Hello to Canada’s SaaS and AI Community,
AI can produce an answer in seconds, but how much time does your customer spend checking it?
For Aaron Harris, CTO at Sage, that question gets to the heart of AI adoption. Speaking with SAAS NORTH AI, Aaron shared why trust is becoming a competitive advantage, how verification eats into productivity gains and what founders can do to build AI experiences customers are willing to rely on.
Key takeaways:
- Trust remains a key barrier to AI adoption.
- Sage and IDC found 26% of AI productivity gains are lost to verification.
- Transparency and user-controlled automation build confidence.
- Trusted experiences could set SaaS companies apart.
Co-Founder/Producer, SAAS NORTH Conference Editor, SAAS NORTH NOW
The Hidden Cost of Checking AI’s Work
While AI has made generating answers easier than ever, convincing customers to trust and act on them remains a greater challenge.
For Aaron Harris, CTO at Sage, the problem is particularly apparent in financial software, where customers need to know that the information they’re relying on is correct.
“Trust is the biggest barrier to adoption,” he says.
That hesitation has a measurable cost. Research from Sage and IDC found that 26% of productivity gains from AI are lost to reviewing and verifying its output.
Consider a finance team using AI to process transactions. The technology might complete the initial work in seconds, but if employees must independently check every figure, a substantial portion of the time saved disappears.
“The name of the game now is reducing that 26% as close to zero as possible,” Aaron explains.
For SaaS founders, this changes how productivity should be measured. The time an AI feature saves is only part of the equation. The time customers spend checking its work matters too.
Reducing that burden requires product experiences that make answers easier to understand, verify and act upon.
A Correct Answer Can Still Be The Wrong Experience
When Aaron tested an early version of Sage Copilot, he asked what should have been a straightforward question: What is my cash balance?
The answer came back almost immediately: zero.
Aaron knew something wasn’t right. After investigating, he discovered that Copilot had interpreted cash as petty cash, rather than the combined balance across his bank accounts.
The answer was technically correct, but it was of little use to someone trying to understand their financial position.
The experience exposed a problem that goes beyond model accuracy. Customers need enough context to establish whether an AI-generated answer addresses the question they intended to ask.
Sage subsequently changed how Copilot presented financial information. Instead of returning a single figure, it provided the total cash balance, individual account balances and direct links to the underlying records.
Users could trace the answer to its source without conducting a separate investigation.
Aaron calls this approach glassbox AI, giving customers visibility into the information behind an answer rather than expecting them to accept it at face value.
“AI will never be 100% accurate. It’s the nature of the technology,” he explains.
That’s particularly consequential in accounting and finance, where discrepancies can affect reporting, payments and business decisions. Aaron argues that accurate outcomes depend on AI working alongside supporting systems and human oversight.
For SaaS founders, the practical lesson is to build verification into the feature itself. Source links, account-level breakdowns and clear explanations of ambiguous results can help customers investigate discrepancies without leaving the product.
The easier an answer is to verify, the less work customers have to do before using it.
Give Customers Control Over Automation
Once customers can verify AI’s work, the next question is how much responsibility they’re comfortable delegating.
Aaron describes this as a journey of trust, where users gradually increase automation as their confidence develops.
Consider invoice processing.
Initially, a finance professional might want to review every invoice and confirm each extracted value before approving it.
After seeing the system perform reliably, they may choose to review only invoices containing uncertain information or amounts above a particular threshold.
Eventually, they might authorize the system to process routine invoices automatically, notifying them when an exception requires attention.
“The human controls that based on their level of confidence,” Aaron explains.
This progression has implications for companies developing AI agents. Customers may trust an agent to summarize financial information long before they’re comfortable allowing it to update records or initiate payments.
Rather than making automation an all-or-nothing decision, founders can design adjustable permissions, approval thresholds and exception alerts.
These controls allow customers to delegate more responsibility as they gain experience with the system, without surrendering oversight before they’re ready.
Why Speed Alone Won’t Win
For much of the software industry’s history, bringing new technology to market quickly has been a significant competitive advantage.
Aaron believes AI introduces another consideration.
Powerful models and development tools are increasingly accessible, allowing companies of different sizes to build sophisticated capabilities. Access to the underlying technology is becoming less distinctive.
What remains difficult is understanding how customers need to experience AI before they’ll rely on it.
“You can’t rush trust,” Aaron says.
A company might release an impressive agent ahead of its competitors, only to find that customers hesitate to use it for important decisions.
That gives SaaS businesses with deep knowledge of their customers an opportunity. They understand which tasks carry the greatest risk, where mistakes are most costly and what evidence users need before handing over control.
Aaron encourages founders to examine the moments that strengthen confidence and those that undermine it.
“Trust arrives on foot and departs on a horse.”
The consequences of a mistake depend on what the AI was authorized to do. An incorrect summary might be inconvenient; an unauthorized financial action could be considerably more damaging.
For founders, speed to market still matters, but so does the quality of the experience customers encounter once a feature is released.
The User’s Job Is Changing
As AI assumes more routine work, Aaron expects people to spend less time completing individual tasks and more time reviewing outcomes, managing exceptions and making decisions.
In finance, that could mean less manual expense processing and fewer hours chasing colleagues for missing information.
“The doer becomes the reviewer,” he explains.
AI could also expand who can use sophisticated business software.
Enterprise systems often require training and familiarity with complex interfaces. Natural-language interactions could allow occasional users to retrieve information, submit expenses or complete routine tasks without learning the underlying system.
For SaaS companies, that opens opportunities to serve people who previously relied on specialists to navigate their products.
It also raises a design question. If customers are increasingly reviewing AI-generated work rather than completing every step themselves, the interface needs to support that responsibility.
Users will need clear explanations, accessible records and straightforward ways to correct mistakes or intervene.
Aaron’s argument ultimately comes down to understanding the experience customers need before they’re willing to give AI more responsibility.
As sophisticated AI capabilities become easier to build, the companies that stand out may be those that make customers comfortable enough to use them.