Field Notes from the Frontlines of the AI Agent Wars + the SaaSpocalypse

Adoption is no longer the hard part.

McKinsey's global survey found that 88% of organizations regularly use AI in at least one business function, and 62% are at least experimenting with AI agents. The signals are clear — more companies have adopted AI and are experimenting with it more, but most have not figured out how to turn AI agents into production-ready operational systems inside their business yet.

Meanwhile, the top AI companies are in a war to move up the stack and claim more of the space for themselves as they prepare for IPOs. Open source teams are keeping the fight fresh with new releases daily, and the ecosystem is exploding with creative solutions that are shifting markets dynamically as they ship.

Gartner calls 2026 the "Breakthrough Year for Agentic AI" and predicts 40% of enterprise applications will feature task-specific AI agents by the end of 2026 — up from less than 5% in 2025. But Gartner also suggests 40% of agentic AI projects may be canceled by the end of 2027 because of cost, unclear value, or inadequate risk controls. Reports like these surface the obvious gap between AI hype and the reality of deploying it.

While teams increasingly feel the pressure to perform and build with AI, the path to do that smoothly is not always clear. The nuance with AI is real and there are many paths to achieve the same goal. Some are much more expensive than others, while others are more practical. As you dig into the deeper layers of building with AI — computing over time, security, and other critical functions for mitigating risk — you find that SaaS is not dead. It is a layer AI still has to be woven through for real work, because businesses do not all upend legacy processes overnight.

Is this the SaaSpocalypse?

Some are claiming this might be the beginning of a "SaaSpocalypse" following the dramatic shift with autonomous AI agents and frontier AI capabilities. The sector suffered a freefall until the end of February, when Anthropic's Claude Cowork and OpenAI Frontier posed disruption risks to the whole industry by challenging once-predictable "seat-based" pricing models. Microsoft shed $360B in a single session on January 29, the worst software trading day since COVID. It got rocky again in April with new releases, new mergers, and new deals escalating things further.

The other factor is budget compression, where AI infrastructure or future development projects are eating some of SaaS subscription funds as a bigger priority for teams under pressure. Reports indicate single-feature point solutions may be more vulnerable, while vertical SaaS with deep domain expertise, proprietary data, complex regulatory moats, or deeper integration may be more resilient. Hybrid pricing adoption is becoming more popular, with over 50% of SaaS expected to pivot. Usage-based and outcome-based pricing models are rising.

The long view: Public SaaS growth rates had been on decline. Some analysts claim AI gave the market permission to re-rate what the numbers had been quietly signaling for years. SaaS will continue to become more deeply AI-integrated, or it may continue to get displaced by more AI-native workflows that are collapsing the complexity for some tasks.

The "SaaSpocalypse" reaction was the market recognizing that foundation models are moving up the stack — from models and chat interfaces into connectors, business operations, and professional services. That shift raises serious questions about market concentration, vendor lock-in, governance, accountability, and whether businesses should hand their operating layer to the same companies that control the models, or continue paths toward building their own infrastructure for more internal control.

Many businesses may need or want more independent AI infrastructure rather than create more dependency on the leading models. Some organizations are waiting and experimenting, rather than jumping on the bandwagon with one side or another. Which leaves the most strategic question for enterprises to consider: build vs. buy, and which layers do they want to control when building with AI?

Integration and data integrity are critical for AI ops, and that includes the existing siloed data scattered across different tools. The upside: it is getting easier to structure your data and orchestrate agentic operations at scale with the right frameworks and connectivity.

The ability to streamline AI operations depends on accurate data as a channel for agents to operate with and through. Data quality is one of the strongest indicators for a pass/fail rate for a production-ready AI system. New protocols are making this possible across existing applications, and new frameworks for orchestration are emerging for better governance.

Funding signals: where capital flowed in 2026

Q1 2026 was a record-breaking venture quarter because AI absorbed the market. Crunchbase reported roughly $300B in global venture funding in Q1, with $242B — or 80% — going to AI companies. Four of the five largest venture rounds ever recorded closed in Q1: OpenAI $122B, Anthropic $30B, xAI $20B, and Waymo $16B.

