No.24: From Seats to Agentic Work

This week, the usual suspects of enterprise software made the same announcement in different costumes: Salesforce going “headless” at TDX (turning the platform into APIs and tools that AI agents call directly, with no human in the browser), Google relaunching its enterprise platform around agents at Cloud Next 26, Microsoft putting GPT-5.5 inside Foundry (their enterprise AI platform for hosting and governing models) as the default, Anthropic shipping Managed Agents (a hosted service that runs AI agents on the Claude platform), OpenAI unveiling Workspace Agents in ChatGPT (shared team agents for business workflows), and Adobe pulling CX Enterprise out of the drawer (their agent platform for marketing and customer-experience workflows). These launches amount to the same argument: seat-licensed application software is no longer the default unit of value, and the people who used to sell it are racing each other and the model makers to own the fabric of the agentic enterprise.

As I covered when "SaaSpocalypse" kicked off in February, the market is selling the lot, indiscriminately. In my experience the core issue is that Finance and the operating side of software live in different worlds. The framework most of the market is reaching for treats "AI is really good at Excel" as if it were a theory of both AI capability and enterprise software value, and applies one SaaSpocalypse discount across the stack on that basis. Software stocks are down c.21% on the year, with Salesforce down 30% and Adobe 26%. Meanwhile, Salesforce just disclosed that its Agentforce business hit $800m ARR, up 169% year-on-year. Both pictures are real: the agentic line growing fast, the seat-licence base eroding underneath; but $800m on a c.$42 billion revenue base doesn't yet decide which one prevails. The next eighteen months are the test.

That’s not to say the bear case isn’t half right. The thin, single-workflow apps that wrap a database with a UI - the kind of company whose entire defensibility was “we built the form” - are in real trouble. We know how it plays out: an agent calls the underlying system directly, the form-as-product becomes redundant, the seats stop being used, the company runs out of reasons to renew. But the bears have missed two things. First, software companies have the option to reinvent themselves on top of agents, and they are. Second, software is not one bucket. Systems of record, the platforms that hold a company's customer data, employee records, and core workflows (Salesforce, Workday, ServiceNow) can have a different fate from the thinner stuff sitting on top of them. Systems of record have the data and daily workflow to attempt the move. The deciding factor is whether they go on to own the orchestration layer that runs the agents acting on that data. Win it, and they become "systems of action". Lose it, and they get relegated to high-value agentic databases inside someone else's system of action. The gap between those two outcomes is most of the value at stake here.

Subscribe now

Salesforce Moves Down the Stack

Headless 360 is Salesforce voluntarily stripping out its own user interface. Marc Benioff: "Our API is the UI." His co-founder Parker Harris: "Why should you ever log into Salesforce again?" Salesforce was founded on "no software" - the 1999 slogan they used to kill the previous-generation incumbent. Headless 360 turns the same move on themselves.

Every Salesforce capability: reading customer data, creating opportunities, running workflows, sending Slack messages, is now exposed for any AI agent to call directly. No human needs to open a browser tab for any of it to happen.

Alongside Headless 360, Salesforce announced Agent Fabric: the layer that runs and governs agents inside an enterprise, including agents built by competitors. This reinforces their attempted pricing model change twelve months ago: away from pure per-seat to a hybrid that adds a per-action metric called Agentic Work Units (AWUs). The meter ticks every time an agent does something inside Salesforce. AWU volume hit 2.4 billion last quarter, up 57% QoQ (that’s the usage running underneath the $800m ARR figure). The enterprise buying pattern is shifting too: 75% of Salesforce’s top 100 deals last quarter included both Agentforce and the underlying data platform (Data 360), showing customers are starting to treat “agents-plus-data” as the new architecture they have to buy together. Per-seat is compressing while the agentic line is growing fast from a small base. Whether the maths gets to net-positive on total revenue long-term is the central financial question for the company over the next 12 to 24 months.

Salesforce is voluntarily destroying its own application layer to be reborn as the layer underneath whichever AI agents end up doing the work. The neat trick is that they get paid either way: whoever's agent does the calling, the meter ticks, and they can price third-party agent calls high enough to keep Agentforce as the path of least resistance. (Raphaëlle d'Ornano's Decoding Discontinuity Part IV is a very good strategy analysis of the move, with more detail from the latest investor results.)

Betting for Value Capture

The same pattern runs through every announcement this week, with vendors rebuilding around agents as the new user, racing to occupy a specific layer of the stack underneath them. Three different bets are being placed on where defensible value lives in the agentic era:

The control plane is the layer where AI agents are run, governed, and connected to enterprise data. Think of it as the operating system for agents. Whoever owns it owns the customer relationship, on the same logic that AWS owned the SaaS era by owning the underlying cloud. It is what Agent Fabric does for Salesforce. Google is going for it as hard as anyone: at Cloud Next 26 last week they unveiled Gemini Enterprise, an agent platform that retires their previous AI infrastructure (Vertex AI) as the main enterprise surface. Microsoft is positioning Foundry as the neutral platform for running and governing enterprise agents, wired into the Azure stack that enterprise IT already deploys (a structural advantage no AI lab on its own can match). Anthropic is playing the most stripped-down version of the bet with Managed Agents - their platform hosts other companies' AI agents and charges for the runtime. That's the whole bet: no apps, cloud business, or productivity suite attached to it.

