Software Categories AI Is Killing: A PM's Field Guide

This is a field guide, not a prediction. I am a CPO. Every category in this post is one I have either bought from, sold to, or built a competitor inside at some point in the last fifteen years. The pace of what is happening to software right now is faster than anything I have seen since the move to mobile, and most of the destruction is happening underneath product teams that have not yet noticed they are exposed.

The point of the post is to give you, the person actually shipping product in 2026, a way to read your category honestly. Which categories are already over. Which are reshaping. Which are safer than they look. And what to actually do this quarter inside each.

I am going to leave specific company names out. You can map them to your own category. The pattern is more useful than the body count.

The short version

Three forces are collapsing entire software categories at once. Search is becoming answers, so anything that lived above a click is at risk. Tools are becoming agents, so any product whose value is "we make it easier for a human to do this task" is at risk. Subscriptions are becoming outcomes, so any pricing model that charges by seat or by usage of the tool, not by the work the AI does, is at risk. Twenty-plus categories of software sit directly in the path of one or more of these forces. Some are already mostly gone. Some will look unrecognizable in eighteen months. A short list of categories has moats AI cannot cross. This post is the map, by bucket, with the action to take this quarter if you build in one.

For the deeper handbook chapters this guide leans on, see SaaS Is Becoming Service-as-Software, The Product Operating Model, When Not To Use AI, and the planning artifacts in The CPO Mandate 2026.

The three forces doing the killing

Before we walk through the catalog, it helps to name what is actually doing the work. Categories are not dying because "AI is disruptive." They are dying because three specific dynamics are unwinding the assumptions the categories were built on.

Force one. Search is becoming answer.

For twenty years the web was organized around the click. A user typed a query, got a page of links, and clicked into one of them. Every business model above that click ran on attention paid in clicks. That is now collapsing into a different shape: the user types a query, the answer engine reads ten sources, the user gets a paragraph with citations, no click required.

Google's own AI Overviews already cover a large share of queries and Anthropic, OpenAI, and Perplexity have all built consumer answer engines that pull from the same web. The traffic stops at the answer. Anything whose value was being a high-ranked link on a results page is being demoted to a citation inside an answer, or worse, omitted from the answer.

This kills any category whose distribution depended on SERP clicks. It also rewires how content is written: optimize for citation in the answer, not for click on the link.

Force two. Tool is becoming agent.

The dominant shape of enterprise software for thirty years has been "we give you a tool that makes your job easier." Spreadsheets, dashboards, CRMs, project management boards, ticketing systems, design tools, IDEs, all of them are tools wielded by a human. The PM job in this world is making the tool easier to use.

That assumption is collapsing. An agent can read your messy data, take an instruction in plain English, do the work, and report what it did. The human moves from operator to reviewer. The product moves from "how do I use this" to "what did it do." I wrote about that shift in detail in SaaS Is Becoming Service-as-Software, and I am living it as a CPO right now.

This kills the category of "tool that helps a human do task X" whenever an agent can be wired up to do task X end to end. It does not kill the surface area of oversight, audit, exception handling, and trust calibration. That is a new surface and it is rich. But the old operation-first GUI is fading.

Force three. Subscription is becoming outcome.

The seat-based SaaS pricing model rewards a vendor when the buyer adds users. That works when more users means more work being done. It breaks the moment one operator with agents can do the work of ten. The vendor's pricing then drops as the product gets better. That is upside down.

The next-generation pricing model charges per outcome. A successful translation. A resolved support ticket. A booked meeting. A generated piece of code that passed eval and shipped. The vendor's revenue grows with the work the AI does, not with how many humans are logged in to watch the AI do it. The buyer side of this shift is already happening, procurement teams at large enterprises now ask about outcome SKUs by default.

This force kills any business model whose unit economics depended on selling more seats. Even if the product survives the answer-engine collapse and the tool-to-agent flip, a per-seat-only vendor will be undercut by a per-outcome competitor who can deliver the same result for less because they took the AI productivity gain into pricing.

Now to the catalog. Three buckets.

Bucket one. Already gone or going fast

These categories sat on a single thin layer of value that an LLM call now provides for free or near-free. The category is not collapsing in slow motion. It already mostly collapsed and most product teams in these categories are working through some form of pivot or layoff cycle. If you build in one of these, your one task this quarter is to figure out which adjacent category you are pivoting into.

Q&A reference communities

The "ask a question, get an answer from a stranger" model is over for anything programming-related, and is shrinking fast for everything else. Question volume on the canonical sites has dropped to a fraction of what it was. The moat used to be aggregated answers over time. The aggregation no longer matters because the LLM can produce the same answer instantly, often better, and tailored to the asker's exact context. This quarter: if you build here, kill the question-asking flow and ship as a curated, citable knowledge dataset that the answer engines will pull from. Become the source the LLM cites, not the destination the user visits.

