What Dies, What Survives, and What Has Never Existed Before

I have spent the last several weeks writing about the mechanics of self-evolving agents: how the retention layer works, why it naturally becomes the service layer, what agencies built on it need, how the best ones differentiate, why execution traces matter more than historical data, and why the sovereignty concern is more solvable than people think.

All of that has been mechanics. How the engine works. This post is about where the car goes.

I am going to make specific predictions about what the service-as-software landscape looks like by the end of 2027. Some of these will be wrong. Predictions about technology transitions always are. But I am making them specific enough to be falsifiable, because vague predictions are worthless and unfalsifiable predictions are cowardice.

What dies

The "SaaSpocalypse" narrative that has wiped roughly $2 trillion from software stocks since January is directionally correct but categorically imprecise. It treats SaaS as a monolith. It is not. Some categories are structurally doomed. Others are structurally advantaged.

Single-purpose workflow wrappers

The clearest casualties are SaaS products whose entire value proposition is wrapping a user interface around a sequence of steps that a human performs manually. Task trackers. Survey builders. Basic CRM data entry. Email sequencing. Social media scheduling. Appointment booking. Invoice generation.

These products have no unique data, no network effects, and no domain logic that an agent cannot replicate through API calls. Their moat was convenience: it was easier to buy Calendly than to build a scheduling workflow. When an agent can build the scheduling workflow in minutes, the convenience moat evaporates.

Gartner predicts 35 percent of point-product SaaS tools will be replaced by AI agents by 2030. I think this is conservative for the categories I am describing. The market is already pricing it in. Atlassian is down roughly 35 percent from its peak. Monday.com down 40. HubSpot down 50. These are not speculative moonshots, they are mature, cash-rich platforms being repriced for maximum disruption risk.

The implementation consultant for standard workflows

This is the one I feel most confident about because the ACP retention layer directly eliminates the need.

A significant share of consulting revenue comes from customization: making generic software work for specific clients. Configuration, integration, workflow optimization, user training. This work exists because software cannot adapt itself. When the self-evolution loop handles the customization that previously required a $300/hour consultant, the specific engagement type disappears.

Not the consulting firm. The engagement type. The McKinseys and Deloittes survive by moving upstream to strategic advisory and organizational transformation. TSIA's 2026 State of Reports confirms this is already beginning: implementation work is commoditizing while strategy, design, and continuous optimization are becoming premium.

Per-seat pricing as the default model

This is a business model death, not a company death. Per-seat pricing collapses when the "seat" is an agent that makes 10,000 API calls per minute. TSIA captures the dynamic precisely: when AI reduces the number of human users needed, revenue tied to user counts declines. The vendor's success becomes its own revenue problem.

By the end of 2027, every major SaaS vendor will have at least one pricing tier that is not per-seat. The holdouts will be the ones with enterprise penetration so deep that renegotiating thousands of existing contracts is operationally prohibitive. PwC's M&A research notes that this transition introduces "short-term financial pain, new volatility in usage and revenue, and pressure to redesign metrics, incentives, and teams." This organizational pain is exactly why startups that were born into outcome-based pricing have a structural advantage.

What survives

Platform SaaS with deep data moats

Salesforce, Workday, ServiceNow are not going away. They are transforming from software that humans operate into infrastructure that agents consume.

Bain's Technology Report draws the critical distinction: SaaS products whose core role is deterministic system-of-record functionality, storing authoritative data, enforcing business rules, maintaining audit trails, become more valuable as agents proliferate. An agent needs a system of record to write to. It needs a compliance infrastructure to validate against. It needs an authoritative database to query.

The agents do not replace these platforms. They replace the human users who operated them through dashboards and forms. PwC's M&A analysis confirms this at the valuation level: the market is correctly repricing point solutions while preserving platform valuations.

Strategic consulting

Novel problem-solving, organizational design, M&A strategy, market entry planning. This work remains human because it requires judgment that current AI cannot replicate. The consultant who helps a CEO decide whether to enter a new market is doing fundamentally different work from the consultant who configures their CRM.

What changes is how strategic consultants work. They will use agents extensively to gather data, model scenarios, analyze competitors, and draft deliverables. A three-person team with agents will produce the output that previously required twelve. The work gets better. The teams get smaller.

Regulated compliance infrastructure

Regulators will require it. When an AI agent autonomously processes an insurance claim, approves a loan, or files a compliance report, someone must be accountable. The EU AI Act already requires transparency and human oversight for high-risk AI systems. The compliance infrastructure that bridges human accountability and agent execution is made more necessary by agents, not less.

What has never existed before

This is the section that interests me most.

The service company with software economics

By the end of 2027, there will be at least one company valued above $1 billion that delivers professional services entirely through AI agents, operates at software-like gross margins above 70 percent, and prices exclusively on outcomes.

