Will AI Kill SaaS?

Which SaaS categories AI replaces first

Exposure depends less on the vendor than on the work the software supports. Routine, observable, rules-based work goes first; regulated, knowledge-heavy work last.

By The editors · Updated · 4 min read

Three customer-support agents with headsets work at their desks while a team lead stands behind them reading a printed sheet.
Support desks were among the first places where software priced per seat met AI priced per outcome.
Contents
  1. Why category matters more than vendor
  2. Bain's two questions
  3. The categories going first
  4. The categories going last
  5. Score your own stack
  6. A sequence for the next twelve months
  7. Questions buyers ask
  8. Sources

Every SaaS tool in a company's stack now faces the same question from its buyers: could an agent do this work instead of a person using this software? The answer differs sharply by category, and it can be worked out before the next renewal rather than after.

Why category matters more than vendor

The market reaction of early 2026 treated software as one block: Salesforce, ServiceNow, Microsoft, SAP, Workday and Atlassian all fell together after Anthropic's Cowork plugins were released⁠1. The evidence on the ground is more selective. Omdia's Lian Jye Su expected the long-term correction to hit feature-focused or easily replicated SaaS hardest, describing an existential problem for "undifferentiated and overvalued" offerings⁠1.

Bain's two questions

Bain's 2025 Technology Report scores each SaaS workflow on two independent axes⁠2:

  • How far can AI automate the user's tasks? Measured by indicators such as how structured and repetitive the task is, the cost of an error and whether the data is available.
  • How easily can outside AI penetrate the workflow? Measured by indicators such as how observable the workflow is from outside, how deep its proprietary data runs and how hard it is to switch.

The combination produces four scenarios, each with Bain's examples:

ScenarioAutomationPenetrationBain's examples
AI enhances SaaS ("core strongholds")LowLowProcore project cost accounting; Medidata trial randomisation
Spending compresses ("open doors")LowHighHubSpot list building; Monday.com task boards
AI outshines SaaS ("gold mines")HighLowCursor's code editor; Guidewire claims adjudication
AI cannibalises SaaS ("battlegrounds")HighHighIntercom tier-one support; Tipalti invoice processing; ADP time-entry approvals

Source: Bain & Company⁠2.

The categories going first

Tier-one customer support. Bain lists it as a battleground⁠2, and the pricing has already moved: Intercom charges $0.99 per Fin outcome, billing a resolution only when no further help is requested, and nothing when a conversation passes to a person without an outcome⁠3. A support team's software cost is starting to track resolved conversations rather than agents employed.

Back-office processing. Invoice processing and time-entry approvals are Bain's other battleground examples⁠2. In accounting more broadly, AI-native ledgers are appearing: Rillet reports that 99.7% of transactions post automatically on its platform, a vendor figure⁠4.

Glue software. Tools whose job is to move data between other tools are exposed because agents can call the underlying systems directly. Their pricing is changing too: n8n charges per workflow execution, however many steps a workflow has, and offers a self-hosted version⁠5.

Task boards and list building. Bain's "open doors" are workflows where the software stays but outside agents do more of the work through its interfaces, compressing what buyers will pay per seat⁠2.

The categories going last

Systems of record built on regulated or proprietary data, with deep domain logic and high switching costs, are the most protected⁠2. In most companies that means the general ledger, payroll, regulated industry systems and anything an auditor or regulator inspects. AI is added to these systems; it rarely replaces them in the near term.

Klarna, the best-known company to drop Salesforce and around 1,200 other SaaS tools, did not replace them with a language model. It consolidated its data into an internal stack first⁠6.

Score your own stack

Five questions per tool

  • Is the work it supports routine and rules-based?
  • Could an agent see and act on that work from outside the tool, through its interface or data?
  • Does the tool hold regulated or hard-to-replicate data?
  • How much of its cost is per seat, and how many of those seats do routine work?
  • Is an AI-native replacement priced by outcome, usage or company available today?

A tool that scores high on the first two and low on the third is a candidate for replacement or renegotiation at its next renewal. A tool that scores high on the third is a candidate for AI added to it, not for removal.

A sequence for the next twelve months

  1. Start with the battleground tools. Support and back-office processing offer the clearest measures, such as cost per resolution or invoices processed per hour.
  2. Renegotiate the open doors. Where outside agents can do the work, use AI-native alternatives as leverage on per-seat prices.
  3. Leave the strongholds alone, and add AI to them where the vendor offers it.
  4. Consider consolidation where several exposed tools overlap. OpenLazarus, for example, presents each of its modules (mail, phones, CRM, invoicing, bookkeeping and others) as replacing a subscription⁠7.

The end state is a software estate sorted by evidence rather than by headline: each tool marked as replace, renegotiate or keep, with a date and a measure attached. The main analysis ranks the AI-native replacements by category.

Questions buyers ask

Which software will AI replace first?

Software that supports routine, rules-based work that an outside agent can easily observe and act on. Bain's examples include tier-one customer support, invoice processing and time-entry approvals. Glue software that moves data between other tools is also exposed, because agents can call the underlying systems directly.

Which SaaS categories are safest from AI?

Software built on deep domain knowledge, proprietary or regulated data and high switching costs. Bain names project cost accounting at Procore and clinical-trial randomisation at Medidata as examples where AI enhances the software rather than replacing it.

Is CRM software at risk from AI?

Parts of it. Bain places list building in HubSpot among the workflows exposed to outside agents, while the CRM as a record of customers remains necessary. The likely change is in how CRM work is done and priced, with agents doing data entry and research, rather than the end of the CRM.

Sources

  1. New AI tools tank software stocks. Information Age (ACS), 10 Feb 2026. Accessed 8 Nov 2026.
  2. Will agentic AI disrupt SaaS?. Bain & Company, 23 Sep 2025. Accessed 8 Nov 2026.
  3. Fin pricing. Intercom. Accessed 8 Nov 2026.
  4. Rillet: the AI-native ERP. Rillet. Accessed 8 Nov 2026.
  5. n8n plans and pricing. n8n. Accessed 8 Nov 2026.
  6. “No, I don’t think it is the end of Salesforce”: Klarna CEO clarifies why it stopped using Salesforce. ITPro, 24 Mar 2025. Accessed 8 Nov 2026.
  7. OpenLazarus: the AI operating system your company owns. OpenLazarus. Accessed 8 Nov 2026.

Updated 8 Nov 2026: First published.

Further reading