What a Real AI Implementation Looks Like, Week by Week

Most descriptions of AI implementation fall into one of two categories: the vendor pitch (six seamless steps to transformation) or the academic model (a flowchart with no time scale). Neither maps to reality. This article describes what a production AI implementation actually looks like in a mid-market regulated company: week by week, gate by gate, with the uncomfortable decision points that most methodologies gloss over.

The structure here is drawn from three years of production deployments across financial services, insurance, and government, combined with implementation frameworks from IBM Consulting, MIT's Project NANDA, and peer-reviewed stage-gate research published in December 2025.[1][2][3] The timelines are realistic. The decision points are real. And the most important feature of the process is the one most organisations skip: the willingness to stop.

RAND Corporation's research identifies the primary cause of AI project failure as a fundamental misunderstanding of the business problem.[4] This isn't a technical assessment. It's a diagnostic: does your organisation have a specific, measurable business problem that AI is genuinely the right tool to solve?

This is where most engagements go wrong. The pressure to move forward (from boards that have approved budgets, from technology teams that want to build, from vendors that want to sell) overwhelms the discipline to ask whether the foundation is sound. IBM Consulting's stage-gate framework makes this explicit. Each gate asks "should we spend more money?" not "is this good enough?"[1]

The critical discipline here is specificity. RSM's middle-market research found that 62% of executives said generative AI was harder to implement than expected.[5] The primary reason? Expectations were set at the wrong level of abstraction. "Use AI to improve customer service" isn't a problem statement. "Reduce average claims processing time from 14 days to 3 days by automating initial document review, while maintaining a 99.2% accuracy rate and full audit compliance" is.

The KPIs must be agreed before building starts. Don't retrofit them afterwards. This is a lesson from every failed pilot: if you don't define success before you build, you'll redefine it afterwards to match whatever you built.

This is the point where the "controlled environment" problem that DAS Advanced Systems has documented becomes critical. The difference between a pilot that succeeds and one that fails? Whether it runs against production conditions. A system that works with clean sample data in a demo environment proves nothing about whether it will work with the messy, incomplete, legacy data that your organisation actually has.[6]

A properly run Phase One in a mid-market regulated company typically takes three to six weeks and includes five parallel workstreams:

Three outcomes are possible. SCALE means the MVP demonstrated sufficient value to justify production deployment and expansion. PIVOT means the results were mixed. The problem is worth solving, but the approach needs adjustment. STOP means the initiative didn't produce sufficient evidence of value to justify further investment.

The willingness to stop is the single most important differentiator between organisations that waste money on AI and those that invest it productively. Every methodology claims to include decision gates. In practice, the economic incentives of most vendor relationships push towards continuation regardless of results. The gate only works if the decision-maker has genuinely independent data and the discipline to act on it.

Track A: Scale the solution

Moving from MVP to production deployment involves expanding to additional users and workflows, hardening the infrastructure for reliability and scale, implementing continuous governance monitoring aligned to EU AI Act and ISO 42001 requirements, and integrating deeply with existing operational systems.

The timeline for Track A varies significantly by complexity. For a mid-market company deploying a single AI system (say, an automated document review tool for an insurance underwriter), production deployment might take eight to twelve weeks. For more complex initiatives involving multiple interconnected systems, it can extend to six months.

Track B: Build internal capability

This is the track that separates implementations that last from those that become expensive shelf-ware. Prosci's research is unequivocal: organisations investing in developing their own people see consistently better results than those depending heavily on outside consultants.[7]

Track B includes structured, role-based training (not a single workshop, but an ongoing programme tailored to different user groups), fostering an experimentation culture (which Prosci identifies as the single most significant factor in AI adoption success), democratising AI expertise across the organisation rather than concentrating it in IT, and providing frameworks for identifying future AI use cases independently.

Graduation criteria are specific and measurable. Your client team can operate the AI system without external support. They can identify when the system needs updating or retraining. Governance processes are in place for ongoing compliance. And, critically, the team has a framework for evaluating future AI opportunities without needing to hire external help to do it.

If these criteria are met, the engagement ends. Not because the partner has nothing more to sell, but because the purpose of the engagement was to build capability, not dependency. If the criteria aren't met, the engagement extends. But only the specific gaps are addressed, not a wholesale continuation of the full programme.

DAS Advanced Systems has documented the typical Big 4 engagement cost: $500,000 to $1 million for strategy, $1 million to $2 million for a pilot, and $3 million to $10 million for implementation. For a mid-market company with $300 million in revenue, that represents 1.7% of total revenue. It's an enormous commitment with no guaranteed outcome.[6]

Boutique firms specialising in regulated industries typically operate at 40–60% lower cost, with faster delivery timelines (weeks, not months) and more hands-on implementation support. But cost is only one dimension. The more important question is: what do you own at the end?