One real AI project, told properly. What we built, what went wrong, what fixed it, and what we would want you to know before you start your own.
A client drowning in tender documents wanted three things out of them, quickly enough to decide whether a bid was worth chasing at all.
Pull out the specific things a bid team needs, milestone dates and commercial detail, rather than producing another generic summary nobody reads.
Score each opportunity against the client's own criteria so the pipeline could be prioritised, and a bid or no-bid call made early.
Separate the front-end administrative questions from the deeper technical ones that need a subject expert to sit down and answer.
None of this appears in a vendor case study. All of it is normal, and all of it costs time and money if you meet it for the first time halfway through a build.
Every one of those problems was solved with ordinary engineering discipline rather than a better model. Careful implementation, validation and human oversight are what turned a promising demo into something a bid team could rely on.
Three techniques did most of the work. None of them is exotic, and all of them are worth budgeting for at the start rather than discovering halfway through.
Process the document in segments and merge the extracted results, with a moving window across the text so context is not lost at the seams.
Run the extraction several times and combine the outputs, excluding the outliers automatically. Three passes in the eighties and nineties collated to a materially better answer than any single run.
The client marked up a set of perfect results. Every change was then checked against that baseline, using a mix of AI and plain regular expressions to validate. Without it you cannot tell an improvement from a regression.
Not all of these bit us on this project. All of them are worth a conversation before you put an AI system anywhere near a client, a regulator or a tender response.
Where your documents go, and who trains on them.
What the model was trained on, and what you now own.
Inherited from the training data rather than from you.
Confident, fluent, and wrong.
The same input need not give the same output.
A document that instructs the model reading it.
It does not know what happened after training.
Improving, still not arithmetic or formal logic.
All of the above, in front of a client.
Every case study we have run hit non-trivial challenges. Careful implementation, validation and oversight are the job, and not an afterthought.
Better accuracy and lower costs have taken the edge off the early problems. Progress may be plateauing, which means smarter engineering rather than simply waiting for a better model.
The best results come from AI augmenting people rather than replacing them. Explainability, validation and control are what make that work.
As the industry moves towards agentic AI, hallucinations, errors and unpredictability matter more rather than less. Governance, safeguards and human oversight become more important as systems become more autonomous.
Adoption should be driven by a real business need rather than by the hype. Cost, feasibility and return belong in the conversation before implementation.
We can now solve problems that were not solvable a few years ago, processing complex data at a scale and cost that simply was not available before.
Thirty minutes, no charge. Tell us where your pilot is stuck and we will tell you what we think, based on having been stuck there ourselves.
We build AI and data systems into construction, infrastructure and property businesses, and our engineers work inside your systems for as long as that is useful and no longer. We will tell you when you do not need us.
Alex will come back to you within a working day.