Six Months of AI Strategy, Zero Lines of Code
- Production AI in 4 weeks
- Exception backlog cleared
- Board demo replaced slide deck
- 4 weeks Time from engagement to production AI
- 78% Exception routing accuracy at launch
- 260 to 80 Weekly person-hours (before/after)
The results
- Time from engagement to production AI
- 4 weeks
- Exception routing accuracy at launch
- 78%
- Weekly person-hours (before/after)
- 260 to 80
- Average exception handling time
- 8 min to 2 min
- Additional headcount required
- $0
- Live demo that replaced the 40-page deck
- 1
The situation
A UK logistics company with 400 employees and a board-level AI budget of $500,000 had spent six months trying to implement AI. They had hired an AI strategy consultant, evaluated 12 vendors, shortlisted 3, started two proof-of-concept builds, abandoned both, and produced a 40-page strategy deck. Nothing was running in production. This is the rule, not the exception: a 2025 S&P Global survey of more than 1,000 enterprises found that 42% had abandoned most of their AI initiatives, up from 17% a year earlier, and the average organization scrapped 46% of its proof-of-concepts before they reached production. The board was losing patience. The operations team, who were supposed to benefit from the AI investment, had stopped attending the AI steering meetings because they did not believe anything would ever ship.
The real problem
The fundamental problem was scope. The AI strategy deck covered seven use cases across four departments. Each use case had a business case, a vendor recommendation, and a timeline. But no one had asked the most important question: which one do we build first, and what does "done" look like for that one thing? The strategy treated all seven use cases as a portfolio to be delivered in parallel. In practice, this meant that every meeting became a debate about relative priority, vendor selection, and resource allocation across seven initiatives. No single initiative had enough focus or resources to reach production. McKinsey's 2025 AI survey found that organizations reporting significant financial returns were twice as likely to have redesigned an end-to-end workflow before choosing a modeling technique. This deck had inverted that order: it picked vendors and architectures for seven workflows it had never actually redesigned.
The two abandoned proof-of-concepts illustrated the problem clearly. The first was a demand forecasting model that required 18 months of historical data to be cleaned and normalized before training could begin. The data team estimated three months for data preparation alone. The project was shelved at month two when the data cleaning proved more complex than expected. The second was a chatbot for customer service that required integration with seven internal systems. The integration work was estimated at four months. The project was shelved at month one when the first integration took three weeks instead of the estimated three days. Both followed the textbook failure pattern: a demo built on clean sample data that breaks the moment it meets real-time systems and messy production records. Industry research is blunt about the gap. Of the AI proof-of-concepts that do reach production, the average journey from prototype to live system takes about eight months, and most never make the trip at all.
The operations team had the most acute pain. They processed approximately 2,000 delivery exceptions per week: packages that were delayed, misrouted, or damaged. Each exception required a human operator to review the details, determine the appropriate resolution workflow, and manually route it to the right team. The process took an average of 8 minutes per exception, consuming roughly 260 person-hours per week. The operations director had been asking for automation for two years. The work was exactly the kind of structured, repetitive classification that machine learning handles well: reading a handful of exception attributes and mapping them to the correct resolution workflow, the same triage that 2026 logistics platforms now automate to move from reactive firefighting to proactive routing. The AI strategy deck listed it as use case number five out of seven, scheduled for Q4. The operations team had given up waiting.
The approach
Single-use-case session
We ran a two-hour session with the operations director, the data team lead, and the CTO. The goal was to identify one use case that met three criteria: high business impact, available training data, and achievable in four weeks. Delivery exception routing scored highest on all three. The data was already structured in the operations database, the business impact was quantifiable (260 person-hours per week), and the matching logic was pattern-based rather than generative, meaning a simpler model architecture would work. The four-week cap was deliberate. A use case you cannot ship in a month is usually a use case that has not been narrowed enough, and a stalled pilot teaches the organization nothing except that AI does not work here.
Data validation and model architecture
We spent three days validating the exception data: 18 months of resolved exceptions with operator notes, resolution codes, and routing decisions. The data was clean enough for training without extensive preprocessing. We chose a classification model that mapped exception attributes (type, carrier, origin, destination, value) to resolution workflows. No LLM required, no prompt engineering, no token costs that scale with volume. This is a pattern the broader industry keeps relearning: the strongest results in logistics come from combining classical operations methods with historical data and a focused model, not from reaching for the largest available system. The model needed to be accurate, not creative.
Working integration build
The integration was built to run alongside the existing manual process rather than replacing it. When a new exception came in, the model predicted the resolution workflow and pre-populated the routing fields. The operator could accept the suggestion with one click or override it. This is the human-in-the-loop design that production AI teams now treat as the default: the model proposes, a person disposes, and every override is captured as a labeled training signal rather than lost in the workflow. It meant the model could be deployed without any risk to existing operations. Wrong predictions would simply be corrected, and the corrections would feed straight back into the model over time.
Production deployment and measurement
The integration was deployed to production in week three. By week four, the model was correctly routing 78% of exceptions without operator intervention. The remaining 22% required manual review, but the pre-populated fields still saved time even when the operator changed the routing. The operations team went from 8 minutes per exception to under 2 minutes on average. Weekly person-hours dropped from 260 to approximately 80, freeing roughly 180 hours a week for the exceptions that genuinely needed human judgment. The operations director presented a live demo to the board instead of the usual slide deck.
In their words
We spent eight months talking about AI. You shipped something in three weeks. The operations team stood up and clapped when they saw the routing predictions working. I have never seen that happen in a steering meeting.
Everything you would ask in the first meeting
Do you always recommend starting with one use case?
Yes. Every AI initiative we have seen succeed started with a single, well-defined use case that reached production quickly. Broad strategies that try to cover multiple use cases simultaneously consistently stall, which is one reason the published enterprise failure rates sit so high. Ship one thing, prove value in real operations, then expand from a working foundation instead of a slide.
Did the accuracy improve over time?
Yes. The model accuracy went from 78% at launch to 89% after 8 weeks as operator corrections provided additional training data. The feedback loop was built into the workflow from day one, so every accept and every override made the next prediction better. Models without that loop tend to drift the other way as real-world data shifts; this one improved precisely because the corrections had a route back into training.
What if we do not have clean historical data?
Data quality is the most common blocker for AI projects. If the data is not clean, we scope a data preparation sprint first. We are honest about timelines: if data prep takes 8 weeks, we say 8 weeks, not 2. Poor data readiness is the single most cited reason AI initiatives are abandoned, so we would rather quote it accurately than discover it in week three. The logistics company had 18 months of clean, structured exception data, which is what made the 4-week timeline possible.
Do you build custom models or use off-the-shelf AI?
It depends on the use case. For this project, a custom classification model was the right choice because the problem was well-defined and the data was structured. For other projects, we use OpenAI, Anthropic, or open-source models where they fit better. A generative model would have been slower, more expensive, and less accurate for a fixed-category routing task like this one. The technology choice follows the problem definition, not the other way around.
4 weeks Time from engagement to production AI
Get in touch
Tell us what you want built. We tell you where to start and what it takes.
ByKostadin StambolievFounder and Imaginator
Published