From Madhur Jain | Product & Market Analysis
Ten SaaS Categories AI Is Erasing in 2026, and Five That Get Stronger
On this page
Chegg's Academic Services revenue fell 57% in a single year, and the company named generative AI in its own filing. That is what erasure looks like. Most of the SaaS categories AI is supposedly killing are not dying at all. This post scores 15 categories against 4 questions, and the split has almost nothing to do with features.
Key takeaways
- Erasure follows a pattern, and the pattern is not "AI is in the product". The 10 categories scoring 4 or below on this test all charge for work a model produces directly, and none of them owns the authoritative record.
- Write authority is the single strongest predictor. Software that other systems obey survives. Software that reads other systems and presents the result is where margin goes first.
- The reported numbers are already this stark. Chegg's Academic Services revenue fell 57% year on year in the quarter ended 31 March 2026, while Datadog grew revenue 36% and Snowflake grew product revenue 30% over comparable periods.
- Being in an erased category is not a death sentence. RWS, a translation company, grew half-year revenue about 5% while AI products reached 32% of group revenue. Category scores predict pressure, not outcomes.
The 4 questions that decide whether a category survives
The short answer
Score any software category from 0 to 3 on 4 questions. Does it originate data? Does it hold the authoritative write? Does anyone face liability when it is wrong? And how much of the paid work is text a model can produce? Totals of 9 or more strengthen. Totals of 4 or less erode.
This post is written for founders and directors deciding what to renew. That reader converts on scoring frameworks and renewal maths, not on tool reviews, so everything below is built to be applied to a contract list.
Every question is deliberately about the category's position in the data flow, not about its roadmap. Roadmaps are cheap. A vendor can ship an assistant in a quarter, and it changes nothing about whether the vendor holds anything a model cannot regenerate.
| Question | Score 0 | Score 3 |
|---|---|---|
| Does it originate data? | It reads records other systems created | It is the only place the record is created |
| Does it hold the write? | Read-only surface over someone else's store | Other systems obey what it writes |
| Does anyone carry liability? | Nobody is fined if it is wrong | Named regulatory, audit or contractual exposure |
| Does the work resist substitution? | Almost all the paid output is text | Almost none of it can be produced by a model |
The 10 categories AI is erasing
These are the categories where I would shorten every contract term at the next renewal. Not cancel, shorten. The distinction matters, and the reasoning is in the final section.
| Category | Score | Fails hardest on |
|---|---|---|
| Study answer and homework subscriptions | 2 | The answer itself is now free at the point of search |
| Developer Q and A and community knowledge | 2 | Contributors stopped writing, so the data stopped renewing |
| General copywriting and content production tools | 2 | The paid output is text, end to end |
| Meeting notes and transcription point tools | 1 | No origination, no write, no liability, fully substitutable |
| Template website and landing page builders | 3 | Templates were the moat, and generation removes template scarcity |
| Keyword research and SEO content tooling | 3 | The metric it optimises is losing its link to revenue |
| Translation and localisation workflow | 4 | Substitution, though liability rises in regulated content |
| Tier 1 support ticketing sold per seat | 4 | Seats measure throughput, and throughput is what gets absorbed |
| Resume screening layers inside hiring tools | 4 | It reads documents rather than originating the hiring record |
| Standalone dashboard and BI presentation tools | 4 | It renders someone else's warehouse and writes nothing back |
Answer retrieval was a toll booth, not a business
Chegg is the cleanest case in the public record because the company says it plainly in a filing. Its 10-Q for the quarter ended 31 March 2026 reports Academic Services revenue down $59.6 million, or 57%, with total net revenue down 48%.
The filing names the cause. It cites Google's AI Overviews and rising adoption of free and paid generative AI as headwinds that reduced traffic, which reduced subscribers. That is a business whose product was standing between a question and an answer.
Stack Overflow is the same shape without the share price. New questions fell to 3,862 in December 2025, down 78% year on year, against a 2014 peak above 200,000 a month. Two trackers published different July 2026 figures, which is why this post uses the December number that a named outlet published alongside its data source.
Substitutable production work reprices first
The best evidence here is peer reviewed rather than anecdotal. A study of more than 3 million postings on a global freelancing platform, published in the Journal of Economic Behavior and Organization in January 2025, measured demand before and after ChatGPT.
Demand for substitutable skills such as writing and translation fell 20% to 50% against the counterfactual trend. Postings for "About Us" page writing dropped about 50%. Western European language translation fell about 30%. Machine learning programming rose 24% in the same window.
That study covers freelance labour, not software licences, and I am extending it by inference. The inference is defensible because both are priced against the same underlying task. If the task reprices, the tool that packages the task reprices with it.
Seat-priced human throughput
Support software is the category where the pricing model itself is the vulnerability. A seat priced against ticket volume is a bet that ticket volume stays human, and vendors are already moving to per resolution pricing because that bet is no longer safe. This mechanism is unpacked in detail in the analysis of what seat compression does to SaaS pricing.