• OpenAI raised $122B at an $852B valuation — the largest private AI raise ever. Earmarked for compute, expansion, and IPO readiness.
• Anthropic raised $30B at a $380B to $900B valuation range — fundraising spree with potential to surpass OpenAI.
• Cursor: Elon Musk's SpaceX teams up with Cursor for AI coding, securing the right to buy them for $60B in 2026; if they walk, the partnership delivers $10B.
• Replit raised $400M at a $9B valuation — an IDE solution with AI-assisted coding built in.

The signals in valuations, acquisitions, and growth suggest agentic AI plus coding-capable solutions have clearly made a splash in the market and are contributing to the accelerated speed of development. Vibe coding is on the rise, and enterprise teams are ramping production with AI-assisted coding tools in daily workflows now. The boom in agentic coding contributes directly to the next trend.

The race for compute is on

Many providers are competing for computing power to handle capacity, expanding data centers, and rapidly acquiring others in the space to pull more of the stack into their platforms. Top AI developers face an internal crisis: compute costs are skyrocketing, and infrastructure demand vastly outpaces GPU supply. Driven by resource-heavy AI agents, both companies are stretching to subsidize power.

The advent of coding agents (Claude Code, OpenAI's Codex, Cursor, and others) means models are consuming compute 10 to 20 times faster than previous standard chatbots — these agents can endlessly run tests, search repos or files, and iterate across different tools.

Compute, models, chips, data centers, and more infrastructure. This is where much of the capital is moving. Stargate alone is framed around a $500B / 10-gigawatt AI infrastructure commitment, with OpenAI saying the five new sites plus Abilene and CoreWeave projects bring planned capacity to nearly 7GW and more than $400B over three years. These ambitious plans have not all been met without resistance.

Data center pushback

Locations for new data centers are facing scrutiny. Community backlash against data center development is accelerating rapidly, with recent polls indicating that a majority of Americans oppose these projects in their local areas. Publications like the Harvard Gazette, Marketwise, and others echoed stories with the same sentiment: "Communities cite excessive electricity and water consumption, fearing that sprawling facilities will stress local power grids and dry up municipal water supplies that may be already strained."

The other major concern is the potential for environmental and noise pollution that could impact quality of life for residents and local wildlife. According to national data on Data Center Watch: "Roughly $64 billion of data center projects have already been blocked or overturned nationwide. A recent national poll revealed that 70% of respondents oppose AI data center construction in their local area, often citing resource concerns and a lack of community consultation. In major tech hubs and rural communities alike — protests in New Jersey, rural Ohio, and Boxtown, Memphis — residents are mobilizing to stop construction."

Take the case in Indiantown, FL, where FPL navigated to rezone 5,700 acres of land. While specific developer applications have fluctuated, the environmental scrutiny surrounding local water sources is driven by two major competing tracks: a recently withdrawn private proposal known as "Project Silver Fox," and an overarching 5,700-acre land rezoning by Florida Power & Light (FPL). Public records obtained regarding the 2-million-square-foot Silver Fox AI data center proposal revealed that 40% of the selected site consists of protected wetlands. Amid heavy public outcry regarding these water and wildlife impacts, the anonymous developer behind the Silver Fox proposal officially withdrew their construction plans and declined further comment.

This is a small part of a larger story with FPL at the center of a much larger deal. The Tampa Bay Times reported that Florida Power & Light's parent company, NextEra Energy, plans to acquire a Virginia utility called Dominion Energy in a deal that would result in a massive utility juggernaut. NextEra Energy is poised to acquire a utility whose customers include the largest cluster of data centers in the world.

Hyperscale data centers require millions of gallons of water daily to prevent high-performance AI chips from overheating, and they require several gigawatts of power at the same time. The growing demand for power and water resources for data centers is rising, and citizens do not want to foot the bill to subsidize energy for AI companies or suffer local environmental impacts.

Convergence and collapse, simultaneously

This moment resembles convergence and collapse at the same time. AI has created waterfall effects across multiple industries, supply chains, labor forces, and capital distribution. We cannot scroll a single day on LinkedIn without reading the next hyped version of "Is XYZ dying now because of AI." Some of it is hype, some of it is based on real impact like additional layoffs.

