Johnny Than, CEO and founder of Appficiency and founder of AskCipher (Photo courtesy of Monia Khan)
Gareth Doherty asked the room a question most leaders are sitting with but rarely say out loud.
What does it really mean to trust something you can’t fully explain?
“I don’t know exactly how all the neurons in your brain work,” said Doherty, executive leader at Think/able Solutions, at a Toronto Tech Week panel discussion hosted by Appficiency, SnapOn Software, and AskCipher.
“But if I can reliably predict based on some ask of you that you’re going to behave in the exact same way, which is every person in your organization. We don’t get so tripped up about what’s going on in your thought process. What I care about is that you behave in the right ways in the right context.”
Unsurprisingly, Doherty was talking about AI. Specifically, what’s paralyzing enterprise adoption right now.
You don’t need to understand how something works to trust it. (Bold, but let’s hear it out.)
You just need it to behave predictably. (Of course.)
With most enterprise AI, at this stage, it doesn’t. (Ding ding ding!)
Moderated by Johnny Than, CEO and founder of IT consulting firm Appficiency and founder of AskCipher, the discussion brought together Doherty, Alex Miles, partner at 180 Systems, and Shayan Rastgou, co-founder of AskCipher, all practitioners with enough deployment chops to know what is and isn’t working, and why most organizations haven’t moved beyond a few haphazard tools.
Think of this as a field report from the front lines of enterprise AI, built from the yet-to-be-answered questions many decision makers are carrying into their next meeting.
What nobody talks about in the press release
Than opened the session by walking through the AI-assisted hiring process step by step, from job description to final candidate selection, as an exercise in showing where enterprise AI breaks or brings friction.
AI has been involved in hiring for a while now, and each stage surfaces a different kind of failure. Off the bat, there’s technical friction when the model doesn’t know what a job really requires at your company, so it takes a guess.
Human friction comes along when a recruiter can’t trust what the screening agent lets through.
Bringing up the rear, organizational friction comes calling when the scheduling tool books the first available slot for an interview. And nobody wants their first available slot of the day to be an interview.
Than named these failures systematically, before coming in with the big guns.
AI reveals risk at scale.
“Everyone’s using it,” he said, “but nobody’s trusting it. That’s really where we’re going.”
The trust problem shows up in practice faster than most organizations expect.
In an interview after the panel, Than explained how a client tried to build a live executive dashboard by connecting an AI directly to their Enterprise Resource Planning (ERP) system using Model Context Protocol (MCP), an open-source standard that lets AI models pull live data from external applications.
While simple enough on the surface (budget versus actual, year to date), the data they needed lived across multiple systems, not just the one the AI was connected to.
What started as a dashboard project became an integration project, then a data rationalization project. Like many teams that end up too deep, the project was abandoned.
A second client ran into a different wall. Their AI was consolidating purchase orders to build a bulk vendor order, the kind of exercise that can unlock better pricing at scale. The model hallucinated the order total, and while to an outsider it seems like a small error, one or two per cent, the trust was gone and the brakes were pumped.
In the panel discussion, Rastgou offered what it looks like when an organization works through the friction instead of abandoning the project.
AskCipher, an AI-powered interface layer, was rolled out to their major client, Appficiency, where it sees about 200 to 300 active daily users. These employees use the AI across three main areas: ERP implementation, code development, and communication.
Rastgou explained that humans expect AI to act like traditional, deterministic software that flawlessly does the same thing every time. Because AI is probabilistic and inherently hallucinates, users get frustrated when it makes mistakes.
“When you put humans in front of an AI, they go into two categories: they either expect it to do everything very well every time, or they expect it to be able to reason like a human,” Rastgou said. “They become angry, and then they get into their seats, and they yell at the AI, and the AI won’t do anything good for them”
By testing the tool with a diverse group to find all the ways it fails, though, the team was able to adjust expectations and build around the failure points.
What they did see was a 25% improvement in run rates, yielding roughly $5 million a year in efficiency gains against an investment of over $2 million over two years.
“If we take the time to actually provide it with the right context, both the AIs and the humans interacting with it,” Rastgou advised, “then it would be a lot less of a friction implementation.”
Doherty pointed to one organization he’s working with, where a team of PhD-level qualitative analysts resisted using AI to help analyze unstructured customer feedback.
Their concern is, of course, whether AI tools could put them out of a job.
But he argued the larger cost is what the organization loses while that resistance continues. These analysts could be spending more time on deeper research and analysis, with AI speeding up the first pass. He estimated the opportunity cost of not redeploying that expertise at roughly $600,000 to $800,000 a year.
The adoption problem, he said, is that while these tools are in the organizations, and employees are getting creative, we aren’t quite there “in terms of creating the culture and the know-how to be able to engage.”
“If you’re going to start small, you want to start with your more successful teams and try to make them better. It’s the best way to pilot. You’re going to learn a lot more,” Doherty said.
“You’re going to have engaged people, they will know their business processes better than poor performing parts of your organization as well. They’ll understand their workflows much better and to be more engaged at being able to get successful and learn as you grow. Because I think, honestly, adoption really is a learn-as-you-grow.”
The knowledge economy just changed
For most of our professional lives, knowledge was the moat.