My position is that seat pricing is not dying, it is separating. Seats attached to a person with approval authority hold their value. Seats attached to throughput do not.
Presentation layers over data you do not own
Dashboards, meeting notes and template builders share one property. They render or reshape information that lives somewhere else, and nothing downstream depends on them being correct.
That is why they score lowest. A category with no origination, no write and no liability has only convenience to sell, and convenience is precisely what a general model provides for free. The absorption pattern this creates across a stack is covered in the piece on point SaaS absorption into agent ecosystems.
What erasure looks like in reported numbers
Datadog reported second quarter 2026 revenue of $1.12 billion, up 36%, with about 4,720 customers above $100,000 of annual recurring revenue against roughly 3,850 a year earlier. Snowflake reported Q4 fiscal 2026 product revenue of $1.23 billion, up 30%, with remaining performance obligations of $9.77 billion, up 42%.
Remaining performance obligations are the number worth watching in that release. They are contracted revenue not yet recognised, so they tell you what customers have already committed to, and they grew faster than revenue did.
The comparison has an obvious weakness. These are different company sizes at different stages, and picking winners after the fact proves very little. The framework is the claim. The companies are illustrations of it.
The 5 categories that get stronger
Each of these becomes more valuable as agent deployment rises, which is the opposite of the assumption that AI compresses software spend uniformly. Every autonomous action needs data it can trust, somewhere legitimate to write, and a record of what it did.
| Category | Score | What carries it |
|---|---|---|
| Regulated systems of record | 12 | Payroll, ledger, claims and clinical records originate the truth and carry statutory liability |
| AI governance, audit and GRC | 11 | Obligations arrive on legislated dates, independent of demand |
| Identity and access, including non-human identity | 11 | Every agent is a new principal that needs credentials and limits |
| Governed data platforms and catalogues | 10 | Model output is bounded by the quality of what it reads |
| Observability and incident response | 10 | Autonomous systems fail in ways only telemetry explains |
Governed data and identity get scarcer, not cheaper
A model is only as good as the data it reads, which makes the governed store a bottleneck rather than a commodity. Snowflake's growth in contracted future revenue is consistent with that, though a single vendor is not proof of a category effect.
Identity is the less obvious one. Each agent granted access to a system is a new principal with credentials, permissions and an audit trail, and nobody has fewer of those next year than this year. The interoperability work that makes agents portable across tools, examined in the piece on agent interoperability standards, increases the number of identities rather than reducing it.
Compliance surface is the one thing agents create more of
The EU AI Act sets 2 August 2026 as the date for high-risk system obligations. A provisional agreement reached in 2026 would move Annex III high-risk systems to December 2027. Formal adoption was still pending at the time of writing. Treat the exact date as unsettled and the direction as settled.
This is the strongest position in the whole framework. Compliance demand is created by a statute rather than by a buyer's discretionary budget, which makes it the only one of the 15 categories whose demand does not depend on anyone being persuaded.
Write authority is the dividing line
Read layers compress, write layers do not
A read layer competes with whatever can read the same source. That set now includes a general model with a connector, which is a competitor that costs a fraction of a seat licence and improves without a roadmap.
A write layer competes with nothing, because the write is the record. Replacing it means migrating history, retraining staff and re-certifying controls. Buyers rationalise read layers first for exactly this reason, a dynamic covered in the analysis of app sprawl and rationalisation.
Agents make the audit trail more valuable, not less
The common assumption is that autonomous execution reduces the need for tooling around it. The opposite is happening. An action taken by software still needs attribution, reversal and evidence, and the volume of actions is rising.
This is why platform vendors are fighting so hard to be the place agents run rather than the place they are built. That argument is set out in the piece on whether Salesforce is a platform or roadkill in the agent era.
Where this framework is weakest
Three genuine problems, in order of how much they should worry you.
RWS is the counter-case
Translation scores 4 on my test, which predicts erosion. RWS, a translation and localisation company, reported half-year revenue of £360.3 million, up about 5%, with organic growth near 7% at constant currency and adjusted profit before tax up 33%. AI related products and services reached 32% of group revenue.
That is a company inside an eroding category growing by selling the thing eroding it. The framework scores categories, and companies are free to move between categories. Any founder reading this should treat that as the interesting result, not the exception.
The scores are judgement, not measurement
Every score in the chart above is mine. There is no survey behind them, no sample size, and no independent panel. I have published the rubric so the scores can be argued with, which is the most I can honestly offer without first-party data.
The second-order problem is that the empty middle of the distribution may be an artefact of a single scorer. A panel would almost certainly produce more categories in the 5 to 8 band, and that band is where most real decisions live.