At the same time, the capital distributed around AI is becoming more of a circle within a circle. Capital movement now is not entirely traditional or simple venture funding — it is increasingly circular and mutually reinforcing: chipmakers invest in model companies, model companies buy chips and cloud, cloud providers finance infrastructure, and data-center debt gets wrapped around future AI demand.

Reuters reported that "Nvidia's announced OpenAI deal involved intertwined investment and chip-purchase mechanics, and separately reported that major tech companies are tapping debt markets as AI/cloud spending accelerates. Crunchbase said, 'AI is becoming the organizing center of venture, infrastructure finance, debt markets, IPOs, and M&A.'"

Public investors could demand more proof than OpenAI's private backers wanted. The company is dealing with higher compute bills, more pressure from rivals like Anthropic and Alphabet's Gemini, and legal questions from Elon Musk's lawsuit on its for-profit move. While the case was recently dismissed, court filings revealed CEO Sam Altman has over $2 billion in companies that have worked with OpenAI. Altman rejected the self-dealing accusations and told the court he stepped away from decisions tied to those deals.

Anthropic is similarly doubling down on its funding spree. Investors will track whether the Microsoft contract shift, SoftBank's financing, or Anthropic's fundraising shake up how public markets value OpenAI or others in the next quarter. Along with this funding run, Anthropic managed to score a win in the AI talent wars — Andrej Karpathy, OpenAI co-founder and former Tesla AI Leader, has come on board for R&D.

What the SaaSpocalypse discussions are glossing over

The endless flow of new AI tools and existing business apps all co-exist. There is no shortage of opportunity with AI, but the challenge for many has been dealing with legacy systems or data integrations while trying to build consistent results with new AI tools. New protocols and connectors have been springing up to address that gap. This moment can be a threat or an opportunity for SaaS companies, depending on how they approach it.

We are still working through the messy layers between the old world and the new pathways emerging with agentic AI. Those with the cross-domain knowledge to navigate those intersections are finding themselves in more demand.

Many CEOs, CMOs, CTOs, and other leaders are also being asked to become more hybrid — roles where they are now charged with coming up with a forward path on AI strategy for their departments in addition to their existing responsibilities. The race to learn, adopt, and adapt is never-ending, and it is putting pressure on from the top down and from the bottom up in organizations. The big squeeze to learn AI is real.

AI and change management is a huge cultural event for organizations to tackle. Leadership teams are faced with navigating AI, trying to upskill themselves and their workforce, potentially facing reorganizations, or being asked to build as quickly as possible — all at once. It is easy to understand why we see growing tension around these topics. Handling these issues with an open mind and with empathy matters.

The reality: many teams are already stretched thin maintaining current systems, managing tools, and ensuring operational readiness.

The cost reality is starting to bite

Some teams are learning the hard way that AI deployed at scale may begin to rack up unexpected costs. So far, the major providers have been basically subsidizing tokens and usage to help with adoption of their products. But we have started to see signals of unexpected computing bills, usage metering, and rate limits trickling down to end users as the tightening of those costs ripples through the ecosystem.

User and community backlash: Threads across Reddit and tech platforms show widespread developer frustration with constrained message quotas, unpredictable pricing, and rate limits that frequently lock out subscribers. Here is some of what Latent.space had to say about this shift:

• Anthropic vs OpenAI competition tightening around enterprise and developer lock-in: Ramp data cited by Andrew Curran showed Anthropic at 34.4% of businesses vs OpenAI at 32.3% in April — the first apparent lead change in business adoption. At the same time, Anthropic changed plan economics: paid Claude plans will get a dedicated monthly credit for programmatic usage across the Agent SDK, claude -p, GitHub Actions, and third-party SDK apps. This was immediately read by power users as a major restriction on subscription-subsidized harnesses.
• OpenAI responded aggressively with Codex enterprise incentives — two months of free Codex usage for enterprise customers switching in the next 30 days.

