The Path Is the Easy Half

Patrick Salyer sold Gigya before the 2021 peak. In February he published a survival guide for software CEOs living through what traders had started calling the SaaSpocalypse, and six months later he went back to check whether the advice held. Most of it did. His revision narrowed the question down to which path the company is taking, given that the old one has closed.

The guide is a strategy document and a good one. Ten items, each of them a decision about product, team, pricing, or buyer. Not one of the ten concerns control, and the omission is deliberate rather than careless, because he set out to answer what a CEO should choose. This brief takes up the half that follows the choosing. What follows covers what the market priced, the three paths it now rewards, the four ways a transition of this kind fails, and the capability all three paths turn out to share.

What the market priced

The first quarter of 2026 took roughly a fifth off the iShares software index and about a trillion dollars off enterprise software capitalization inside six weeks. A Jefferies trader named it, and the name stuck. Salesforce fell around thirty percent across the year and Workday around a third. Software forward earnings multiples dropped below the S&P 500 for the first time in the modern era.

The revenue multiples tell the same story with more precision. PitchBook's set of ninety-nine public enterprise software companies moved to a median of 3.3 times trailing revenue at the end of March. The same set read 4.9 at the close of 2025 and 6.2 at the close of 2024. Aventis Advisors reads 3.4 times over a different sample and the same quarter. Two independent samples land within a tenth of each other.

One correction to the shorthand. The move from thirty times to three times describes the best names in 2021 against the median today, so it compares two different things. The median public software company peaked near eighteen times, not thirty. The direction is right and the magnitude is close enough to plan against, and the distinction matters when a board asks what the company was ever worth.

What investors repriced is the assumption underneath a seat. The selloff followed a frontier lab shipping a product rather than an earnings miss, and the accounts still disagree on which release did it. Investors watched a lab ship general-purpose work automation and marked down every business that bills per human user. Goldman Sachs Research had published the same conclusion in slower form. The software market keeps growing through the end of the decade, and by 2030 more than sixty percent of it may run through agents rather than through seats.

The market has since sorted rather than simply fallen. Windsor Drake's first-quarter read puts top-quartile AI and security platforms at eleven to fifteen times while mature horizontal software compresses toward historical lows. By August the dispersion inside single categories had widened further. Companies with a working AI playbook are being re-rated upward, and the gap between them and their neighbors is now larger than the gap between sectors. Something is being rewarded, and it is visible enough to describe.

Three paths the market is rewarding

Salyer's revision names three, along with a fourth answer for a company that cannot commit to any of them.

Re-found around an AI-native outcome. Fin is the worked example. Intercom spent years at a flat valuation while its category commoditized around it. Then it built an agent that resolves support conversations end to end and priced it at ninety-nine cents per resolution. At announcement the company ran about four hundred million in annual recurring revenue. Roughly three hundred million of that was the legacy seat-based business, close to flat. The agent accounted for the other hundred million and was growing around three hundred and fifty percent. Salesforce paid about 3.6 billion dollars, which prices the blended business under nine times revenue. Strip out the flat legacy line and the growing line prices in the mid-thirties. The buyer paid for the part that does not bill per seat, and the company changed its name to match.

Move into the path of AI demand. MongoDB is the working example rather than the finished one. Coding agents now participate in architecture decisions, and MongoDB has spent the year making itself reachable by default from the tools where those decisions happen. Its managed MCP server powers native plugins inside Claude Code, Codex, Cursor, Grok Build, and Devin. The existing MCP server runs above thirty thousand npm installs a week, and Citi named the company a top software pick for the year. The path suits infrastructure and data companies more than application companies, and it rests on a judgment about where demand routes rather than on a change to the product.

Turn the system of record into a system of action. Samsara launched Agent Studio in June. Customers build agents over Samsara's own operational data, starting from prebuilt templates or from scratch, without a developer. The path leans on an asset the incumbent already has and a newcomer does not, which is a decade of proprietary operational history and the customer relationship that produced it.

The fourth answer is to sell, and it is an honest one for a company that will not do the work. The option that has stopped working is the status quo. The market has now had six months to look at companies holding position, and it has priced them accordingly.

Why the path is the easy half

Read the three examples again for what each one cost. Fin needed a new pricing system, a new billing model, and a buyer who had never bought from it before. MongoDB needed to ship and support integrations inside five agent tools it does not control. Samsara needed to put policy, preview, and an audit trail into a product before it could let customers point agents at fleet data. Each of those changes how the company works rather than what it sells.

The four challenges below are how a change of that kind fails. Three are already inside most software companies. The fourth arrives at exit, in the room with the buyer.

1. Ungoverned agent deployment is a self-inflicted risk

Deploying agents broadly without identity, bounded permissions, or evidence of what they did is the operational equivalent of hiring fifty people and handing them production credentials with no job description. A company operating this way has roughly six months from the point its agents reach production scale before the approach produces a material failure of its own. The failure tends to land at the worst moment, during a sale process or an audit.

Uber exhausted its entire 2026 AI budget by April, four months into the year, as coding-agent token consumption outran any measure of shipped outcome. By May, Uber's engineering organization had published the governance the rollout had skipped: an agent registry, a cryptographic identity for every agent, and a gateway mediating each call into internal systems.

Salyer is right that the engineering organization has to be rebuilt around agents, and his revision puts the target at roughly a third of current headcount. That rebuild is where this challenge lands hardest. Five engineers each running ten agents ship faster than fifty engineers, and they also produce fifty unattributed actors inside the repository unless someone decided in advance who owns each one. The answer is not to slow adoption; it is to make governed execution the precondition for scale.