This company will not look like a consulting firm. It will not look like a SaaS company. It will be a new category. Foundation Capital calls it "Service as Software." The label does not matter. The economics are unprecedented: outcome-level value delivery at scale, with margins unconstrained by human labor costs.

The verticals where this emerges first will be the ones with high workflow variance, expensive current customization, and measurable outcomes. My bet is insurance operations, commercial property management, or financial back-office services.

Evolved state as an asset class

When an AI agency's most valuable asset is not its code but its accumulated customer-specific evolution, the versioned prompts, tools, memory, and workflow adaptations that represent months of operational learning, what does M&A look like?

PwC's M&A research already notes that buyers struggle to price AI assets correctly. Finro's AI agent valuation report finds that valuation has shifted from "how intelligent is the agent?" to "how reliably does it behave as software?" None of these frameworks account for evolved state. It is neither a dataset (it was generated through operation, not collected), nor a model (it is versioned resource configurations, not trained weights), nor code (it is dynamic, not static). It is a new category of asset: accumulated operational intelligence, specific to each customer relationship.

I predict that by the end of 2027 at least one acquisition will explicitly price evolved customer state as a line item in the deal.

The sovereignty-as-a-service layer

In Your Consultant Already Knows Your Competitor's Playbook, I argued that provable tenant isolation is a stronger sovereignty guarantee than anything consulting has ever offered. I believe this capability becomes a standalone infrastructure business.

By the end of 2027 there will be at least one company selling provable tenant isolation and audit infrastructure as a horizontal service to AI agencies. Cryptographic provenance chains. Tenant-scoped registries with self-service audit. Compliance documentation for agent accountability. This is the Stripe play for the service-as-software era.

The evaluation function marketplace

I argued in What an AI Agency Actually Needs that the evaluation function is the product. Building a good one requires deep domain expertise. Most agent builders have engineering talent, not ten years of adjusting insurance claims.

A marketplace will emerge for pre-built, calibrated evaluation functions organized by vertical and workflow type. These become components that agent builders purchase and plug into their evolution loops, the same way they purchase foundation model access today. The companies that build them will be staffed by domain experts, not engineers.

The agent credential standard

SOC 2 certification tells an enterprise buyer that a SaaS vendor meets minimum security standards. There is no equivalent for AI agencies.

By the end of 2027, at least one credentialing body will publish a certification standard for AI service providers, covering tenant isolation rigor, evolution methodology, regression testing requirements, evaluation function calibration, human escalation protocols, and data deletion procedures. Enterprise procurement will demand this credential before signing outcome-based contracts. Whoever writes the criteria shapes the market.

The timeline

By end of Q4 2026: At least three major SaaS vendors launch non-seat-based pricing tiers. At least one mid-tier implementation consultancy publicly acknowledges AI-driven revenue decline in earnings. At least one credible industry analyst publishes a framework for evaluating AI service providers that includes evolution methodology criteria.

By mid-2027: The first service-as-software company crosses $50M ARR with software-like margins and outcome-based pricing. At least one M&A transaction includes evolved customer state as an explicitly priced asset.

By end of 2027: The agent credential conversation is active with at least one draft standard in public comment. Per-seat pricing is no longer the default for new SaaS contracts in at least two major categories. At least one sovereignty-as-a-service company has raised a Series A.

What I might be wrong about

The timeline might be too aggressive. Enterprise adoption moves slowly. Procurement cycles are long. Regulatory uncertainty creates friction. The BPO industry took fifty years to reach mature adoption. Even compressed, this might take five to seven years rather than two.

The incumbents might move faster than I assume. A Salesforce or ServiceNow that successfully transitions to outcome-based pricing with agent-delivered services could dominate in ways that leave little room for startups. The cannibalization dilemma is real, but some companies do successfully cannibalize themselves. Apple did it with the iPhone.

The sovereignty concern might be harder to solve in perception than in practice. If a high-profile data breach involving an AI agency occurs, even one that would not have been prevented by traditional consulting ethical walls, it could set back enterprise adoption by years. Trust is slow to build and fast to destroy.

The evaluation problem might be harder than I think. If evaluation functions cannot be made reliable enough for autonomous commit gates in most verticals, the self-evolution loop requires human-in-the-loop evaluation, which changes the economics. The loop still works but at human speed, and the cost advantage narrows.

These are real risks, not hedging language. I believe the direction is correct. I am less certain about the pace.

Why I am writing this

Predictions are bets. Publishing them is accountability.

A year from now I will revisit this post and score it. Which predictions came true. Which were wrong. What I missed entirely. This is how intellectual credibility compounds: not by being right about everything, but by being specific enough to be evaluated and honest enough to revisit the results.