Building Intelligence: Harnessing AI Responsibly

Oliver Fisk, Chief Information Officer (CIO) at calfordseaden, reflects on how a multi-disciplinary consultancy is adopting Artificial Intelligence (AI): balancing innovation with governance, commercial realism, and long-term responsibility.

A Deliberate Approach

The built environment is inherently long-term. Decisions made today shape safety, performance and quality of life for decades. As we celebrate our 85th year, our focus remains unchanged: professional, personal service delivered by people our clients trust. The challenge is how we harness emerging technologies such as artificial intelligence without undermining that trust or introducing unnecessary risk.

As CIO, I lead an IT function embedded in the business and am responsible not only for secure, dependable systems, but for innovation that delivers competitive advantage. While techniques such as machine learning have been used in engineering and analysis for years, the rapid advance of Large Language Models (LLMs) demands a different level of scrutiny.

Data, Security, and Governance

As an RICS‑regulated consultancy working across residential, commercial, and infrastructure projects – governance comes first. We ensure we know where AI is being used, for what purpose and on which projects, enabling transparency and assurance for clients, partners, and regulators.

Digital sovereignty underpins this approach. We have never outsourced our core data stores and retain full control over how project information, correspondence, and records are stored and transmitted. In a sector where responsibility, auditability, and traceability matter, this is not optional.

LLMs embedded into everyday tools can blur boundaries between datasets. In a typical enterprise environment, AI assistants can present confidential material in inappropriate contexts not because permissions are wrong, but because they are technically valid. AI does not understand commercial sensitivity, professional judgement or contractual boundaries unless explicitly constrained.

These risks increase when considering autonomous agents. Agents operate within a user’s security context, with access to everything that individual can see and do. Their actions can be indistinguishable from those of the account holder for audit purposes. In a regulated, safety‑critical industry, that combination of power and opacity presents a material risk. Until these limitations are addressed, we believe agents should only operate within tightly controlled environments, using constrained permissions and auditable service accounts.

In an environment of rapid technological change, expectations change quickly too. Countering the cultivated ‘fear of missing out’ generated by multibillion dollar companies’ marketing budgets is hard enough, but second order effects abound too. Not long ago, no one would think to bring a random application or script they “found” on the internet and attempt to run it inside the network. Yet a new normal is positioned where non-technical users “write” code they cannot read using an LLM in order to automate or improve some process. In these instances, to support this creativity safely we have invested time in code review and provided sandboxed testing environments beyond the normal developer job role, but security comes first.

Commercial and Budgetary

In the great gold rush of AI no one can fail to notice how every product is getting an LLM shoehorned into it, which is only slightly better than the prior status quo which was an AI badge on the back and an IF statement or two under the bonnet. The more pernicious current commercial threat has tended to come from “AI first” new products where the ease of access of cloud software as a service proposition makes serving or marketing to niche use cases a very high margin opportunity.

Only a small number of organisations operate frontier LLMs. Most AI‑powered tools used in construction and consultancy are wrappers, thin interfaces layered over models owned by third parties. Understanding how much proprietary value a vendor adds beyond the base model is critical.

We have seen both ends of this spectrum. For instances where the use of AI is explicitly permitted to support the research of bids and proposals, we were pitched products to assist us. The interface was compelling, but the underlying capability relied primarily on our own data (free) and a non-proprietary LLM. By building a retrieval‑augmented generation (RAG) pipeline anchored in our project and bid history, we replicated the core functionality at a fraction of the cost; under £1,000 per year compared with a quoted £50,000 annual licence cost.

By contrast, we happily invest in AI‑enabled legal and contract review tools that combine LLMs with curated construction‑specific content, templates and embedded expertise. In those cases, the value added is real and directly supports risk management and commercial decision‑making.

Tools, Not Toys

More and more businesses are recognising the risk of indulging in “AI Theatre”, spending time and effort producing or procuring products which work but don’t add to the bottom line. That are independently efficient, but not effective in helping the organisation reach its goals. For this reason, we have taken a focused approach, testing the return of investment of proposals and measuring post deployment outcomes carefully.

As has been observed elsewhere, “the market for feeling productive is orders of magnitude larger than the market for being productive.” It is easy to build elaborate AI workflows automated summaries, agent chains, and dashboards without adding a penny to the bottom line.

I recently met an established CIO of a large company, pushed by his board to make the organisation more “AI enabled” who contended that he could prove he had achieved his aims through the evidence of co-pilot engagement reports. What were people using co-pilot for? He did not know.

Cost models matter too. AI tools priced per user can become expensive when adoption or even required use is sporadic. We have deliberately developed internal tools priced per use, achieving outputs comparable to commercial offerings while avoiding subscription sprawl and external data silos. In one case, we have been able to replicate the core functionality of commercial meeting note takers at 1p per hour vs £30 per user per month.

Looking Ahead

Our approach remains consistent with our values: protect the security and integrity of project and client data, adopt commercial AI where it demonstrably adds value, and build bespoke solutions where that delivers a better fit for our business and our clients.

In a sector defined by regulation, safety and long‑term responsibility, the challenge is not whether to adopt AI, but how to do so intelligently. That requires detailed understanding of both emerging technology and the environments in which it is applied.

While artificial intelligence may assist in building some of these solutions, it is the authentic intelligence of our people across calfordseaden that ensures technology enhances, rather than replaces, professional judgement.