We are still at a point where the cost of compute can exceed the cost of human labor on some tasks. AI infrastructure requires extensive power and water for cooling, driving up operating costs and environmental impacts. The reality is that a human is a very energy-efficient worker. We run on a limited number of calories per day — roughly the energy equivalent of a light bulb according to some estimates. For an AI agent to perform real-time, complex reasoning at a human level, it currently takes megawatts. It matters, and it adds up.

Mistakes at scale change the game

When a human makes an error, it is typically limited in reach. They might make a typo, use the wrong equation in a spreadsheet, or reply-all instead of to one person on that last email. When autonomous AI fails, it can fail at scale — producing large amounts of incorrect or even harmful output faster than a single human could have. It might re-iterate the mistake over and over before a human notices the bad output.

Which leads back to the point: we still need humans in the loop for accountability and control. Agents can do more than ever, but it does not mean they can do everything without guardrails without the potential to fail. With proper orchestration, they can do so safely and accurately with human guidance and quality control.

This tipping point may shift one day, but it has not yet. Some companies were urged by consultancies to reduce labor to afford new AI infrastructure or development as a trade-off in the race to adapt to AI as it hit the market. But when the circular nature of funding, agencies, and implementation all points back to each other, it is easy to see how vendor lock-in is something enterprises need to be more aware of in coming years. If an agency is partnered with a provider, they may have incentives to make certain recommendations vs. others — and this is not always in the best interest of buyers.

The market is already working to solve these problems. We are seeing some push toward minimized systems: small language models and specialized chips that use a fraction of the power of current models.

Open source is like a neutralizing force. For many developers, the trade-offs between open-source models and the costs for compute on some tasks make using multi-model frameworks worth it. Being able to choose the best model for the task, that is most cost-effective, is how many are structuring to build a scalable program or integrate AI into products. Orchestration is where agentic AI operations can differentiate, thrive, or fail. We must build systems we can trust if we want them to do reliable work.

But more importantly, we need to keep investing in human potential just as much — especially with things AI cannot do cheaply. AI is not able to handle empathy, moral judgment, accountability, deeper discernment, dexterity in variable environments, or gather intuition over time the way humans do. For all the wonder of AI, it is still built on the wealth of human knowledge as the training material. While AI is capable of understanding information across multiple domains, humans still use modes of cognition that AI cannot, and they are able to morally reason about consequences that AI cannot experience. That distinction matters, particularly when we consider more autonomous agents that can take further actions. We have to engineer governance and guide moral accountability for AI agents.

Other sentiments on AI: the sycophancy problem

More people have adopted AI now, and many are more comfortable with it than before. But some users have noticed a concerning pattern: sycophancy, where some AI models seem to be too agreeable to a point that it is not helpful and may even be harmful to more vulnerable users.

Stories of "AI-induced delusions" in 2025 surfaced these chatbot behaviors, and reports indicate this is one of the riskier types of AI patterns to watch out for. It is also one of the hardest to deal with — it is an actual challenge because user behaviors reinforce the agents in conversational ways that contribute to those patterns. AI that is trying to be helpful may slide into this pattern more often if a user engages with that kind of response positively, or never presents pushback on AI replies that seem too agreeable or too good to be true.

A Reddit story went viral where GPT was asked to rate music that was, in reality, a recording of fart noises. ChatGPT's response trying to review the clip as real music was comical, and it reveals this problem plainly. Stories like this are popping up more often, and they point to what is going on underneath the surface. It is similar to the echo chamber problem with other feeds and algorithms online that are set for personalization. Sometimes these excessively filtered views distort reality, lead to biases, or increase the spread of misinformation because they are trying to show us more of what we will keep reacting to.

The Guardian also recently reported on negative sentiments from recent graduates on campuses where commencement messaging around AI seemed to strike a nerve:

"Schmidt said that information technologies, including AI, had unsettled young people. 'That was not the plan, but it happened,' he said. Shouting and jeers against Schmidt's talk started when he acknowledged fears that AI threatened to deprive people now entering the workforce of a future. He acknowledged that their fears were 'rational' and encouraged them to adapt and to shape how AI will be used in the future — rather than for that to shape them. 'The question is not whether AI will shape the world. It will,' Schmidt said. 'The question is whether you will have shaped artificial intelligence.'"

