Behind the Build: Building a governed, source-backed AI research mode on the Irish Tax Institute's existing platform

An AI research aid, a rebuilt advanced search and a modern interface, added to a trusted tax research database without disturbing the decade-old foundations that already worked. 

 Client: Irish Tax Institute | Sector: Professional body, tax research and training | Engagement: Discovery, delivery, ongoing support 

Immediate outcomes

ing contractors

At a glance

Who ITI are

The Irish Tax Institute is a professional body in the Republic of Ireland, providing training and information about tax law to the people who need to know it: accountants, tax advisers, company leaders and Revenue librarians. It also runs professional certifications, so a significant share of its users are students rather than practitioners, and a trainee wants something different from the same software than a practitioner with thirty years behind them.

TaxFind is ITI's Irish tax research database, and 67 Bricks built it. Irish tax legislation arrives as structured XML, is stored in a MarkLogic database and served through a Scala application, alongside hundreds of pieces of commentary and guidance written by ITI's own experts. The law is one half of the value; what a tax adviser is supposed to do about it is the other.

The problem

Nothing was broken, and that was the point 

There was no failure to fix here. The platform 67 Bricks built more than a decade earlier was still doing its job, the ingestion process ITI's editorial team used to publish new legislation was working well, and no internal team was struggling. This was not a case for re-engineering. 

What had aged was the surface and the search. The site had drifted out of step with ITI's corporate branding and was starting to look dated. Underneath that sat a shift in what users assume a research tool can do: conversational, natural language querying had become the default expectation, rather than knowing the precise phrase to search for. 

 "A tech stack from 2015 works well, so that wasn't the problem we were trying to solve. We kept the stuff that works, and sometimes that's the right thing to do."   —  Stephen Nichols, Technical Lead, 67 Bricks

Search that made members work too hard

The substantive problem sat with end users rather than with ITI's staff. TaxFind holds a large volume of content, much of it legislation referenced by number, and members could not combine filters: if you wanted legislation, expert commentary and Revenue guidance in the same result set, the platform would not give it to you. An advanced search existed but was difficult enough to use that people avoided it. Some members were running searches outside the platform and returning to TaxFind to retrieve what they had found. Separately, small editorial and commercial jobs, showcasing the authors behind the content or running promotions, all had to come through 67 Bricks under the support contract.

Why 67 Bricks

 ITI had worked with 67 Bricks since around 2014 and came back to them, which is the short version. The longer version is that the specialism fitted. 67 Bricks makes sense of and surfaces the right data from large volumes of information, so the work would have been in scope even from a cold start. Knowing the system already removed real cost: nobody needed briefing on how TaxFind weighted content types or how legislation was structured, and the colleagues who wrote the original system were still there where knowledge had aged out. The team delivering this phase was deliberately not the team running TaxFind support, so knowledge was cross-pollinated rather than the work bottlenecking on whoever had built it first. 

The solution

Discovery first, and a recommendation to the board 

At the start of 2025, 67 Bricks ran a discovery project: user interviews and an investigation into where TaxFind could usefully go next. It identified user needs in the platform and the strategic areas ITI wanted to explore, and produced recommendations that went to ITI and their board. Those recommendations shaped the delivery phase that followed, and the principle carried into the build: confirm the problem is real before writing code to solve it.

Evolution, not revolution 

The 2014 architecture was sound, so the work built on it. The existing MarkLogic database became the foundation for natural language search, with a new layer added to let large language models interrogate the content. This is a point 67 Bricks makes to clients repeatedly, and TaxFind demonstrates it: the quality of the AI depends on the quality of the structure underneath it. 

Thirty seconds, and no hallucinations 

The AI search was new territory. Nobody knew in advance what members would use it for, or how to make the answers trustworthy. What made it tractable was that ITI set a hard, measurable brief. 

"The most important instruction we got from the customer was that the AI search needed to take less than 30 seconds and have no hallucinations in. That was a very clear goal, and something we can measure and monitor against."   —  Stephen Nichols, Technical Lead, 67 Bricks

Both halves are engineered rather than hoped for. Search times and answer quality are recorded, and human and automated testing were used to confirm the search was finding the relevant material and summarising it correctly. The reference handling is where the no-hallucination requirement is enforced: the model is not allowed to invent citations. Every reference is a hyperlink to content that already exists in TaxFind, and the system verifies the content is really there before the reference is shown. The AI's job is to get a professional to the source faster, not to stand in for their judgement. 

A refresh tested with real users, and tools handed back 

The interface was modernised and brought into line with ITI's branding, with extensive user testing throughout by ITI's in-house testers and external users recruited to check the direction. Author profiles and the management of on-platform ads and promotions were built as self-service tools, on the principle that anything ITI's editorial team can sensibly run themselves should not need to come through 67 Bricks.

ITI's editorial and project teams were brought inside the delivery process rather than reviewing it from outside: sprint reviews, planning, test recruitment and their own QA layer on top of 67 Bricks' testing. Because ITI run training and user support for TaxFind themselves, that also left them ready to support their members from launch.

Managing risk: protecting trusted content 

TaxFind's value rests on the quality of what sits inside it, so the risks here were about the content rather than the architecture. Adding large language models to a content platform raises an immediate question about where that content ends up, and TaxFind is hosted so that its content cannot be fed back into large language models as training data. This was not an existing gap being closed; it is a condition 67 Bricks now works to whenever it builds natural language search over a client's data. The second risk was availability: aggressive scraping by bots serving language models and search engines is a fact of life for any content platform, and members sit exams and do billable research on this one, so it needed robust protection against high-volume automated traffic.

Against both, the answer was conventional technology used well: MarkLogic, Scala, modern language model tooling, cloud hosting, security best practice. Nothing proprietary, nothing mystifying, and a competent development team could pick the system up. 67 Bricks continues to support TaxFind, but that is a continuation of the relationship rather than a dependency created by it.

Outcomes

This was the biggest change made to TaxFind since 67 Bricks built it, covering both the search service and the look and feel of the platform, and it landed quietly. The launch went to ITI's own timeline and went smoothly.

In the first month after release, AI mode usage was healthy and appeared to be producing the right output, and the site was robust under it. The clearest signal was the absence of noise: very few members rang to complain, which for a platform professionals use daily is a result in itself.

ITI's own headline measure is membership renewals, which fall at the end of August. Those figures will be the real test of whether the investment has paid for itself.

"The nice thing is that we didn't have to make tons of technological changes to the underlying architecture, because that was sound."  —  Rachael Lammey, Strategic Technology Consultant, 67 Bricks 

Why the partnership worked

• A decade of relationship, and a client who came back by choice rather than lock-in. 

• A decision to keep what worked, so ten years of sound architecture became the foundation for new technology rather than a sunk cost to be written off. 
  

• A discovery phase that reached ITI's board as recommendations, so delivery was shaped by evidence about users rather than a feature list. 
  

• A hard, measurable client brief, under 30 seconds and no hallucinations, enforced in the reference-handling code rather than promised in a specification. 
  

• ITI's team inside the delivery process, which also left them ready to support their members from launch. 
  

• Industry-standard, documented technology throughout, with no hidden knowledge and no permanent dependency on 67 Bricks. 

Project credits