The front door is the surface where employees sit to do agentic work - and what OpenAI is trying to own with Workspace Agents. The agent is built and run inside ChatGPT itself, with optional surfacing in Slack. The integrations are curated and controlled by OpenAI as they’re ignoring the open integration standard the rest of the industry is rallying around (MCP, the open standard Anthropic seeded and handed to the Linux Foundation last year). The strategic logic is straightforward: ChatGPT has consumer reach in the hundreds of millions of weekly users plus a real enterprise distribution motor. If employees only ever invoke agents through ChatGPT, and identity, permissions, context, and routing sit there too, the control plane race becomes somewhat moot.

The vertical workflow. Concede the platform layer entirely and defend on proprietary data and embedded process. Adobe’s CX Enterprise is the marquee example. They partnered with OpenAI on the model layer and designed its own agent platform to be embeddable inside Anthropic, Google, Microsoft, and OpenAI’s tooling. The bet is that owning the marketing-and-CX workflow itself is defensible even if the platform underneath belongs to someone else.

Microsoft is the only player credibly fighting two lanes at once: Copilot as the front door, Foundry as the platform. The structural position makes them the most dangerous full-stack competitor on paper - though their own shares are also in a slump and Copilot's proof of execution is still pending. Salesforce has explicitly conceded the front door (Headless 360 lets agents render in Slack, Teams, ChatGPT, Claude, Gemini, etc). They decided they had to pivot strategically to agentic enablement to defend the business, thus becoming the layer underneath whichever front door wins. Headless 360 is a strategic retreat to the plumbing layer as Salesforce races to make themselves unavoidable infrastructure before OpenAI and Microsoft lock the gates on the front door.

The company is betting that Agent Fabric becomes the control plane and Agentforce the default agent, so customers route through Salesforce when working with Salesforce data. The risk is that customers reach for ChatGPT, Claude, or Copilot as the front door AND control plane, calling Salesforce as a (very valuable, highly priced but relegated) agentic database via MCP. They would still capture the data premium, but the orchestration premium goes to an AI lab - which is the fatter pool. Headless 360 makes Salesforce callable from anywhere; whether it makes them the caller is a different question.

The Market is Pricing this Wrong (Again)

A single SaaS discount sits across the entire sector while the architecture has split into a stack of distinct economic positions: control plane, front door, vertical workflow, agentic database - each capturing value at a different scale.

For the current Salesforce discount of ~30% on the year to be right, three things have to be true at once: per-seat revenue eroding faster than AWUs can compensate, Agentforce losing the agent-buying decision to ChatGPT, Claude, or Copilot inside Salesforce’s installed base, and Headless 360 commoditising to plain data-fee pricing that caps the AWU upside.

The disclosed numbers contradict the first and third directly. AWUs at 2.4 billion units a quarter (+57% QoQ) show per-action revenue accelerating from a small base. The 75% Data 360 attach rate on top-100 deals shows enterprise customers paying for the bundled stack at premium prices. Agentforce ARR up 169% year-on-year is the early signal that the agent-buying decision is currently going Salesforce's way inside its installed base.

The bear case that survives that narrative: Salesforce captures the agentic database revenue but loses the orchestration premium to whoever owns the model that does the calling. That thesis remains unresolved, and is the legitimate reason for an above-sector Salesforce discount. But pricing the company today as if every bear claim is settled - including the ones the disclosed numbers already contradict - applies a SaaS-era framework to a stack the framework cannot read.

A company being steamrolled by agents does not voluntarily destroy its own UI. It does not invest billions in building a layer that governs competitors’ agents. It does not aggregate thousands of third-party apps and AI integrations into a marketplace it owns.

Five things that should keep us honest

First, AWUs are an unproven economic substitute for seat licences. Per-action pricing might not even survive as the dominant unit: it tends to get optimised away (batching, caching, routing around the meter), and buyers will push for outcome-based pricing once agents are doing meaningful work. AWUs may end up as an intermediate model that gets competed down or replaced before they ever reach the scale needed to compensate for seat-licence erosion. Even if they hold, the maths is steep: per-action pricing only outperforms per-seat when agents expand the work past what humans could do alone. If agents just do the same work faster, per-action at best recovers the seat revenue and adds nothing on top. Going from 2.4 billion AWUs a quarter to a level that compensates for ~75% of seat-licence erosion is a different game that requires upselling every client, one at a time. Investors who don't believe AWUs scale linearly with value are not being unreasonable.

