Enterprise teams often postpone generative AI projects because their data is incomplete, inconsistent or trapped in mixed formats. Joe Rose, president of technology provider JBS Dev, argues that this is the wrong threshold: modern models can begin creating value before an organisation completes a perfect data transformation.
The claim is not that poor data stops mattering. It is that large language models, optical character recognition and agentic workflows can help interpret imperfect inputs while a controlled deployment improves over time.
That shifts the practical question from “Is every record clean?” to “Can the system recognise uncertainty, route exceptions and produce a measurable result at a sustainable cost?”
Imperfect inputs can support a staged start
Rose describes a medical-sector migration involving billing records distributed across PDFs and images, with information sometimes placed in the wrong fields. Generative AI extracted text, used OCR on images and separated cleaner records from problematic ones.
Agents then compared customer records with insurance contracts to assess whether billing rates were correct. The workflow did not need to automate everything immediately. It could begin around 20%, expand toward 40% and continue as teams learned where models succeeded or failed.
This incremental approach is more realistic than treating AI as conventional software that is built once and forgotten. Model output remains probabilistic, and new document patterns or business rules can expose weaknesses after launch.
Human review therefore becomes a design component rather than an admission of failure. The system should assign confidence, preserve source evidence and send uncertain or high-impact cases to qualified employees.
The last mile is operational, not just technical
Combining OCR, extraction and contract comparison demonstrates why useful AI often consists of several narrow capabilities layered together. The model’s raw intelligence matters less than how each step hands evidence and exceptions to the next.
Teams need guardrails for incorrect output, monitoring for changing data and a feedback loop that turns human corrections into better prompts, rules or evaluations. Progress should be measured through completed cases, error severity and review time rather than a headline automation rate.
Rose also challenges the assumption that every organisation needs another specialised SaaS product. Major cloud platforms already provide tools for building agentic workloads, and companies with an established cloud presence may be able to start without adding a new vendor.
Building internally is not free. It transfers responsibility for security, maintenance, evaluation and cost control to the enterprise. The choice should follow available expertise and the strategic importance of the workflow, not a general preference for buying or building.
Cost and portability become the next constraints
As model capabilities mature, Rose expects attention to shift toward running them efficiently on laptops, phones and other smaller systems rather than continually expanding data-centre demand.
That movement could make AI more private, responsive and affordable. A smaller model close to the data may be sufficient for extraction or classification, while a larger hosted model is reserved for harder reasoning. Portability reduces dependence on one provider and lets architecture follow the economics of each task.
Sustainable deployment requires measuring the entire pipeline. OCR, repeated model calls, retrieval, human review and failed attempts all contribute to cost. A technically impressive workflow can still be uneconomic if it consumes more compute and oversight than the manual process it replaces.
The best starting point is therefore a bounded business problem with accessible evidence and a clear baseline. Teams can tolerate imperfect data if they know which mistakes are acceptable, which require escalation and how performance will improve.
Perfect data remains a worthy goal for critical systems, but it should not become an excuse for indefinite preparation. Modern AI can help organisations discover and structure the disorder it encounters.
The last mile is reached when that capability becomes a reliable, portable and affordable operating process. Success comes not from pretending the data or model is flawless, but from building a system that can recognise imperfection and still deliver controlled value.
