (2025). Survey of 600 Enterprise CIOs on AI Performance and Career Impact. Dataiku Stories.
• Deloitte. (2026). The State of AI in the Enterprise, 7th Edition. Deloitte Insights.
• Fortune / ServiceNow. (2025, October 29). AI Doesn’t Fail on Tech—It Fails on Leadership. Fortune.
• Gartner, Inc. (2024). Survey on AI-Ready Data Management Practices. Gartner Research.
• Gartner, Inc. (2025, June 25). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Gartner Newsroom.
• IDC Research. (2025). AI Pilot-to-Production Scaling Analysis. Referenced in AI Smart Ventures, Why Do AI Pilots Fail?
• Korizis, G. (2025). Interview: How Companies Are Escaping Pilot Purgatory. EnterpriseDB.
• McKinsey & Company. (2025). The State of AI in 2025. McKinsey Global Survey.
• MIT NANDA Initiative. (2025). The GenAI Divide: State of AI in Business 2025. Massachusetts Institute of Technology.
• NStarX Inc. (2025, December 16). The Next Frontier of RAG: How Enterprise Knowledge Systems Will Evolve (2026–2030). NStarX Blog.
• S&P Global Market Intelligence. (2025). AI Project Failure Rates on the Rise. CIO Dive.
• Sweep. (2025). Why Enterprise AI Stalled in 2025: A Post-Mortem. Sweep Blog.
• VentureBeat. (2026, January 3). Six Data Shifts That Will Shape Enterprise AI in 2026. VentureBeat.