Earlier this year the Goodhart Future Leaders team discussed how it believed the market was overestimating the threat AI posed to established software businesses. Since then, the share prices of several of the companies discussed have risen sharply, but the more interesting story is what has happened underneath. Russell Champion revisits the original thesis: what we got right, what we got wrong, and what six months of company-level evidence tells us about where AI is really creating and destroying value.

Executive Summary

Six months ago we argued the software and classifieds sell-off was mispricing durability, that AI was reshaping value within the universe rather than uniformly destroying it, and that businesses with proprietary data, high switching costs and system-of-record status would prove more resilient than the market was pricing. The fear embedded in valuations at the time was that AI-native entrants would disrupt and take share or that enterprise customers would bypass the software layer entirely by building their own tooling, “vibe-coding” their way out of SaaS subscriptions. Neither has happened at scale yet, and looks unlikely to in the next 12 months. At the height of the sell-off, some specialist technology investors were describing software as “toxic”, a useful indication of just how extreme sentiment had become.  The old adage that people overestimate change in the short term and underestimate it in the long term is playing out.

You can see from the chart above how extreme sentiment had moved within the technology sector.  Software had been sold heavily to fund increased semiconductor exposure.  At that point our fund's exposure to software and classifieds stood at ~42% today it stands at ~52% having added two new software positions. I feel proud that we stood our ground when others were panicking.

What We Got Right, What We Got Wrong

The core framework held. None of the system-of-record businesses in the portfolio were displaced by an AI-native entrant; each instead absorbed AI into the existing platform, exactly as we argued in February. Enterprise buyers so far prefer sourcing AI from incumbents, although we admit it’s still early days.

The specific fear we were pushing back against, that revenues would soon fall as users churned to AI-native alternatives or vibe-coded their own solutions, has not materialised in the system-of-record names. What has been more nuanced is how the investment cycle played out. Software companies with genuine innovation capability spent the first half of 2026 investing heavily in their own AI products, and the market initially penalised them for it. Higher R&D spend, compressed margins, analyst downgrades. The market read investment spend as evidence of threat. By the third quarter, those investments are beginning to generate returns, new product launches, new revenue streams, and the first disclosed AI commercial traction numbers, and the re-ratings in Zeta and Money Forward reflect the realisation that profits might reaccelerate soon.

One topic we maybe didn’t fully address in our original note: how differently the AI story resolves company-by-company. For some incumbents, AI is close to a pure product tailwind, with no real threat to the underlying revenue model. For others, like GitLab and NICE (a new position for our fund), it’s a genuine race to grow consumption revenue faster than seat-based revenue erodes, and both are still working through it in real time. GitLab’s dollar-based net retention has ground steadily lower over recent quarters, to 117% (still a good number!) in its most recent results, as customer-side layoffs and M&A shrink seats within existing accounts; the base is still expanding in dollars and logos, but seats per account are softening. And NICE’s Non-GAAP operating margin had already fallen ~20% from its peak in 2024, driven by a sharp rise in operating costs as they aim to prepare the business for the AI opportunity. Across the two quarters we’ve held it, cloud growth has been stable but not accelerating: NICE however has been discounting to extend contract length, locking in customers in a competitive environment. Will these issues derail a return to growth as new AI products become a larger part of the revenue? That’s why we own a portfolio of names; remember we do not expect to get every investment decision right.

Returns Since Last Note

Software Index vs DM benchmark vs Our Names Average

The chart above shows the performance range of our portfolio names against the Software Index and the developed markets benchmark since 13 February. Money Forward, Twilio and Zeta Global have roughly doubled or better since the note, and Pinewood returned in a similar region before being exited at deal value in August. Veeva, Kinaxis and GitLab are all up over 40%, with Dynatrace and CarGurus close behind in the 30s. The middle of the pack, nCino, Visional, Oro Co, Scout24 and Baltic Classifieds, sits in the low teens to high twenties, consistent with a re-rating thesis that remains a work in progress. Phreesia, Doximity, RaySearch and NICE are all close to flat.

Two names have been weak, and both for reasons that have nothing to do with the AI thesis this note is built around. Craneware fell ~30% in a single day on 3 July after a 340B drug-pricing revenue miss, before another setback on a disclosed data breach; neither event was AI-related, and the position is down about 11% since the note despite announcing new AI products and services. It is the year’s clearest reminder that a company-specific, non-AI risk can do more damage to a position than anything on the AI roadmap. Angi remains the standout laggard in the classifieds bucket, a company going through a business model pivot driven by AI.

Overall, given we had large positions in those names that have done well these names have driven considerable performance for the fund which we will detail in our next quarterly.

What We Did In The Portfolio

We kept buying into the panic rather than retreating from it. The Pinewood position was rebuilt during the post-takeover-withdrawal trough and exited in full in August at roughly deal value, approximately double our average entry price. GitLab was added around its April low and continued to be built into strength through June; it's now up materially from the low. Zeta was initiated around the time of the original note and added to repeatedly through the spring and summer, with the Palantir-powered business intelligence product providing fresh conviction along the way. Two new positions, nCino and NICE, were established without waiting for confirmation. Scout24 was pre-existing in the February portfolio and added to steadily over the following months.

The miss was Money Forward: trimmed in two tranches in April right after its first AI-driven earnings pop, only to watch the stock run another 50%+ from there as the re-rating fully played out. We sold Lumine in full after a large acquisition changed the thesis; it has outperformed the market but underperformed the average of the software book since.