Online tutoring and homework help

The homework-help category was built on a labor arbitrage: cheap subject-matter expertise on demand. That arbitrage is now zero. The category will not survive in its current shape. What might survive is "verify my work" tutoring, evaluator-as-a-service for student outputs, and human coaching layered on top of agent tutors. This quarter: stop selling answers, start selling certification, accountability, and human coaching wrapped around the AI tutor.

Standalone grammar and spell-check

Every writing surface now has an LLM baked in. The standalone category, "open this app, paste your text, get suggestions," is being absorbed into the editor itself. The category that survives is style and brand governance: enforcing voice, terminology, and compliance across a team, integrated into the writing tools the team already uses. This quarter: stop building features that compete with the LLM inside the editor, and start building the governance and rule surface the LLM does not get for free.

Voice transcription as a service

Transcription used to be a paid service tier with human reviewers. Now it is a near-zero-marginal-cost API call. The category that survives is high-stakes transcription with compliance, custom vocabulary, and speaker diarization for regulated industries (legal, medical, broadcast). This quarter: move every feature that competes with raw transcription into vertical compliance, and price on outcome (a court-admissible transcript, not a per-minute fee).

OCR as a service

Same story as transcription. The horizontal OCR API is a commodity. What is not a commodity is vertical extraction pipelines: invoices, receipts, lab reports, signed contracts, where the output has to feed into a downstream system with structured fields. This quarter: stop selling the OCR layer and start selling the structured workflow on top of it.

Generic translation services

Generic horizontal translation is a commodity API. What survives is enterprise localization at scale with brand voice, terminology databases, in-context review, and outcome SLAs. I run product for one of these. The shift from "we sell you translation" to "we deliver localized content end to end" is the same shift this post is about, applied to one category. This quarter: stop selling words per minute. Start selling content delivered on a schedule.

Resume builders

A resume builder used to be a layout app with templates. An LLM can do better in one prompt. The category survives only as part of a larger "career assistant" or "match me to opportunities" service where the resume is one artifact, not the product. This quarter: if you sell resume building, ship a career-management product or get acquired by one. Standalone is over.

Standalone copywriting tools

Generic ad copy, headline, and product description generators are a single LLM prompt with a template. The category is mostly absorbed into general-purpose LLM interfaces and into the marketing platforms that need the copy. What survives is brand-specific generation systems with brand voice training, eval harnesses, and integration into the publishing destination. This quarter: stop selling "generate copy" and start selling "generate copy that ships through our integration with your CMS, your ESP, and your CRM."

ATS keyword optimization tools

Applicant tracking systems are themselves being absorbed by AI evaluators. Selling job seekers "keyword optimization to beat the ATS" is a market that disappears the moment ATS replaces keyword matching with semantic eval, which is happening now. This quarter: if you sell to job seekers here, pivot to interview prep, networking, and signal generation. The ATS keyword market does not exist in two years.

Stock photography for generic content

Generic stock photography for blog posts, landing pages, and slide decks is being replaced by generated imagery on demand. The categories that survive are licensed celebrity and editorial photography, where the value is the right to use a specific real image, and authenticated photojournalism, where the value is verifiable provenance. This quarter: if you sell generic stock, pivot to licensed and provenance-verified inventory, and price on rights cleared, not images served.

Standalone scheduling intermediaries

The single-purpose "send me a link to book time" category is being absorbed into the calendar layer and into general-purpose agents that can negotiate calendars on both sides of a meeting request. The standalone product is a feature now, not a category. This quarter: if you build here, ship a meeting-intelligence and follow-up product. Scheduling alone is over.

Generic chatbot platforms

The "build a rule-based chatbot for your website" category was already shaky before LLMs. It is now over. The market is being consumed by vertical agent platforms and by general-purpose LLM agent runtimes. This quarter: if you still sell rule-based chatbot infrastructure, pick a vertical (support, sales, recruiting) and rebuild on top of an LLM, or get out.

Bucket two. Reshaping at the core

These categories are not disappearing, but the shape of the product is changing so much in the next eighteen months that the team building it today is likely shipping a different product by end of next year. The work this quarter is to figure out which shape your version will take, and start migrating.

Search engines

The search engine itself is being rebuilt as an answer engine. The link-based SERP is becoming a small part of a larger response surface, dominated by direct answers with citations. The user behavior is changing fast: the click is no longer the goal, the answer is. For PMs building anything that depended on organic search traffic, you are now in the GEO business, Generative Engine Optimization. The optimization target is being cited in answers, not being ranked in links. This quarter: rewrite your top thirty pages with citable summary blocks at the top, add FAQ and HowTo schema, and add answer-extractable bullets near the lede. Measure citations in AI answers, not just SERP rank.

