Everyone in the industry is talking about artificial intelligence, but mortgage tech boss Zahid Bilgrami is more interested in what nobody is discussing, writes Kevin Rose.
As chief executive of Mortgage Brain, Bilgrami is enthusiastic about AI’s potential, but believes too many firms are rushing to adopt it without understanding the long-term consequences.
His concern is not that AI will replace advisers overnight, but that brokers, lenders and technology providers may be making decisions based on assumptions that will not hold true in a few years’ time.
“We see lots of really innovative solutions coming to market, and this is really great to see,” he says. “But I don’t think that the providers of the tech, as well as the potential customers, necessarily understand what they’re buying and the inherent risks that come with it.”
The company’s recently launched AI charter was designed to challenge that thinking. It is built around four principles – cost, intellectual property and data sovereignty, consistency, and speed and fit-for-purpose deployment – intended to provide a framework for how AI should be built, governed and deployed within mortgage advice and lending.
One of the questions the charter is designed to force firms to ask is where client data ends up. During our own conversation, Bilgrami points out that the interview is being recorded and transcribed by an AI-powered video application. There is nothing sensitive being discussed, but he says the same technology could easily be used in a broker’s conversation with a client.
“If you can imagine a mortgage broker talking to that client, and there are sensitive financial discussions happening, then that sensitive data could end up training the next version of ChatGPT, Claude or whatever else,” he argues.
For Bilgrami, that is only one example of the risks that accompany the opportunities AI presents.
COST
According to Bilgrami, human beings subconsciously assume tomorrow will look like today, and technology providers are no different. They build solutions around what’s cheap and easy right now, which gets products to market quickly but leaves the underlying economics unexamined. That gap underpins the first pillar of the charter – cost.
“If you are buying a tech product which has a third-party AI cost in it, you need to understand that’s not going to last for very long”
The major AI providers, including OpenAI, Anthropic, Google Gemini and xAI’s Grok, are offering access at very low prices, from free tiers up to around £180-£200 a month for a premium model. He argues this bears little relation to the true cost of running the technology: once data centre electricity, water and manpower are factored in, he believes current revenue covers only a fraction of what it really costs to serve customers.
“The cost to serve today is not going to be the cost to serve tomorrow,” he warns, suggesting firms should understand what happens if the underlying cost of third-party AI rises ten to thirty times over the next three to five years.
“If you are buying a tech product which has a third-party AI cost in it, you need to understand that’s not going to last for very long,” he notes.
He also raised the possibility that AI providers could turn to sponsorship or advertising to plug that gap, comparing it to Amazon’s search results, where most of the first few pages is now paid placement. Applied to AI-powered mortgage research, that could mean outputs that look neutral but are shaped by commercial arrangements.
IP AND DATA
The second pillar is intellectual property and data sovereignty, covering both customer data and the accuracy of information used by AI systems.
Bilgrami said firms need to understand where sensitive customer data is going, which jurisdiction it may end up in, and whether that’s even permitted under regulation. “If you don’t know, you need to ask – that’s the minimum you need to do. People don’t know the right questions to ask.”
“If you are reliant upon third-party AI to help power that advice, you need to be certain that it is whole of market and that it is accurate”
He argued that legal agreements with a technology provider may not be enough if the AI capability is powered by another company’s backend model. A small tech provider built on an OpenAI, Gemini or Anthropic API may leave a financial institution with a watertight agreement, but no visibility of the backend it doesn’t control. “Your legal agreement is with your counterparty and not with the party they rely on at the backend,” he explains.
The second issue relates to the information public AI models are trained on. Frontier models are updated only every few months, so to answer questions about anything current they run a live web search and stitch together whatever is publicly available – information never guaranteed to be complete, current, or ‘whole of market.’
“If you are reliant upon third-party AI to help power that advice, you need to be certain that it is whole of market and that it is accurate,” he states, questioning how any firm can be certain of that when relying on a model that cannot itself guarantee it.
CONSISTENCY
While data and governance dominate AI discussions, Bilgrami believes one of the biggest challenges for regulated financial services is more fundamental: can the technology deliver the same answer every time?
He explains that large language models are probabilistic rather than deterministic by design. Ask the same model the same prompt in two separate, clean sessions and it will give a broadly similar answer, but rarely an identical one. Push it further and, he estimates, roughly one time in a hundred it may give something noticeably different.
That variability is fine for creative work, he says, but becomes a problem the moment AI is embedded in a process shaping customer outcomes. “If there are 100 clones of you giving 100 customers advice, you don’t want that advice to be different one in 100 times,” he says.
