A review of the strengths & limitations of Construction AI

1. Tireless automation of repetitive tasks: from manually completing reports and forms to image recognition detecting defects, our eyes get tired but an AI's doesn't. A good way to analyse if you are facing a problem which AI can solve is to see is the task you want to automate within a repeatable process with a set of rules or approvals which need to be achieved before passing to the next step, which can usually be solved by reading or visually checking the status of something. If this is the case, then it is usually a solveable AI problem. The next question is how complicated are the rules and approvals which need to be followed and how frequently does the task come up, to decide if it is cost effective to automate using AI.
 

2. Search capabilities and options analysis: Early Chess computers like 'Deep blue' invented in 1995 proved that computers can search and analyse a lot more possibilities than humans ever could. To continue with the chess analogy, a human novice player can analyse 1-2 moves deep, a human GrandMaster is capable of analysing 5-8 moves deep, whereas a modern Chess Engine such as Stockfish can analyse 20-30 moves deep in some positions. This same principle can allow AI applications to analyse potential outcomes on a construction project better than human counterparts and retrieve information from project archives which a human could not find by searching more effienctly than we can.
 

3. Filling in staff shortages - Construction has a skills shortage with many companies reporting that they have listed vacancies for more than 12 months, automation can help us to work more productively with the resources we have. This problem is multiplied by many countries facing challenges around managing aging infrastructure, stretching resources more thinly and pushing us to work more productively. AI automation can allow us to create draft outlines of our work, write minutes and summaries, find information for us and save our time to be spent on higher value work.
 

4. Cutting through messy and unstructured information - In construction we work with lots of archived PDF’s and messy folder structures with many versions of the same document. AI can help us cut through the noise and bring structure using methods like semantic similarity, which we have implemented at Civils.ai. This allows you to search through 1000's of construction documents all at once to retrieve information and answer construction questions instantly.
 

1. Difficulties in communicating uncertainty - Large Language Model’s in particular have faced challenges around expressing uncertainty. This is especially a case since AI developers want to provide succinct answers to reduce their cost in running the LLM and users want succinct answers to save time and find information quicker
 

2. Over-expectation - Linked to the first limitation, in many cases AI applications take time to fine-tune to certain use cases and are rarely out of the box solutions. This can often mean that expectations of users need to be managed and clear milestones need to be reached on smaller samples or datasets before deploying live on the project.
 

3. Specialist applications - In the construction industry we engage many subcontractors or specialists to complete a specific set of tasks. Given expensive costs associated training AI models and a limited amount of public data being available for AI training, this often means that users of AI applications within specialists fields find the results of the application underwhelming.
 

4. Barriers to entry - Many companies do not have their own internal data scientist teams to develop and train their own models and hiring talented data scientists is costly and difficult. This can lead to in-house developed prototypes falling short of expectations.
 

5. Privacy concerns - Many construction projects use sensitive company data meaning that additional measures need to be taken such as running the AI application on the construction companies own cloud environment to ensure data is not sent to external 3rd parties or anonymisation of any data being sent externally to 3rd parties.
 

6. Silo'd data - Ultimately all AI applications require data and in construction we have lots of it, but it is nearly always silo’d or locked away in our PDF documents. This is the reason why at Civils.ai we’ve spent the past two years developing tools to help you unlock all the data in your construction documents.