Where We’re Not Sure Yet

GitLab

A share price reflects expectations, not quality, a company priced for disaster can rally just by turning out to be merely bad, no improvement required. We think GitLab is better than that. Its core DevOps business is seeing real seat erosion, AI story or not, but it is a good business priced as a threatened one, with real upside if usage-based pricing and the new agent platform work, we would win twice, once as low expectations recover, again if the business genuinely improves. The risk runs the other way too. If the seat losses prove permanent and the pricing shift stalls, a good business priced as threatened simply turns out to be a threatened business, fairly priced. Our case rests on the first outcome, not the second; we are watching usage revenue growth and retention closely to find out which one we get.

NICE

NICE is a different kind of question to GitLab. We did not buy on a view that the market had overreacted; we bought after the margin reset had already happened, cognizant on where the business stood. The case for owning it is labour replacement. If NICE’s AI agents can genuinely substitute its new products for human contact-centre staff at scale, value per seat rises sharply, not because NICE raises its price, but because a customer converts a slice of its labour budget into a NICE subscription instead, a bigger prize than simply flexing between AI and human seats in one contract. Two quarters in, cloud growth is stable but not accelerating, and NICE has been discounting to lock in longer contracts rather than pushing price, which tells us little either way. We are watching for the AI product to start converting labour budget, not just defending the seat base it already has.

Angi

Angi is the highest risk AI impacted stock in the portfolio, a genuine strategic pivot is required to prepare the business for the future. Management has effectively torn up the old local-services marketplace model and is repositioning around trust, becoming the verified point of contact professionals need as AI increasingly handles the sales and marketing work around finding and winning customers. If it works, Angi becomes infrastructure for a trades industry that no longer trusts what an AI-generated lead or review actually represents. It is early, the execution risk is significant, and the numbers do not yet show whether professionals or customers are buying into the new positioning. We hold a small position sized for exactly that uncertainty, enough to have a stake in the pivot working, small enough to protect us from failure.  If we see traction we will increase our position size.

What We Have Learnt Since

Across company reviews this summer, consistent tailwinds and risks emerged.

Positives

• Vertical leading SaaS companies are responding fast. They are winning the race to own the control layer, not just defending share. nCino’s Agentic Operating System, GitLab’s Duo Agent Platform, NICE’s Mpower orchestration layer, Scout24’s Agentic OS and Craneware’s “guided, not autonomous” AI are the same strategy, sell the governance and logic that agents run on rather than compete with the underlying model.

• Enterprise buyers still prefer sourcing AI from trusted incumbents. NICE customers want to flex between AI and seats inside one contract; Pinewood sells its AI modules into rival dealer-management systems precisely because trust travels with the incumbent relationship.

• Proprietary and regulated data is being reinforced by AI, not eroded. Scout24 deliberately withholds its property data from external LLMs; nCino’s customers’ loan-level data underwrites AI outputs a bank credit committee will actually trust; Money Forward’s Japanese regulatory complexity is a genuine barrier foreign AI entrants have not cracked.

• AI is expanding addressable markets rather than just defending existing revenue. Money Forward’s AI Cowork targets customers’ back-office headcount budgets, a different and larger pool than software budgets. Zeta’s new Business Intelligence product opened a substantial new market and a new buyer: the CIO rather than the CMO.

• Monetisation is now real and disclosed, not just promised. GitLab’s Duo Agent Platform added more net-new ARR in its first quarter after launch than Duo Pro and Duo Enterprise combined had generated in any prior quarter, though management cautions it’s still too early to model; NICE’s AI ARR is growing strongly year on year; Scout24’s AI-feature customers show materially higher ARPU.

Negatives

• GitLab’s own numbers show real seat erosion, not just an AI narrative. Net revenue retention has declined for eight straight quarters, slipping to 117% in the latest print, as customer-side layoffs and M&A shrink seat counts within existing accounts. The dollar and logo base is still growing, but seats per account are softening, a genuine structural pressure sitting underneath whatever view we take on the opportunity.

• Incremental IT budget has been diverted toward AI, and much of that spend was pushed straight into raw token consumption on platforms like ChatGPT and Claude, with some teams effectively maxing out usage rather than deploying it sensibly. That crowded out spend on everything else, including the software our holdings sell. The experimental phase is now maturing and budgets are being refocused on where AI actually adds value and drives ROI. As that discipline takes hold, the pressure on other software budgets should ease, with spend flowing back toward platforms that convert AI into measurable outcomes.

Conclusion

The original argument, that AI is reallocating value within software rather than destroying it uniformly, has held up. The more interesting lesson from six months of company-level work is that the investment cycle itself was the evidence the market was slowest to read, initial AI R&D spend attracted downgrades; the same spend is now generating disclosed commercial traction and driving re-ratings. We are still early in that sequence.

Durability is still being mispriced, just not always in the direction the AI narrative implies. Sometimes the market underprices a name because it is scared of an AI threat that barely applies, and sometimes it correctly prices a real one while missing an unrelated, more mundane risk entirely.

GitLab is a good illustration of the opportunity. Its entry-level seat price of $30-50 a month (Premium, with Duo Pro layered on) looks cheap next to what heavy users already pay for AI subscriptions elsewhere: Claude Max and ChatGPT Pro both top out at $200 a month. Even GitLab’s own Ultimate tier, which accounts for the majority of its ARR, runs closer to $100 a month, so the real headroom is smaller than the entry price alone suggests, but it is still there. That gap is the prize. A company priced as a seat-erosion casualty only needs its usage-based revenue to capture a fraction of that per-seat ceiling to stop looking like a loser and start looking like a business that found a new, larger revenue line inside an old one. It won’t work for every name carrying an AI-threat discount, but it’s the kind of mispricing we’re looking for next.

On our five-year assumptions, exit multiples for quality software incumbents remain undemanding, and we believe the IRR arithmetic is attractive at the prices where we have been buying. The cases where we pick the AI winners correctly may still have a great deal of value to unlock.  The edge remains the same, patience and company-specific work, and the discipline not to let the “AI risk” label substitute for either.