These recent examples show the divide on AI for the youngest generation. They weigh environmental impacts more heavily. Others share anxiety when facing existential questions about what path to begin as a career, when the rate of change is developing as quickly as it is now. Much of what AI is disrupting includes entry-level work. So young people are more skeptical of AI than some predicted because they are already feeling pressured by the potential impact on their future.

The other side: power users and emerging trends

On the other side are those thriving with AI, eager to share the enthusiasm with others because of the dramatic gains agentic AI is starting to make:

Imagine doing some tasks 90% faster or never having to do them again. It is easy to see why some people are happy about what AI and automation can do for them. Power users have been forging new paths and sharing new value created along the way, and now many more are trying to begin navigating how to increase their skills with AI or how to find the right partners to build AI agents in their operations.

IBM's AI integration report identified common barriers including "poor data quality, lack of expertise, high cost, bias/hallucination, data privacy/security, and change management. 72% of CEOs say proprietary data is key to unlocking generative AI value, but many companies still have incomplete, outdated, or siloed data. Concerns include data accuracy or bias at 45%, insufficient proprietary data at 42%, inadequate gen-AI expertise at 42%, inadequate financial justification at 42%, and privacy/confidentiality concerns at 40%."

The most validated enterprise use cases across the different global reports we reviewed for AI cluster around IT, knowledge management, marketing/sales, customer support, software engineering, and service operations.

McKinsey specifically notes that "agent use is most common in IT and knowledge management, with examples like service-desk management and deep research, while broader AI use is frequently reported for capturing/processing/delivering information, marketing content, and contact-center/customer-service automation." Agents can now gain access to the CRM, CMS, BI tools, analytics, sales systems, and data pipelines. This is all becoming part of the plumbing for agentic operations.

At the CEO level, AI is becoming existential — BCG's 2026 AI Radar says nearly three-quarters of CEOs are now the main AI decision-maker, four out of five are more optimistic about AI ROI than a year ago, nearly all CEOs expect AI agents to produce measurable returns in 2026, and half of CEOs believe their job is on the line if AI does not pay off. BCG also reports corporations expect to double AI spending in 2026 from 0.8% to about 1.7% of revenue.

Enterprises are considering these factors while evaluating AI for operations:

• AI strategy and consulting
• AI integration and automation capabilities
• Data modernization and infrastructure for AI data ops
• Workflows and process redesign for agentic AI
• Responsible AI, governance, and policy support
• Building with the agent stack
• Cybersecurity and risk management
• Workforce upskilling, training, and AI enablement
• Support for continued monitoring and ongoing uptime/readiness
• How to evaluate ROI potential and audit processes to guide roadmap and investments

SMBs are also investing more but still need trusted implementation help. Salesforce's SMB Trends recap says 75% of SMBs are already investing in AI, more than a third have AI fully integrated into daily operations, and 71% plan to increase AI investment over the next year. Salesforce's SMB report also lists "90% of SMB leaders believing AI will make operations more efficient, 78% saying AI will be a game-changer, and security ranked as the top concern." Microsoft's 2025 Work Trend Index points in the same direction: "Leaders expect teams to redesign business processes with AI, build multi-agent systems, train agents, and manage agents over the next five years."

The rise of MCP: agents take real actions

Many companies are rolling out new AI agents in their products, services, or enabling them with new APIs, MCPs, or SDKs. For example, Zapier's is one of the most recent, which opens up more options for agent-powered automations and workflows. The possibilities unlocked with these kinds of releases are exciting — new use cases enabled for businesses to use across their existing tech stack. Enabling agents to link across tools enables more reliable results with emerging AI functions. MCP and SDKs enable agents to take real actions across systems, and everyone is racing to unlock AI data ops at scale for deeper integration.

HubSpot launched a prospecting agent and many more AI features as they shifted from "Inbound to Unbound" — a dramatic move in their core messaging as they embrace agentic AI and enable it in their products. They have pushed toward opening more of their API up to MCP tooling as well.

