Constellation Software’s AI Bet Still Has No Forecast

Constellation Software shares fell 4.9 percent on September 22, 2025, after founder Mark Leonard hosted a public webcast on AI and refused to forecast its path. The Toronto serial acquirer of niche software firms had already dropped more than 18 percent from a May peak on the same fear. Leonard still would not give the street a map.

Three days later he was gone. A year on, president Mark Miller is running the same bet: ship useful tools, buy more vertical software, and print no AI target.

Leonard Opened With a Radiologist Who Was Not Replaced

Leonard, Constellation’s founder, president and director, began the session with Geoffrey Hinton, the University of Toronto researcher often called a godfather of AI. In 2016, Leonard said, Hinton told people to stop training radiologists because machines would take the work. Nine years later, Leonard said, the count of U.S. board-certified radiologists had risen from 26,000 to 30,500, a 17 percent increase that outpaced population growth, from 7.9 to 8.5 radiologists per capita.

Firstly, you and I will never know a tiny fraction as much about AI as Jeff did. Secondly, despite his deep knowledge of AI, he was unable to predict how it would change the structure of the radiology profession.

Mark Leonard, founder and president, Constellation Software AI webcast, September 22, 2025

He applied the same caution to programmers. A tenfold jump in coding output could spark a renaissance of custom software, or a recession of surplus talent if demand stays still. A 10 percent gain, he said, would leave the status quo almost intact. “We don’t know which way this is going to go.”

That line was the product. Constellation buys, holds and compounds vertical market software, the unglamorous systems that run a city utility, a clinic, a school board. The model never required a 10-year technology forecast. The webcast was that model, spoken into a microphone.

Four Specialists, First Names Only

Beside Leonard sat four internal AI leads he introduced only as Cam, Chris, Henk-Jan and Paul. He kept surnames off the line so they would not be “overwhelmed by outreach from our shareholders or competitors.” They were long-tenured application people working on AI full time, he said, not data scientists writing CUDA on NVIDIA machines.

Thanos Moschopoulos of BMO Capital Markets asked whether head office would impose metrics so laggard business units could not hide. Leonard said some operating groups already track AI in detail and others prefer case studies. There was no Constellation-level scoreboard. He expected a rolled-up subset later. He also warned that radical decentralization produces duplicated experiments, usually modest ones, and sometimes useful lessons.

Cam said the early wins were narrow: unit testing, test-plan writing, stored-procedure tuning, vulnerability scans, code comments. Vertical systems often run to tens of millions of lines, and the models’ context windows were too small to hold that whole stack for debugging. Henk-Jan added that a clean-sheet build can look wildly productive, while maintenance and bug-fixing stall on the same window limit. Vendors were racing toward multi-agent setups that carve the job into pieces each agent can hold.

WHERE THE TOOLS HELPED, AND WHERE THEY STALLED

• Narrow chores: Testing, comments, scans and procedure tuning showed the first real lift.
• Greenfield code: Starting from a blank page was the cleanest use of the models.
• Giant codebases: Context windows could not yet swallow a full vertical system for troubleshooting.
• Lifetime cost: Faster first drafts can raise later maintenance if the generated code is buggy.

One large business unit had used fairly simple AI coding tools for about a year. Suggested lines of code rose. Adopted lines stayed flat. Efficiency, in the sense that pays the bills, had not moved.

Leonard then read figures from one operating group that had sent him data, not from the whole company. They are a snapshot of how far that group had got by the morning of the call, nothing more.

ONE OPERATING GROUP’S AI USE, SEPTEMBER 2025

Leonard said 3 percent was lower than he had expected, and probably a good thing. Other units told him they had more work, not less, once the tools sped people up. A second group had been told to experiment, so its answer was “all.” A third said about 20 percent of units were seeing meaningful gains. Customer-service use was 29 percent in the first group and over 50 percent in a fourth. Output quality, he said, was all over the map, depending on the tech stack and how well public models already knew that niche.

Why Constellation Software Hosted a Public AI Call

The company is usually quiet. It does not do product keynotes. Leonard has long avoided the press. The webcast happened because one of its largest shareholders asked CSI for a private question-and-answer session on AI’s effect on software businesses. Management figured the talk would slide from the industry to Constellation itself, so it opened a live audio webcast at 9:00 a.m. ET on September 22, 2025, for every holder.

Leonard called that “a nice example of an enterprising shareholder catalysing corporate action that benefits all shareholders.” He also said the quiet part: AI had created uncertainty for employees, shareholders and customers, and neither CSI management nor the specialists “purport to know the future of AI.” North American listeners could dial 1-877-879-1183, international listeners 1-412-902-6703, conference ID 2442103. CSI said a replay would stay up for 12 months at its media-server link. That window closed in September 2026.