BI and analytics dashboards

Twenty years of BI investment produced dashboards that nobody actually reads on the cadence they need to. The conversational analytics shape, "ask a question in English, get an answer with the chart and the SQL it ran," is now technically realistic and is going to absorb most of the dashboard category. The dashboard does not disappear, but it stops being the primary interface. The primary interface becomes a chat where the answer is sometimes a chart. This quarter: if you build BI, ship a conversational answer surface backed by your existing semantic layer, and start measuring "questions answered" instead of "dashboards viewed."

CRM

CRMs in 2010 were systems of record. In 2020 they added systems of engagement (sequencing, dialer, etc.). In 2026 they have to become relationship intelligence: pulling signal from email, calendar, calls, support tickets, and the broader web, then surfacing what matters and acting on it. The data-entry-heavy CRM is over because no rep types into a CRM if the CRM can read their email. This quarter: if you build CRM, ship a "we will populate the CRM from your actual work artifacts" flow and replace the data-entry surface with a review and approve surface.

Helpdesk and customer support tier one

Tier one support, the part that handles "I forgot my password, my invoice looks wrong, where is my order," is being absorbed by support agents. The category survives, but the ticket queue becomes an exception queue, and the human work becomes coaching the agent, reviewing escalations, and handling the long tail. This quarter: ship a support agent over your existing knowledge base, instrument the deflection rate, and rebuild the agent dashboard around "what did the agent do and where do humans need to step in."

Email marketing platforms

The send-and-template ESP category is being squeezed from two sides: AI is generating the copy and creative inside the ESP, and outcome-driven competitors are pricing on engaged delivery, not sends. The ESP is becoming a campaign-generation, segmentation, and outcome-attribution agent that the marketer steers. This quarter: ship a "generate the campaign end to end and recommend the schedule" flow and move at least one pricing line item from per-send to per-engagement.

SEO tools

SEO as a category is being rebuilt as GEO. The tooling shifts from "rank tracker, backlink analyzer, on-page optimizer" to "answer-engine citation tracker, prompt coverage analyzer, schema validator." Search Console matters less. Citation share inside answer engines matters more. This quarter: ship the citation-share view and the prompt-coverage view alongside (or instead of) the SERP rank view. You will be ahead of the category by twelve months.

Project and task management

The category is reshaping into autonomous coordination: tasks, statuses, and updates are read from the work itself (commits, docs, calendars, threads) rather than typed in by a PM. The board becomes a generated artifact, not the source of truth. This quarter: ship "we pull status from the actual work" alongside your manual board, and start measuring the percentage of updates that arrived without human typing.

Knowledge bases and wikis

Wikis decay because nobody updates them. An agent can read the source artifacts (docs, decisions, code, messages) and answer the question, eliminating the wiki as a queryable destination. The category survives only as a canonical source layer the agent reads, not as a user-facing destination. This quarter: kill the standalone wiki product and ship as a knowledge layer that an LLM answers from, integrated into chat and search.

Survey and form tools

Generic form builders are commoditized. What is not commoditized is the agent that asks the right follow-up question, reads the open-ended response, and codes the qualitative data. The category reshapes from form builder into research agent. This quarter: ship an "interpret the open-ended responses" agent and an "ask the right follow-up" agent on top of the form. Price on insights generated.

Coding bootcamps and learn-to-code platforms

The category was built on a labor shortage of junior engineers. That shortage is in fast retreat as AI absorbs the work that junior engineers were hired to do. The bootcamp category survives only as senior-track and specialty (AI-product engineering, evals, agent systems). This quarter: stop selling "become a developer in twelve weeks" and start selling "become an AI-native product engineer in twelve weeks." That is a different curriculum.

Design and prototyping tools

The category is moving from "designer manipulates pixels" to "designer specifies intent and reviews generated variants." The tool that survives is the one with the deepest design-system enforcement, the strongest brand and accessibility guardrails, and the cleanest hand-off to working code. This quarter: ship a generate-and-evaluate flow inside your tool with your design system as the constraint set. Stop racing on canvas features. Race on system fidelity.

Bucket three. Categories with moats

These categories are not safe forever, but they have structural moats AI cannot synthesize on its own timeline. If you build in one of these, your job is not to panic. It is to use AI to widen the moat, not to assume it protects you.

Regulated vertical SaaS where compliance is the moat

EHRs, banking core systems, KYC and AML platforms, clinical trial software, GxP compliance. The moat is regulator approval, validated state, and the audit trail. AI can speed up everything inside these systems, but the regulator does not approve an LLM. Use AI to widen the moat (compliance copilots, validated test generators, automatic regulatory mapping) and you become harder to displace, not easier.