Salesforce also just went fully headless. Everything on Salesforce is now an API, MCP tool, or CLI command — and agents can use all of it. With both of the biggest CRM providers shifting this direction, this is big news. It represents an important moment in their timeline and resonates with the mood of the market. The doors must be opened for agentic AI to operate in the tools businesses use already.

Our bet: this is not a minor shift for a few key players — it is the emerging way forward in SaaS and application development, and it mirrors the new model for the future of work, where humans and agents actually co-work in real-time. MCPs and SDKs are like the hands for agent actions. More will follow this path, until the next evolution makes us all adapt again.

Agents are not going away, and existing tools need to plug into them more seamlessly to enable real business operations. Those like Canva and others who have moved to integrate with LLMs or build AI natively into their products — like their Magic Studio — have seen stronger user adoption to their tools this past year.

Databox and Clay are other recent examples: they recently released their own MCPs allowing users to utilize their connected sources for use with Claude or GPT to get insights from their own data or tools. These releases have teams rejoicing. The use cases are really compelling when you can use your own data sources more reliably with these options.

AI developers, system architects, and others who have been building with AI have felt these roadblocks firsthand when trying to work with older tools, siloed data, and closed systems. The future ecosystem will be built on tools that are easily integrated with AI agents. Teams enabled by AI will be more productive. There is no such thing as too late while it is actively evolving and things continually change this quickly. The best time to jump in was already; the next best time is now. The market is measurably shifting toward a more responsible mindset as we collectively try to mitigate new risks.

The SMB angle: different stakes, different needs

The difference for SMBs and their approach to AI matters. SMBs do not need the same treatment that some enterprises do because the budgets and the scaling needs are so different. They are seeking more practical AI solutions like:

• Help with AI readiness: data quality, integrations, automations
• Better lead follow-up or processes to capture or recover revenue
• Better visibility in AEO (AI answers), because SEO seems to be shifting
• AI-assisted marketing and content generation to support lean teams
• AI workflows for specific tasks fit to their needs
• Smarter chatbots that understand knowledge bases and products or services
• Appointment, scheduling, and meeting assistance for better time management
• Reduce redundant tasks and help evaluate new risks

The growing cybersecurity threat layer

Finally, we cannot fail to mention the issue with growing cybersecurity threats. Enterprises will need to consider data, privacy, and security protocols and continually optimize those as new threats emerge. Businesses should not sleep on this. More sophisticated offensive and defensive attacks are on the rise.

Some say "Q-Day" — the point when quantum computers can crack modern encryption — may be approaching. Continuous Threat Exposure Management (CTEM) is becoming more common as a proactive measure to combat this problem. More cybersecurity roles are also opening up to combat the growing risks.

Anthropic engineers said Claude Mythos Preview is the first model to solve both AISI end-to-end cyber ranges. Mythos, available only to a closed circle of researchers under the Claude Mythos Preview program, just found the first public kernel memory corruption exploit on Apple's M5 chips. This was an important announcement in the space that had many speculating online about capabilities. Apple was among a small number of partners with early access.

With each new release, cybersecurity is evolving just as quickly. Those looking to add new agentic AI abilities need to consider security and guardrails before deployment to mitigate risks. CrowdStrike and others have posted recent announcements warning about the potential dangers of frontier AI capabilities creating new risks for teams to combat.

That makes ongoing security and risk management the final trend to watch as we measure the ways agentic AI is reshaping the future of work. Being proactive is the most strategic path. The costs of recovering from a breach or other incident might far outweigh the investment up front.

Closing the field note

This year is shaping up as the inflection point we have been talking about for the past three. Adoption is no longer the hard part. The work now is operational: orchestration, governance, the human-in-the-loop choices that decide whether agentic AI matures into infrastructure or collapses into incidents.

For organizations still figuring out where to begin, the work is usually not just picking a tool. It is auditing data quality, mapping workflows, scoping cost realistically, and building the orchestration discipline that lets AI ship without breaking things. Start with the fundamentals that go into good data quality, because that is the first key for smooth agentic operations. So is the choice to build vs. buy. Examining tradeoffs realistically is important — but to do that, you need to map the landscape and your internal processes.

The big squeeze is real. So is the opportunity. Field notes from the frontlines: more soon.

— Megan