Analysts on the line included Moschopoulos, Stephanie Price of CIBC Capital Markets, Paul Treiber of RBC Capital Markets, David Kwan of TD Cowen, Samad Samana of Jefferies and Jérome Dubreuil of Desjardins. Questions ran past 90 minutes and covered competitor AI features, whether big customers would build their own tools, and whether AI changed how CSI prices acquisitions. Leonard’s answer on deals was that the company was not short of opportunities and was not content to sit on cash. AI had not rewritten discount rates, he said, though it did raise the chance a given deal would miss the plan. High hurdle rates were the margin of safety.

Shares Fell 4.9 Percent After the Non-Answer

Holders did not come for a radiologist anecdote. They came for a statement that vertical software would be fine, or a concession that it would not. They got a monitoring program, four first names, and a warning to treat bold vendor claims with skepticism. Leonard told listeners to sleuth, run experiments, and think like anthropologists rather than people hunting a single scientific law.

The tape treated that as a non-answer. Constellation’s stock, listed in Toronto as CSU, finished the session down 4.9 percent. The company reports in U.S. dollars; the shares price in Canadian dollars. For a firm that had spent years compounding in public with little theatre, a special-purpose AI call that ended with “we don’t know” was itself a data point. Uncertainty is a discount rate. The market applied one.

David Kwan asked whether large customers were already using generative tools to displace Constellation’s products. Leonard placed vertical software between cheap horizontal apps that do half the job and fully custom systems that do all of it. Clients with regulatory burden and little appetite for building in-house still wanted a vendor who already knew the workflow. The risk, which he did not dismiss, is that the largest, most strategic clients may try to take more of that stack inside if AI makes homegrown software cheaper. His prescribed response was more time with the customer, not a press-release feature list.

He also drew a line between a feature that happens to use a model and “AI washing,” the habit of slapping the label on work the customer did not ask for. If everyone in a vertical uses the same sales tool, he noted, nobody is differentiated. That is a dull operational point. It is also why a holdco of several hundred niche vendors cannot issue one AI revenue figure that means anything.

Three Days Later the Founder Was Gone

On September 25, 2025, Constellation said Leonard had resigned as president for health reasons, effective immediately. He would remain a director. The board named Mark Miller, then chief operating officer, as president. Miller’s other jobs inside the group stayed as they were.

The company did not tie the move to the webcast. The calendar still reads as a shock: notice on a Tuesday, the AI call on a Monday, the founder out on Thursday. Shares briefly plunged 19 percent on the resignation news and wiped out C$4.8 billion ($3.4 billion) of market value in the first hours. Chairman John Billowits wished Leonard a full and swift recovery and credited his “visionary leadership, humility and wisdom” since he started the company in 1995. Leonard said he and the board had “complete confidence” in Miller, a trusted adviser inside the executive team for more than 30 years.

Miller’s own history is the Constellation story in miniature. He co-founded a business in the late 1980s that became CSI’s first acquisition in 1995, and he has been inside the group ever since, including long stretches as executive chair of the Volaris and Trapeze operating groups. On October 1, 2025, he and chief financial officer Jamal Baksh hosted a separate call so investors could hear how capital deployment might change. The short version: the machine was built to outlast one person.

THE NINE DAYS FROM NOTICE TO RESIGNATION

1. September 16, 2025: CSI announces a shareholder Q&A webcast on AI for September 22 at 9:00 a.m. ET.
2. September 22, 2025: Leonard and four first-name specialists take questions for more than 90 minutes. Shares close down 4.9 percent.
3. September 25, 2025: Leonard resigns as president for health reasons. Miller becomes president. Leonard stays on the board.
4. September 26, 2025: CSI sets an October 1 call with Miller and Baksh on the leadership handoff.
5. August 11, 2026: Miller’s CSI reports Q2 2026, still without an AI revenue line.

The health disclosure should be taken as the company wrote it. What the tape priced was the loss of the person who had allocated capital since 1995, arriving in the same week he told holders he could not see AI’s future either.

Miller Still Will Not Print an AI Number

By August 2026 the argument had not moved, even as the operating picture had. Miller told listeners on the second-quarter call that AI is changing two things: customers want the software to do more, and the businesses can build faster. He then shut the door the street still wants opened.

We are not going to give you an AI target, an AI revenue line or an AI time line. If we start reporting a number like that, we will start managing to it.

Mark Miller, president, Constellation Software Q2 2026 earnings call

He added that the software is mostly mission-critical, embedded in how a customer runs the day, and that AI is something CSI adds to make that software more valuable. Decision rights stay in the business units, funded on their own profit-and-loss statements. Head office will not run an AI program from Toronto.

