The Intermediary – September 2026 - Flipbook - Page 102
T E C H N O L O GY
Opinion
The bigger prize
may be what AI
can remove
F
or all the discussion about
artificial intelligence (AI)
transforming financial
services, one of the more
interesting applications
may turn out to be
solving a problem lenders have been
trying to solve for decades.
Mortgage origination has become
progressively more digital, yet beneath
many apparently digital journeys
remains a surprisingly analogue
problem. Customers, brokers and
lenders assemble information
from payslips, bank statements,
identification documents, application
forms and numerous other sources.
Systems capture some of it digitally,
documents provide evidence for some
of it and people spend an enormous
amount of time establishing whether
the two agree.
Many lenders have digitised the
process without necessarily removing
the work meaning there is now
an obvious point for AI to deliver
genuinely transformational rather
than simply fashionable opportunity.
Our Collect & Conclude technology
has roots stretching back 15 years, and
has processed millions of documents.
The significance of introducing AI
into that environment is therefore
not that a large language model
(LLM) has suddenly been pointed at
a mortgage application. It is that AI
is being inserted into an established
architecture that already understands
lending documents, data collection
and validation at scale.
For a CTO, that is probably the
more interesting proposition. The
underlying architecture separates the
problem into three activities: collect,
extract and validate.
Our agent, DocStreet, handles data
collection, DTA2 extracts information
and Orbis applies centralised business
100
The Intermediary | September 2026
rules and validation. The system can
bring together physical documents,
application and back-office data and
information obtained from trusted
external sources, then turn that
material into structured data and
validation results.
AI changes what can happen within
that architecture rather than replacing
the architecture itself.
Complex documents
Historically, extracting information
reliably from documents meant
creating increasingly sophisticated
rules. If a salary appeared here,
labelled in this way, in this particular
document format, the system could
be taught how to find it. That works
extremely well when documents
are predictable, but becomes
expensive and cumbersome as
formats proliferate.
LLMs invert part of that problem
because they can understand what
information means rather than
relying entirely upon where it
appears. The temptation would be to
conclude that AI could replace rules
but that is not the case in lending
where rules remain extremely
important where precision,
consistency and control maer.
LLMs are beer where information
varies and understanding that context
maers. Collect & Conclude combines
the two, using AI where flexibility
is valuable and deterministic rules
where control is essential. The result
is a process that feels like a much
more grown up definition of AI in
financial services.
The objective is not autonomous
lending but the creation of
an increasingly autonomous
administrative layer around lending.
Consider what happens when
a document enters a traditional
JERRY MULLE
is UK managing director
at Ohpen
process. Somebody or something
identifies it. Information is extracted.
That information is compared with
the application. Other documents
may need to be checked against it.
Differences have to be identified. Rules
are applied and exceptions passed to
somebody capable of resolving them.
None of those activities represents
the judgement for which experienced
mortgage professionals are employed,
yet collectively they consume
an extraordinary amount of
operational capacity.
Ohpen already has configurations
covering more than 150 lending and
mortgage document types and Orbis
contains hundreds of document and
cross check validation rules. The
opportunity created by AI is to make
that infrastructure considerably easier
to configure, maintain and improve.
The next generation of the
technology introduces an AI agent
which continuously monitors
extraction performance, analyses
behaviour against real documents,
and benchmarks the results before
improvements are reviewed.
Separately, self-service configuration
allows users to select document
types, fields and validations while
an AI agent generates the underlying
configuration without requiring code.
None of this removes the need for
governance. The architecture maers
precisely because AI is being contained
within controls rather than being
treated as the control.
The first phase of Ai in financial
services has been dominated by what
AI can create. For lenders, the bigger
prize may be what AI can remove. ●