For example, knowing how a specific application worked or being able to compare CRM platforms. Generative AI, however, has commoditized a significant portion of it.
“What used to be ‘I know all the stats of every restaurant in the world’ is now instantly available with one prompt,” Than said. “That concept of knowledge being your moat has changed.”
Instead, judgment is taking its place. Than used a hiring scorecard as an example.
Train an AI on your scorecard and it will apply it consistently at scale. That’s a knowledge economy problem, solved. Ask the AI to build you the scorecard and you’re in different territory.
The output might be better than anything you could’ve made yourself, for all you know. Now, however, someone has to read it, interrogate it, reverse-engineer how it made its choices, and decide whether those choices were right.
Ta-da! Judgment economy.
Every enterprise Than works with has built a series of checks and balances around AI output. Spending time evaluating what the model produces is, for now, the normal state.
Than said the people you need to hire are the ones who can evaluate, challenge, and apply what the models produce.
Miles made the same point through a client example. A company heading into a private equity sale needed a full ERP implementation completed in under two months, a timeline that would historically have been impossible.
Using AI for data migration mapping and automated report generation, the team compressed what is typically a nine-month process, freeing the team to better focus on value for their clients.
“If we didn’t hit that date,” Miles said, “it would have been catastrophic for the organization.”
The practical part
The panel closed with a question Than put to the room: if you walked out from the session today and wanted to start training yourself on AI tools, what should you do?
“The short answer is use it,” Miles said. Pick one business process, something that takes two hours, and see what you can automate.
Doherty quoted Walid Hejazi, a program lead at the Rotman School of Management, who said “When we focus on technology rather than business objectives, the strategy becomes about inputs instead of outputs.” Start with the outcome, he advised, and work backwards, making sure there’s a foundation of AI literacy in your organization.
Rastgou focused on communication.
“The biggest element of getting what we want out of the LLMs is how well we communicate with them,” he said.
Sign up for several models, he suggested, learn each one’s character, and build from there.
Than’s answer to his own question was a different kind of practical. The clearest signal that a company is going to get real value from AI, he said, is that they assigned a human to figure it out. Not a software purchase. A person with a mandate and a question.
In our interview after the panel, he described an eight-person accounts payable team struggling to find efficiency. The company assigned one person to find out where the time was going and what might help.
They interviewed their teammates, looked at available tools, picked one, and ran with it. The tool handled optical character recognition on incoming vendor invoices, routed approved vendors directly into the system, and generated the bank payment files. Two people had been doing that work. One of them was freed up for analysis.
The process that remained ran with AI assistance.
“What would you normally hire someone to do to solve the problem?” Than said. “Take that, and see how much of it you can do with AI first before you hire the person. Because maybe that person can come in and do something way more interesting for you.”
In the judgment economy, the hire still happens, but they just arrive at a different job.
At the risk of sounding like a kindergarten teacher, it’s easier to maintain patience when you remember how new all of this is.
“Widely available LLMs were 2024, and we’re halfway through 2026,” said Doherty, “so the idea that there’s a robust database of experiences to be able to share and help you with your adoptions just isn’t there. We’re all kind of learning together.”
It’s easy to feel left behind, but everyone is somewhere on the curve.
“The things that are really challenging on the AI side [are] less about technology, more about people,” he added, “so one of the main barriers is really about digital literacy and AI literacy specifically.”
Your data has more layers than you think
On data governance, Than offered a framework that could roughly be compared to a traffic light.
Operational data (for example, customer records, contacts, raw transactional data) is generally seen as AI-friendly for organizations, with some error tolerance and auditability after the fact.
It’s a green light, if you will.
Financial data like payables or customer invoices are more proceed-with-caution for organizations. They’d largely want a human to review before anything moves.
You just said “yellow light!” you smart cookie, didn’t you?
Regulatory data like GDPR compliance, licensing, or anything where a single violation can trigger a contractual or legal consequence, is still largely “do not pass Go” territory.
Most organizations, Than said, are afraid to touch this layer with AI at all, but the frameworks that would make teams comfortable don’t fully exist yet.
You probably don’t need me to tell you, but, yes, the red light.
Than also raised the topic of data domains that haven’t been named yet, an issue the panel hadn’t covered.
Beyond public, private, and process data, he sees financial data, judgment data, and identification data as emerging concerns.
“I think that there’s still more domains that we haven’t even learned to describe yet,” he explained, “but it makes us distinctly human and able to execute many complex things without as many checks and balances.”
Intuitive data, he argued, is what CIOs should be thinking about protecting now, before AI finds ways to surface it.
“How many invoices did you cut last month?” asked Than. “Probably lost to the ether already, like it or not. But why is this customer valuable to you? What is your profit with this customer? Those are the intuitive pieces of data that we need to protect, and the CIO needs to think about protecting those now, long before AI comes along and sneakily grabs it from you.”
Final shots
- The organizations with efficiency gains from AI put a human in charge of figuring out where the tools fit before they bought anything.
- The lack of trust does come with a dollar figure. Than’s clients stopped using AI because it got something wrong once, and that was enough.
- The data governance conversation is coming whether organizations are ready or not. The best-prepared CIOs will have already mapped what they’re willing to let AI touch, and what they aren’t.