2. Competitors will turn agents into a widening execution gap

While one company in a category struggles to govern AI, another will deploy it across product and operations and improve at a rate the first cannot match. The competitive unit has shifted from the feature to the speed of the learning cycle behind it. A market-leading product buys less protection every quarter, and the gap compounds because every faster cycle returns more feedback, more operating evidence, and more reusable knowledge for the cycle after it.

Retail already ran this experiment. Shein produces initial runs of one to two hundred units, reads demand in days, and reorders only what sells, while the traditional apparel supply chain still runs six to nine months from forecast to fulfillment. That advantage re-sorted a global category before agents existed.

The benchmark in the revision is five to ten times year-over-year growth, set there simply to keep pace with AI-native competitors. A number that size is the outside view of this challenge. Nobody reaches it by working harder inside the current cycle time.

3. AI-native entrants will attack the pricing unit

The third challenge comes from companies built around what Baser calls the 10–100–3 model: fewer than ten people generating one hundred million dollars in annual recurring revenue within three years of inception. Cursor crossed one hundred million in revenue roughly twenty months after launch with a team of a few dozen. Lovable did it inside eight months with forty-five people. Each generation ships leaner than the one before, and each is AI-native from the first commit.

The pressure lands on the pricing unit itself. Seat-based revenue assumes a human behind every license. When a customer's agents absorb the work of five licensed users, per-seat revenue compresses at renewal even while the product stays fully competitive, and AI-native entrants sharpen the squeeze by pricing on usage and outcomes. Net revenue retention registers the erosion two or three quarters before a deal is lost.

Salyer already assumes the compression. He tells CEOs to give the agentic product away to existing customers, to stop optimizing for the upsell, and to pay commission on transitions rather than on renewals. Those are the instructions of someone who has concluded the seat is finished. The compression arrives whether or not a company chooses the migration, which is why Fin's legacy three hundred million sat flat while the agent line ran.

4. Frontier-model providers are a longer-term threat

Eighteen to thirty-six months out, the frontier labs appear to have entered a recursive cycle in which each model accelerates the next, so dramatic gains from more than one provider should be the planning assumption. As intelligence commoditizes, the providers move up into the application layer, and their conduct to date settles the question of neutrality. In January, OpenAI and Anthropic entered healthcare within a week of each other. In April, Anthropic launched Claude Design into Figma's market. Figma's stock fell seven percent on the day, inference costs have since weighed on its margins, and one of its own model suppliers now sells against it.

The challenge showed up in the tape before it showed up in anyone's profit and loss. Public investors set off the first-quarter repricing on a lab's product release rather than on any operating result at the companies that fell.

The survival guide tells CEOs to inventory their real advantages, and it names distribution, data, and customer trust. This challenge asks which of those survives a supplier becoming a competitor. A buyer will run the test directly. Strip the rented model out of the company and ask what remains. Where the answer is a user interface and a customer list, the buyer prices the franchise as exposed. Where the answer is a domain and entity model, a governed execution system, a proprietary evaluation corpus, and the evidence of what worked, the company owns a layer the labs cannot ship past. That layer carries the multiple.

What every path requires

Look again at what Samsara actually shipped. Agent Studio lets a customer load company policies and documents as the knowledge base behind an agent. It previews the agent's behavior before the agent goes live, holds it inside boundaries the customer sets, and tracks outcomes on a dashboard afterward. Policy sits in the agent's path, and the record survives the run. Samsara built that because pointing an agent at fleet operations without it is unsafe. Every company on any of the three paths needs the same capability internally, and almost none of them have built it.

The same requirement sits under all four challenges. Each one is a race between the pace a company can move and the control it can prove, and each is decided in the operating model rather than the tool stack. A company cannot buy its way out with the same tools its competitors are buying. The system that answers all four is the same in every case: govern execution so speed is safe, prove outcomes so scale is earned, and build the owned layer above the models so the models stay replaceable underneath it.

That system has two halves. The first is a software factory: specification-driven delivery in which agents build against approved intent and acceptance gets verified rather than assumed. The second is a system of intelligence, a Control Plane that puts policy and evidence in the path of every agent call. What the factory ships then runs in production under authority the company can defend. Neither half can be purchased. The build order decides whether the first governed workflow takes a quarter or a year.

For a board or an investor, three questions do most of the work. Which agents are running in production, who owns each one, and what evidence exists of what they did last month. What net revenue retention looks like by cohort once seat counts are separated from price. What the company would own if its model supplier became a competitor. A management team that answers all three has done the work.

Choosing the path takes a quarter of argument and a board vote. Building the capability to execute it takes longer, starts earlier than anyone wants, and decides which side of the dispersion holds the company at the end.

Sources and notes

Median public software revenue multiples through March 2026 come from PitchBook via FE International and from Aventis Advisors. The bifurcation read comes from Windsor Drake's first-quarter 2026 report. The agent share of the software market by 2030 comes from Goldman Sachs Research. Salyer's original survival guide and his six-month revision are the source for the three paths and for the ten recommendations cited throughout. Company facts on Fin, MongoDB, and Samsara come from the announcements and from contemporaneous trade coverage.

Two claims we hold loosely and flag as such. The precise trigger date for the first-quarter selloff varies across accounts, and we describe the cause by type rather than by date. The six-month figure for time-to-failure after ungoverned agents reach production scale is our own judgment from client work rather than a published statistic.

Next step. Baser Potential works with software companies and their investors. We locate the constraint holding a company at its readiness level, prove one governed workflow in production, and build the owned layer that carries the multiple. A first conversation takes an hour. Reach us at hello@baserpotential.com or www.baserpotential.com.