Salesforce in the Agent Era: Platform, or Roadkill?

Salesforce repositioned itself around autonomous agents and consumption pricing, charging roughly $2 per conversation rather than per seat. Critics argue the company's real value was always the system of record and never the interface, which means agents threaten the pricing model rather than the product. Both readings are defensible and they lead to opposite conclusions about the largest company in software.

The actual question, stated precisely

The debate is usually framed as whether AI kills Salesforce, which is not a question anyone can answer or falsify.

The answerable version is narrower. Does an agent that reads the data, decides and acts reduce the value of the product that stores the data, or increase it?

Those pull in opposite directions and both effects are real. Agents reduce the number of people who need a licence to look at a screen. They also increase the value of having one authoritative, well-structured place for the agent to read from and write to.

Which effect dominates decides the outcome, and the honest position is that nobody knows yet.

What the repricing actually changed

Charging per conversation rather than per seat is not a pricing tweak. It is an acknowledgement that the old metric stopped tracking value.

Per-seat pricing worked because software value scaled with the number of humans using it. When an agent handles a conversation, no human was involved, so no seat exists to charge for. Consumption pricing restores the link between usage and invoice.

The broader market moved the same way. Pure per-seat pricing fell to roughly 15% of the SaaS market from about 21% a year earlier, with hybrid models becoming the dominant approach.

Why the metric had to change

A pricing metric works when it moves in the same direction as customer value. Seats did that for twenty years because software value scaled with the number of people using it.

The moment a meaningful share of work happens without a person present, the metric stops measuring anything. Charging per seat for an agent-heavy workload is like charging per driver for an autonomous fleet, and no amount of discounting fixes a metric that has stopped tracking reality.

The system of record argument

The case for the incumbent rests on a distinction that gets lost in the agent conversation.

Software that presents information is exposed. Software that holds the authoritative version of what happened is not, because the authority is the product and an agent needs somewhere authoritative to read from.

An agent updating a customer record needs that record to exist somewhere consistent, permissioned and auditable. Building that is the unglamorous work of enterprise software and it does not become easier because the interface changed.

This is why Gartner's absorption projection covers point products specifically rather than software generally, a distinction examined in the analysis of category absorption. Systems of record sit at the protected end of that spectrum.

The comparison that clarifies it

Think about what happened to accounting software when spreadsheets became universal. The prediction was that anyone could build their own ledger and the category would collapse.

What actually happened is that spreadsheets absorbed the analysis and reporting layer, while the ledger itself became more valuable rather than less, because more people needed a single authoritative version to analyse against.

That is the shape of the bet an incumbent system of record is making. The interface commoditises, the record does not, and the record becomes more important as the number of things reading it increases.

Where the threat is genuinely real

Three exposures, and they are not the ones usually cited.

Seat compression. If a company needs fewer people to run the same process, per-seat revenue falls at exactly the moment the product proves its worth. This is the structural problem and it is why the repricing happened.

The interface layer separating from the data layer. If an agent becomes the primary way people interact with customer data, the value of owning the interface declines. The database remains valuable and the application on top of it becomes a commodity.

Data portability improving. Standards that let agents read from any system reduce the friction of moving between them. Integration depth was frequently what kept customers in place, a dynamic covered in the analysis of agent interoperability.

None of these threatens the existence of the product. All three threaten the price it commands, which is the more realistic risk and the one that shows up in a valuation.

What an agent actually needs from a platform

It is worth being concrete about this, because the abstract version of the argument convinces nobody.

An agent updating a customer record needs four things: a schema that does not change under it, permissions that determine what it may write, a conflict resolution rule when two processes touch the same record, and an audit trail showing what changed and when.

None of those is AI work. All of them are the accumulated, unexciting engineering that enterprise platforms have been doing for two decades, and none becomes easier because the caller is a model rather than a person.

Does the consumption pricing actually work

For a buyer, the question is whether per-conversation pricing costs more or less than the licences it replaces, and the answer depends entirely on volume.

The arithmetic is unforgiving in one direction. A high-volume operation with successful deflection pays substantially more under consumption pricing than it did per seat, because volume is the thing it has most of.

That is not a criticism of the model. It is a reason to model your own volume before signing, and to negotiate a cap or a committed tier rather than accepting pure consumption. Vendors grant these more readily than buyers expect.

What incumbents actually do well

The disruption narrative consistently underrates three things.

Distribution. Selling an additional module to an existing customer costs a fraction of winning a new one. An incumbent shipping an adequate agent to its installed base beats a challenger shipping an excellent one to nobody.

Procurement inertia. Replacing a system of record requires migration, security review, retraining and a business case. Adding a module to an existing contract requires a signature.

Compliance surface. Established enterprise software carries certifications, audit trails and data residency options that took years to build. A challenger must reproduce all of it before a regulated buyer can even evaluate the product.

The legal software case is instructive here, where a well-executed incumbent response held a strong position against a fast-growing challenger, examined in the analysis of legal AI competition.

What that case also showed is that incumbents lose when they treat the threat as a feature gap rather than a positioning problem. Shipping an AI feature into an existing product is necessary and it is not sufficient, because the challenger is not selling a feature. It is selling a different way of doing the work.

What would actually settle this

Three observable things, none of which requires a view on whether AI works.

Direction of seat counts at large customers. If licensed user counts fall across an installed base while revenue holds, the repricing worked. If both fall together, it did not.

Whether consumption revenue grows faster than seat revenue declines. This is the arithmetic the whole strategy depends on and it is visible in segment reporting if you read it carefully.

Whether data migration volumes rise. The system of record argument holds only while moving is expensive. A visible increase in customers actually migrating would indicate the switching cost is falling.

None of these will be announced. All three appear in reported results, quarter by quarter, for anyone willing to track them.

Where this reading is weak

Three honest problems.

I have argued the incumbent case more strongly than the challenger case, partly because the challenger case is already everywhere. That is a corrective rather than a balanced assessment, and it should be read as one.

The system of record argument assumes the record stays where it is. If agents make data migration substantially cheaper, the switching cost that protects incumbents falls, and that is precisely what interoperability standards are designed to do.

And consumption pricing transfers risk to the vendor as well as the buyer. A vendor whose revenue depends on conversation volume is exposed to customers becoming more efficient, which is an odd position for a company selling efficiency.

Frequently asked questions

Where to start this week

If you are evaluating consumption pricing on any platform, do the modelling before the conversation rather than during it.

Take your actual volume for the last twelve months, apply the per-unit price, and compare it to your current licence cost. Then do it again at twice the volume, because successful deployment increases volume rather than reducing it.

Take both numbers into the negotiation and ask for a committed tier with overage protection. A vendor that has modelled its own economics will have an answer ready. One that has not will tell you something useful by not having one.

Related reading