Second, the orchestration piece is still in beta. Agent Broker, which decides which agent handles which request and enforces the rules, ships in June. Without it Agent Fabric is just marketing. Microsoft Foundry is already in production with a comparable governance pitch, which means every month Agent Broker stays in beta is a month of enterprise customers deploying agentic workloads on someone else's control plane. Microsoft has a clear production lead on the lane Salesforce needs to win; and the risk is that they get relegated to the agentic database tier before they get to compete for the control plane at all. Salesforce has been on the wrong side of similar delivery gaps on AI before: their previous AI suite, Einstein, launched in 2016 with similar fanfare and spent years underdelivering on what the keynote had promised. Two months of beta is not nothing.

Third, the senior AI talent is leaving the incumbents. Three Agentforce executives have departed in the past month: one to OpenAI, one to Sierra, one to Anthropic. Talent voting with its feet. Salesforce is fighting uphill against new entrants who get to build greenfield while it has a legacy business to protect and an installed base to convert.

Fourth, orchestration may not be a separate layer at all. The model labs are building it in: Opus and GPT already plan and route their own multi-step work, with audit reliability improving. If that trend holds, Agent Fabric becomes a compliance veneer over orchestration that's happening inside whichever model the customer is calling, and the orchestration premium goes to Anthropic or OpenAI as the model vendor. The case for the layer holding up rests on enterprises needing vendor-neutral governance and audit that survives model swaps; the case against rests on enterprises eventually standardising on one or two models, which makes vendor-neutral governance largely theoretical. How enterprises actually procure decides this, and the jury is still out.

Finally: hybrid will persist. The move from "agent recommends, human approves" to "agent acts unsupervised" at scale, as the default for enough workflows, will not happen this year and possibly not this decade. In the short term that's an opportunity to double-dip: charge per-seat for human UI access and AWUs on top for whatever agent-driven work the customer adds. The open question is pricing and usage of one relative to the other.

The Logic Holds for B2B Info

What is unfolding in enterprise software lands directly on B2B information services, a case I made in "AI and the Transformation of B2B Info Services" earlier this year. Once the agent is the consumer, every human-only access mechanism is a technical barrier to entry. They block the agent at the door, prevent it from retrieving information accurately, and the agent either fails or routes around your platform entirely. If Salesforce has decided that humans should not log into its UI, the same seam is moving for everyone selling research, intelligence, events, or data.

The control plane war is happening one floor up; the front door war one floor above that. For most B2B information businesses those fights are out of reach. The realistic positions are vertical workflow (proprietary data embedded in your customers’ workflows) or agentic database (be the callable data source agents reach for in your sector). The risk of relegation runs at multiple levels: lose the workflow position and you compete on per-call price as a database; lose the database position too and you’re commoditised data agents source from somewhere else, leaving you invisible to whoever’s doing the work. The implication: your value has to be exposed where agents can reach it (machine-readable, API-accessible), governed in a way that makes you trustable to enterprise agents (audit trail, identity, provenance, data residency), and priced in a way that survives the seat-licence erosion (per-call, per-action, per-decision, hybrid). Being visited is good, but being callable, structured, and trusted will matter more. Beyond that, accruing value comes from building a position as preferred agent, which depends on how tightly your data is wired into your customers' decision loop. Brand and raw data quality become secondary to default inclusion in the agent toolsets your customers rely on.

The objection is usually about brand visibility and direct customer relationship. The honest answer: this is partly an exercise in embracing disintermediation. Refusing it means being routed around by the agents doing the work: present in the data layer, invisible everywhere customers work and decide.

But proprietary data and embedded workflow are only the starting line. Building agents on top of your own data is the move that decides whether you become a system of action or get relegated to a database inside someone else's product. For B2B information services that means shipping sector-specific agents: research providers producing decision support, events businesses orchestrating year-round buyer-seller interactions, media delivering personalised role-based briefings, data providers building actions on top of feeds. The companies that ship those agents capture the workflow value on top of the data fee. The companies that stop at exposing their data become the priced-but-invisible layer that someone else's agent calls.

Which lane is realistic depends on four things. First, whether customers currently engage with your product as a destination (opening a portal, running embedded workflows on your side) or as a feed (calling your data into someone else's product). Second, whether your data is genuinely proprietary (either unique to you or transformed and curated in ways hard to replicate). Third, whether the workflow your data feeds into is decision-shaped: whether the data converges on a discrete moment of judgment, regardless of where the customer engages with it today. Research that informs an investment committee, intelligence that drives a procurement decision, event interactions that shape a buying cycle - these are workflows you can build agents around because you already understand the decision shape. A subscription that gets read and filed lives in the data lane no matter how good the analysis. Fourth, whether you have the engineering and AI capacity to ship and maintain agent products. All four lean yes and the workflow lane is open. Two or more lean no and the realistic ceiling is the agentic database lane - which means competing on data depth, governance, and per-call economics, with the workflow value going to whoever wraps your feed in their own agent.

The move Salesforce just made is the canonical playbook for what every system-of-record holder in B2B information services has to run, in some form, starting now.

About

I analyse AI progress beyond the headlines, focusing on enterprise execution, incentives, and real-world economic impact.