Identity, IAM, and security infrastructure

Identity is the trust anchor for everything else. The moat is integrations, certifications, and the network of systems that trust the provider. AI helps inside (anomaly detection, threat hunting, identity graph analysis) but does not change the trust topology. Build deeper into customer infrastructure, do not lateralize.

Hardware-attached software

Software downstream of physical hardware (industrial control, robotics, EV, medical devices, semiconductors) is protected by the hardware itself. The moat is the bill of materials, the supply chain, the certification, and the install base. AI accelerates the software, but does not generate the chip. Build the AI agent for the operator inside the hardware footprint, and you sell the agent because you sold the hardware.

Two-sided marketplaces with strong network effects

Marketplaces with cold-start dynamics on both sides (drivers and riders, hosts and guests, sellers and buyers) have an AI-resistant moat in the network itself. AI can compete on user experience and matching quality, but it cannot synthesize the other side of the market. Use AI to improve match quality, demand forecasting, and trust and safety. Do not rely on the network alone, but the network buys you time most categories do not have.

Data plumbing

Warehouses, pipelines, ETL, reverse-ETL, transformations. The moat is the integration surface and the customer's data living inside the system. AI helps build the pipelines but does not replace the data layer. Use AI to slash the cost of pipeline authoring and maintenance, and price the savings into the offer. The customer leaves only when something demonstrably moves them off their current stack, which is rare.

Financial settlement and payment rails

Card networks, ACH, clearing relationships, banking partnerships. The moat is the rail itself and the regulatory permissions that sit around it. AI optimizes inside (fraud, underwriting, treasury operations) but does not replace the rail. New AI-native payment surfaces will be built on top of the existing rails for a long time.

Industry-specific systems of record with high switching costs

Legal practice management, dental office systems, dispatch and logistics, agricultural management. The data, the workflow, and the integrations are deep. AI absolutely changes the user experience on top, but ripping out the underlying system of record is a multi-year migration that customers will not do for a cosmetic AI overlay. Use AI to build the agent layer on top of your existing data, and bury it deeper into the customer's operations.

How to read the signals in your own category

The categories above are the obvious cases. The harder work is reading your own category honestly. Three signals tell you which bucket you are in.

The first signal is how much of your product's value lives in being a layer over a single LLM call. If your value prop is "we are easier to use than a prompt," your moat is the GUI ergonomics, and GUI ergonomics is the cheapest moat to copy. You are in bucket one.

The second signal is whether your usage is in operation (the user clicks, types, drags) or in oversight (the user reviews, approves, corrects). If your top user actions are operation, you are exposed to the tool-to-agent shift, which means bucket two. If they are already oversight, you are further along the curve and the shift hurts you less.

The third signal is what your pricing actually measures. Per-seat priced products that depend on team-size growth are exposed to outcome competitors. Per-transaction or per-outcome pricing aligns you with where the buyer is moving. If you are still selling seats with no outcome SKU, you are exposed regardless of which bucket you sit in.

What to do this quarter

This part matters more than the catalog. The catalog is a frame. The frame is useful only if it changes what your team works on this week.

Pick one workflow your users do today inside your product. Just one. Look at it from the angle of "what would it take to have an agent do this end to end, with the user reviewing the result." Build a prototype. Ship it as a hidden flag to five customers. Measure whether the agent flow is more satisfying than the manual flow. This is the single most important experiment your team can run this quarter, in any category.

Add an outcome SKU. Even if you keep your seat-based pricing as the primary line, add one outcome unit that the buyer can choose. A successful translation. A resolved ticket. A booked qualified meeting. A passed eval. Get one customer on that SKU. The data you learn from pricing on outcomes will move your roadmap more than anything else.

Rewrite your top pages for citation. Not for SEO rank. For citation inside an answer. Add a citable summary block at the top. Add FAQ and HowTo schema. Add named entities. Measure citation share in answer engines, not just SERP rank. Even if you are in a moat category, you are losing top-of-funnel attention to the answer engines and you need the citation share to recover it.

Write the one-pager. Inside your team, write a single internal document called "What our product looks like in 2028 if we are honest about which forces are hitting us." Three pages. Not a strategy doc. A field note. Share it with two peers and ask whether they recognize the same forces from where they sit. If they do, that is the most credible signal your category is in motion that you will get.

The three forces are not waiting. The answer engines are already here. The agent products are shipping every week. The outcome pricing buyers are signing contracts now. The PMs who will look back on 2026 as the year their career inflected are the ones running these experiments this quarter, not the ones who waited for the category to be officially over.

Pick one thing from the four above. Do it this week. Then do the next one next week. Twelve weeks from now you will be inside a different product, in a different conversation, with a different set of options. That is the only output that matters.

For deeper handbook chapters on the shifts above, see The Product Operating Model, SaaS Is Becoming Service-as-Software, When Not To Use AI, The Anti-Backlog, and the planning artifact in The CPO Mandate 2026.