Published on the 16/09/2026 | Written by Heather Wright
Token shocks, data work, getting ROI and operationalising AI….
“We used three quarters of our allocation of tokens for a quarter,” Annette Rangi admits of an early AI project at HEB Construction.
It’s not the kind of KPI most organisations put in their AI success stories, but for Rangi, HEB Construction head of digital transformation, the experience was a turning point. What began as an AI-powered resume project at the Kiwi civil infrastructure and engineering company became “a really good learning exercise” – and one that has resulted in a repeatable framework designed to minimise token consumption, measure value and provide a foundation for future AI projects.
“We used three quarters of an allocation of tokens for a quarter and now we’ve reduced it down to a couple each time we do the call.”
HEB, which is a subsidiary of France’s Vinci Construction Group, has multiple AI initiatives underway across contracts, HR, onboarding and infrastructure operations as it builds the governance, monitoring and data foundations needed for broader adoption.
The project that provided so many of those lessons began with a relatively mundane problem.
For decades, tender teams had manually assembled bios and resumes for bid submissions. Over more than 50 years in business, that meant multiple versions of employee records had accumulated across the organisation, making it difficult to easily identify the right people, qualifications and expertise for individual tenders.
The company implemented Boomi Enteprise Platform to build two AI agents. Shredder, deconstructs CVs and creates a structured representation of the person’s skills in an easily accessible and reusable format, while Bob is a conversational agent that retrieves appropriate profiles and generates a tailored CV or short bio for tender teams.
“It was a really good opportunity to clean up the process of how we manage our resumes and automate the data and have the ability to prompt and retrieve that data efficiently and with a high level of data integrity to be able to respond to tenders with the appropriate people that had those technical ability or that experience,” Rangi told iStart.
The project is part of a broader technology strategy built around Microsoft technologies including Fabric, Power BI, SharePoint and Teams, alongside Snowflake for data warehousing and Boomi for integration.
“We are wanting to use AI across any of the applications we have internally,” Rangi says.
HEB’s AI approach in New Zealand is centred on low-code and no-code development with support available from a wider pool of 88 data scientists at Vinci.
Everything starts with data
While the resume project delivered a practical use case, it also reinforced a lesson Rangi says applies to virtually every AI initiative.
“Everything starts with data.”
Construction is, by its nature, data intensive. Drone footage, compliance records, environmental monitoring, project information and health and safety reporting generate vast amounts of information. The challenge is less about collecting data than ensuring it is available in a form that can be trusted and used.
“Our ability to have our data in a place that we can actually use automation and AI tooling to automate and optimise processes is significant,” Rangi says. “It’s about optimising information to make timely decisions and freeing up our people to actually do the things that they need to do, rather than sitting in front of their devices trying to do reporting – which wouldn’t be their forte in a lot of cases.”
But while data is foundational, Rangi doesn’t believe organisations need to wait for a perfectly curated data estate before pursuing AI.
“You can’t clean up everything, it’s too big,” she says.
Instead, HEB has focused on identifying the information employees access most frequently and making that data available in a governed, usable format. As part of its migration to SharePoint, the company has been analysing how information is used across the business and prioritising the data most likely to deliver value.
“We’re working across the business to identify what data is accessed more frequently and enabling that to be turned into a data pipeline so our business can self-help and know the data is clean.
“For example, we use SAP for our ERP system and we’re using SAP Analytics Cloud and developing that to enable our business to safely access and do dashboards, but also we are starting to test Claude because the ability to call up dashboards on the fly instead of having to train people to use Power BI and get access to data quickly is something that we’re looking at as well.
“Technology moves so rapidly it’s not a one-stop shop and we might take an approach and then change that, but our our strategy is to enable our business to have access to data in a real, efficient and effective way and making sure the data is clean.”
The resume project highlighted the data challenge early. Before AI agents could retrieve information effectively, HEB first had to standardise how employee skills, qualifications and experience were represented. The company created a common template and loaded the information into Snowflake, creating a consistent foundation for the agents.
Token shock
The bigger lesson emerged after deployment.
As users began experimenting with prompts, token consumption escalated rapidly.
“We soon experienced quite a lot of tokens being used for the prompting,” Rangi says.
The experience forced the team to revisit how information was structured and how requests were made. HEB refined the way data was organised and introduced more targeted prompting, reducing the amount of work the AI models needed to perform.
“We made the calls a lot smarter so that it could be more specific when they did their prompt plan,” she says.
The result was a significant reduction in token consumption – from that initial comment of using three-quarters of the allocation of tokens for a quarter, it’s now been reduced to “a couple” each time a call is made.
Perhaps more importantly, it has also resulted in a repeatable approach HEB believes can be applied to future AI projects.
“The way that we’ve done it is a repeatable way, so that framework or process that we’ve built, we can now look at reusing for other operational-type activity that we do quite regularly,” Rangi says.
Those opportunities are already emerging with “five or six” AI agents in testing or review and one to three AI-related initiatives worked on each month. Projects range from contract analysis and employee onboarding through to systems designed to help identify and forecast potholes on roads.
As AI activity has increased, so too has the need for governance and visibility.
HEB has established monitoring and alerting through Power BI dashboards to track token consumption across projects. The company is also investigating ways to provide visibility across AI use regardless of whether tools are built in Copilot, Claude, Boomi or other platforms.
“For each agent we build, we’ll have it in Power BI, and we can monitor those and have alerting,” she says.
“It’s given us the basis for having a very robust management solution across our agent use and visibility across our projects,” she says of the early learnings.
The company is also requiring business units to identify, with the tech team, expected returns before projects begin and agree on how those outcomes will be measured.
“A lot of AI initiatives don’t get out of test or development or get past the idea or concept because everyone wants to do AI so we’re doing a lot of validation upfront and ensuring that we identify how we can capture value.”
The value isn’t always measured in financial terms. Sometimes it’s time savings. In others, it may be a better experience for employees or easier access to information. The key, she says, is understanding what success looks like before development starts.
As HEB expands its use of AI across contracts, onboarding, asset management and other parts of the business, Rangi says the key lessons from the resume project have remained remarkably consistent.
Start with a clearly defined problem. Be realistic about your data. Measure outcomes from the outset. And don’t try to transform everything at once.
“Start with data, start small, support the journey by training,” she says.