The second-quarter results to June 30, 2026 are the first full-year scorecard on Leonard’s wager. Revenue rose 17 percent to $3,335 million from $2,844 million. Almost all of that was acquired. Organic growth was 3 percent, or 1 percent after foreign-exchange. Net income attributable to common shareholders jumped 386 percent to $274 million, or $12.93 a diluted share, from $56 million, or $2.66; strip out foreign exchange and an IRGA liability revaluation and the increase was 8 percent. Free cash flow available to shareholders, CSI’s preferred cash metric, rose 57 percent to $345 million from $220 million. Cash from operations was $477 million, up 10 percent.

Q2 2026 VERSUS Q2 2025

Miller told analysts the company has not yet seen a real pickup in organic growth from AI, and that pickup is still a ways out. Product teams can now chew through backlogs faster. Selling the extra software still requires a customer budget and a problem the buyer will pay to kill. Maintenance and other recurring revenue, the line Leonard used to call the best health check, is where that lag shows up. For the first half of 2026, total revenue was $6,516 million, up 19 percent from $5,498 million, with organic growth of 4 percent, or 1 percent after FX. The board declared a $1.00 dividend, payable October 9, 2026, to holders of record on September 18, 2026.

On the May 2026 annual meeting stage, operating-group heads sounded more impatient than the Q2 print. One described a path from concept to production that used to be measured in months, then weeks, then days, and now hours. Another said a chat agent had resolved 82 percent of the tier-one tickets that went through it. Those are local facts, not a group forecast. They sit beside Miller’s refusal to roll them into a headline number, which is the same refusal Leonard made in September 2025 with different words.

The Buyers Wrote Bigger Cheques on Smaller Multiples

If the wager has a second side, it is in the deal pipeline. CSI completed acquisitions in the second quarter for $732 million in cash, plus deferred payments it valued at $160 million, for $893 million of total consideration. After June 30 it had closed or committed to another $818 million ($669 million in cash plus $149 million deferred). Fear that AI will cheapen software writing is ugly if you already own the software. It is useful if you are the permanent buyer of niche vendors whose owners now hear the same scare story from their bankers.

That is the part of Leonard’s bet the tape still underweights. Vertical software was never a moat made of code volume. It is a decades-long conversation with a regulated customer, plus the data and workflow that conversation produced. A model trained on public repositories does not have the court filings, the water-meter history, or the trust to run them. Switching a municipality for a demo is a different act from generating a function. CSI’s own experience with a year of coding assistants, more suggested lines and no more adopted ones, is a small version of that gap.

The cost of being wrong is also clearer than it was on the call. If AI does let customers or new vendors reproduce a niche system at a fraction of the old effort, CSI’s acquired cash flows shorten. Leonard’s high discount rates were built for misses of that size. Miller has kept the same hurdle, the same decentralization, and the same silence on an AI line. He has also kept buying.

A year after the webcast, Constellation still will not tell you what AI is worth inside the group. It will tell you what the group collected, what it paid for new businesses, and what it will not manage toward. The radiologist count went up. The share price, on the day Leonard said it might, went down. The bet is still open, and Miller is still not pricing it for you.

Frequently Asked Questions

Who Requested Constellation Software’s AI Investor Call?

CSI never named the holder. Leonard said one of the company’s largest shareholders asked for a question-and-answer session on AI and software businesses, and that management opened the session to every shareholder once it became clear the talk would turn CSI-specific.

Is the September 2025 AI Webcast Still Available to Replay?

CSI said the audio replay would remain on its media-server link for 12 months after September 22, 2025, which means that window ended in September 2026. Written transcripts of the session still circulate on financial-data sites, and the original press notice remains on CSI’s site and on GlobeNewswire.

Did Mark Leonard Resign Because of the AI Discussion?

The company said he resigned for health reasons, effective immediately, and that he would remain a director. The announcement did not link the decision to the webcast, to the 4.9 percent session drop, or to AI strategy. Miller, whose firm was CSI’s first acquisition in 1995, became president the same day.

Does Constellation Software Report an AI Revenue Figure?

No. Miller said in 2026 that CSI will not publish an AI target, an AI revenue line, or an AI timeline, because the company would start managing to that figure. Business units fund their own experiments on their own profit-and-loss statements, and group reporting stays on revenue, organic growth, cash flow and acquisitions.

Disclaimer: This article is news reporting and analysis of Constellation Software’s public statements, its September 22, 2025 AI webcast, and subsequent results. It is informational only and is not investment advice, a recommendation to buy or sell CSU or any other security, or a prediction of future returns. Readers should consult a qualified financial adviser or licensed investment professional who can consider their own objectives, time horizon and risk tolerance before acting on any figure or event described here. Share prices, organic-growth rates, acquisition totals and management comments reflect the cited company notices and call remarks as published and can change with later filings, restatements or further disclosures.