Much of that unpredictability, he explains, comes down to a setting called “temperature” – a dial controlling how much randomness a model injects into its output. Firms running their own AI can turn that dial down for consistency. Those relying on a third-party frontier model have far less control over it, and providers who claim consistency can be achieved through prompting alone, he argues, are offering no real guarantee.
He also highlighted the issue of model changes. If a provider is reliant on a third-party model, every update can change the system’s behaviour, since newer versions may respond differently to prompts that previously worked reliably.
“Every time they change their model, I need to do a new baseline, because all the stuff that I did previously may work differently with this new model,” he warns.
SPEED AND SUITABILITY
The final pillar is speed, or more accurately, choosing the right tool for the right task.
“Faster is not always better, and bigger quite definitely is not always better”
Bilgrami said the frontier labs’ flagship reasoning models are formidably capable – fast, built on hundreds of billions of parameters, able to produce something close to a PhD-level answer on demand. But that scale isn’t always what a use case needs, and can work against firms on cost and consistency as well as performance.
“Faster is not always better, and bigger quite definitely is not always better,” he maintains. A smaller, more tightly controlled model, trained appropriately for the task, is often the better fit for financial services than a general-purpose model built to answer humanity’s hardest questions. “We feel that everyone’s going right, and we’re going left,” he says of Mortgage Brain’s own approach.
AI AND THE BROKER MARKET
Despite the warnings, Bilgrami is not arguing brokers should avoid AI. He believes the opportunities are significant, but that the broker market needs to be viewed in segments.
Smaller directly authorised (DA) firms, often one or two-person operations already using tools like Microsoft 365 Copilot or Google Gemini, may use AI for productivity, admin and customer communication. Larger corporates and networks may use it more extensively, including for compliance, lead generation and adviser support.
He accepts many smaller brokers may not know what to ask and suggests they can learn a lot by watching what peers do on YouTube, following trade press, and – half in jest – using ChatGPT itself to work out the right questions to ask a supplier.
Whatever the size of firm, the consequences of getting it wrong scale sharply. A small DA that mishandles AI faces an operational headache, while a network or lender that leaks client data or embeds a flawed AI process at scale faces the kind of reputational and legal exposure seen in past mis-selling scandals.
“The implications of getting it wrong are dramatically different if you are a small DA compared with a network, larger corporate or lender,” he insists.
THE HUMAN ROLE IN MORTGAGE ADVICE
Asked whether AI could fundamentally change the role of the broker, Bilgrami said the answer depends heavily on consumer behaviour and regulation – two variables he says are impossible to forecast with confidence.
He believes there remains a need for an independent layer between product manufacturers and consumers, whether human or synthetic. In general insurance, that role has shifted to price comparison websites. Mortgages, he argues, are different, given the complexity of an individual customer’s finances and the property used as collateral.
“I think that intermediation layer is not going to go away within mortgages,” he asserts.
He acknowledged that some parts of the market, such as product transfers, could become more direct, but said consumers should still research the wider market before deciding – and beyond that, he is cautious about predicting further, noting it has been less than two decades since the iPhone reshaped how people live.
AN INDUSTRY-WIDE DEBATE
Bilgrami accepts that an industry-wide AI charter may be difficult to achieve. He likens the ambition to Isaac Asimov’s fictional Laws of Robotics – a neat starting point for a debate about ethics, even if reality is messier than fiction. Getting even a small, aligned group of major lenders to agree shared principles is hard enough, he suggests, let alone the whole market. So rather than chase that agreement directly, Mortgage Brain’s approach is to publish its own principles and let others adopt, adapt or ignore them.
“The best way in which I can influence the industry is by taking our principles and making them public”
“The best way in which I can influence the industry is by taking our principles and making them public,” he explains, candid that the charter began as an internal exercise rather than a marketing campaign: “It is inherently something that we are doing internally.”
The charter is intended to give Mortgage Brain a clear set of principles to guide how it develops and uses AI. Externally, the aim is to help the wider market ask better questions and move beyond the hype. Not to slow adoption, but to encourage a more informed debate.
“I’m really passionate about this technology. I’m really passionate about helping people understand this technology,” he concludes, describing himself as equally keen to demystify it – to help people understand, in plain terms, what a large language model actually is, rather than an unexplainable black box.
Ultimately, he hopes the charter encourages brokers, lenders and technology providers to look beyond demonstrations and productivity gains, and think more carefully about governance, reliability and long-term consequences.
In an industry increasingly captivated by AI, Bilgrami’s message is measured: embrace the opportunities, but understand the questions before you accept the answers.
