nPlan's system uses deep learning to model an owner's way of handling projects using data from its previous projects. Having "learned" the owner’s work processes, the model reviews its current project schedules to generate forecasts of outcomes with risk profiles of planned activities.

The company provides "constructive criticism at an actionable level," says Amratia. "We go all the way down to the most granular detail the project makes available," he adds, citing submission of drawings for approval as an example.

A major new nPlan user is an emerging multibillion dollar program to upgrade 76 miles of railroad across Northern England between Manchester and York, encompassing 23 stations, 285 bridges and viaducts and around 6 miles of tunnels. The TransPennine Route Upgrade team could previously review individual projects on the multi-year program only once a month and the whole program at three to six-month intervals, according to nPlan.

"Being able to get more frequent analysis done on a larger volume of schedules is a game-changer," notes Richard Palczynski, the project's head of strategic program controls. "We simply don’t have the resources to use QSRA to generate the insights the program needs to stay on track."

Since there could be "10,000 or more activities" in a large project, QSRA is "cumbersome" and prone to oversimplification, says Amratia. Innate bias among project planners is also a problem, he believes. "You only know what you know ... you will only bring that forward to the table [and] you are eternally optimistic," he says. 

And there is always a risk of focusing on "big sexy things" at the expense of less exciting but important elements. "Overwhelmingly what goes wrong on projects are the ... seemingly mundane activities."

Subjectivity in assessing risk is "very difficult to avoid," agrees Paul Miller, head of risk for transportation at the project management consultant Faithful and Gould, a unit of SNC Lavalin Group. "In an ideal world, you'd run a workshop with representatives of the project team present so you’d have a mixed view," he says. Even then, "there can be an element of group think. If a project had just had a major issue occur, that tends to steer your view that [such an event] could happen again."

And QSRA's time demand is significant, adds Miller. A major analysis could last two to three months with a full-time senior manager plus an assistant assistant, he says. Once a project schedule has been simplified for modeling, risk reviews would require a series of workshops, some with 30 to 50 people, as well as one-to-one sessions.

Recognizing the limitations of QSRA, Faithful and Gould began rolling out its own digital tool for trials late last year, says Lisa Silander, the company's UK technical lead for project controls. "Using historical data it will predict the outcomes of a project," she adds. A trial review of a major job predicted 80% of the risks that actually emerged, she says. 

Silander believes the tool will eventually replace QSRA, but workshops will still be needed to quantify and mitigate the identified risks, she adds.

Amratia says he was driven towards AI not just to build a better QSRA tool but to improve investor trust in major projects. Having seen first hand the failings of current methods in the oil refinery construction sector, he merged his management experience with the AI skills Alan Mosca, now chief technical officer, to found the company in 2017.

The pair retain a controlling interest in nPlan, whose outside investors also include a Google affiliate and Scotland-based Pentech Ventures LLP. On making its investment in nPlan three years ago, Pentech hoped future projects "having an nPlan certainty ‘seal of approval’ would be more attractive to finance than one that did not have their validation."

Amratia believes that investors are now skeptical about rosy performance projections of project promoters. So, "by quantifying your risk in a systemic way ... you actually introduce trust ... which will lead to superior outcomes in projects [and] investor confidence into the execution of the built environment." 

Faithful and Gould's Silander agrees that change is needed. "We keep seeing these major projects overrunning and overspending. With all the best will in the world, humans have got biases that we bring to every project."

Bent Flyvbjerg, professor of major program management at Oxford University's Saïd Business School agrees that "high-quality data that replace assumptions and expert judgement help de-bias projects." 

However, "we have also seen that the use of historic data in forecasting ... make leaders wonder ... how can we outperform history?" he says. While it's prudent to plan conservatively, "we also need ambition to drive innovation, disruption, and productivity in projects to beat the historical track record."