Erasure runs backwards more often than the narrative admits
Klarna deployed an AI customer service agent in early 2024 and cut its outsourced support pool. Forbes reported in July 2026 that the company cut too aggressively, lost expertise on ambiguous cases, and rebuilt human capacity by hiring specialists back.
My reading is that Klarna does not disprove the direction, it prices the transition. The lesson for a buyer is that a category can be structurally eroding and still be premature to abandon in your own stack this year. The market-level version of that mistake is examined in the breakdown of the SaaS selloff.
How to score your own stack
Pull your 10 largest software contracts by annual value. For each one, answer the 4 questions from memory, without opening the vendor's website. If you cannot answer whether the tool originates data, that is the answer.
Then check one number per contract: active users in the last 30 days against seats paid for. A low score plus a low utilisation ratio is a renegotiation, and it is worth more than any category forecast, including this one.
Where a vertical tool scores higher than the horizontal one it sits beside, that gap is a consolidation opportunity. The direction of travel there is covered in the analysis of vertical AI eating horizontal SaaS.
Frequently asked questions
Which SaaS categories is AI replacing in 2026?
The categories being replaced share one trait: they charge for work a model can now do directly, and they do not own the record. Study answers, developer Q and A, general copywriting, template site builders, meeting notes tools and the screening layer inside hiring software all score 4 or below on the 4 question test in this post. Chegg's own filing is the clearest reported case.
Which software categories are getting stronger because of AI?
5 groups score 10 or higher. They are governed data platforms, observability and incident response, identity and access including non human identity, AI governance and audit tooling, and regulated systems of record such as payroll, ledger and clinical systems. Each one gets more valuable as agents multiply, because every agent needs data it can trust, a place to write, and a record of what it did.
How do I tell if a SaaS tool will survive AI?
Ask 4 questions and score each from 0 to 3. Does it originate data that exists nowhere else? Does it hold the authoritative write that other systems obey? Does someone face a fine, a lawsuit or a failed audit if it is wrong? And how much of what you pay for is text a model can produce? Totals of 9 or more survive comfortably. Totals of 4 or less are renewal risks.
Is per seat SaaS pricing dead?
Not dead, but repricing. Seat pricing survives where a seat represents a person with authority to approve something, and it erodes where a seat represents throughput a model can absorb. Support software is the visible case, since vendors are moving to per resolution pricing. Expect hybrid contracts rather than a clean switch, because finance teams budget better against seats than against consumption they cannot forecast.
Did AI kill Chegg and Stack Overflow?
AI accelerated a decline that was already visible in both. Chegg's 10-Q for the quarter ended 31 March 2026 reports Academic Services revenue down 57%, and names Google AI Overviews and generative AI adoption as causes. Stack Overflow's question volume had been falling since 2014, well before ChatGPT, then fell 78% in the year to December 2025. Both had weak defences before the shock arrived.
Should I cancel SaaS subscriptions because of AI?
Cancel on evidence, not on category. Run the 4 question test on your 10 largest contracts, then check usage data for the 3 lowest scorers before the renewal window opens. A low score is a reason to shorten the term and renegotiate, not always a reason to leave. Switching costs are real, and a tool that scores 5 today may score 8 once it holds your data.
Where to start this week
Take the 3 contracts renewing soonest and score them before you speak to the vendor. Bring the score into the call and ask the vendor to dispute a specific question rather than the total. A vendor who can show you what they originate is worth a longer term. A vendor who answers with a roadmap is worth a shorter one.
Then write your own version of the rubric with a fifth question that matters in your industry. Mine has 4 because 4 generalises. Yours should not generalise.
Related analysis
If seat pricing is the mechanism you care about most, the detailed version is in the piece on seat compression and SaaS pricing.
References
- Chegg Inc, Form 10-Q, three months ended 31 March 2026. Used for Academic Services revenue decline and the stated cause.
- devclass, Dramatic drop in Stack Overflow questions as devs look elsewhere for help, 5 January 2026. Used for December 2025 question volume and the 2014 peak.
- Journal of Economic Behavior and Organization, Winners and losers of generative AI: early evidence of shifts in freelancer demand, January 2025. Used for the 20% to 50% demand decline in substitutable skills.
- Snowflake, Financial results for the fourth quarter and full year of fiscal 2026, 25 February 2026. Used for product revenue and remaining performance obligations.
- Datadog, Second quarter 2026 financial results, 6 August 2026. Used for revenue growth and $100,000 ARR customer counts.
- RWS Holdings, Half year results statement, HY26, June 2026. Used for revenue, organic growth and AI share of group revenue.
- Forbes, How Klarna's AI agent strategy backfired but became a useful lesson, 16 July 2026. Used for the support headcount reversal.
The weakest thing about this source base: the 15 category scores are one analyst's judgement with no panel and no first-party dataset behind them. The company figures are all primary filings or company releases, but they were selected to illustrate a framework rather than sampled.
Related reading