AI and the Repricing of Software: What Buyers Are Actually Acquiring

In December 2021, Airtable announced a $735 million Series F (opens in a new tab) that valued the company at $11 billion pre-money. In August 2026, Bending Spoons agreed to acquire it (opens in a new tab) in an all-cash deal at an enterprise value of $1.285 billion. Airtable confirmed on 4 September (opens in a new tab) that the acquisition had completed.

It is tempting to read that as a collapse, and to blame AI for it. I would not, and neither do the documents.

The two numbers measure different things. The $11 billion was a pre-money equity valuation, set by investors buying new shares in a private round at the end of 2021. The $1.285 billion is the enterprise value of the whole company in a negotiated sale; once Airtable's net cash is added, Bending Spoons says the implied equity value is approximately $2.25 billion. The first figure reflects the financing terms and market conditions of one December. The second reflects a different buyer, a different structure and a very different year. Even the two equity figures are not like for like, and neither announcement says AI caused the difference.

Plenty else could account for the distance: lower growth expectations, higher interest rates, a harder market for financing, a more crowded category, the profitability a buyer can underwrite, and the structure of the deal itself. AI is probably part of the picture. No one outside the room can say how much of it.

What can be said is that Airtable was not a failing business. Bending Spoons reports annual recurring revenue of approximately $480 million as of June 2026, growing more than 20% year on year, and Airtable says more than 500,000 organisations use it, including 80% of the Fortune 100. The enterprise value is about 2.7 times that ARR, which is my own arithmetic from two reported figures. A company of that size and growth changed hands at a price its 2021 investors would not have recognised.

That is the signal I find more interesting than the headline. The market has moved from paying for a growth story to paying for what a buyer can underwrite: customers who stay, cash it can improve, and a product it believes it can run better. For anyone acquiring software now, the real question is what, precisely, is being underwritten. AI makes that harder to answer and more valuable to get right.

For most of the past two decades, a software feature cost engineers, quarters and management attention, and part of what an acquirer paid was the right to skip that queue. Tools that write code, connect systems and generate interfaces are now cutting the time and cost of building some of it. In my view, the part of a software company's price that rests on the idea that nobody could build it quickly is the part most exposed.

But a convincing demonstration is not an established business. A demonstration can now be assembled in days. A business has paying customers who rely on it every working day, security reviews it has passed, integrations that keep working when other people's software changes, billing and support processes, and years of edge cases that someone had to find and fix. The distance between the two is where much of the value of a mature product sits, and AI does not close it on its own.

The clearest recent example of the sum in practice is Salesforce and Fin. Salesforce already sold its own agent platform: it said Agentforce reached $1.2 billion in ARR in the first quarter of its fiscal 2027, up 205% year on year. In June it still agreed to pay approximately $3.6 billion for Fin (opens in a new tab), formerly Intercom, and completed the deal on 10 September (opens in a new tab). What it described buying was a base of more than 30,000 customer companies, a technical AI team, proprietary models built for customer service, and an agent that Salesforce says achieves an average resolution rate of 76%. That last figure is the company's own claim rather than an independent measurement.

My inference is that a buyer with the capability to build chose to pay for time, customers and a record of results that could not be produced quickly. That is build-versus-buy working as it should. AI made the build side look cheaper, and the buyer paid for the parts it could not reproduce. Our deal analysis of Salesforce and Fin covers the transaction in more detail.

It helps to separate a software business into layers and ask of each one how long it would take a capable competitor, now armed with AI tools, to match it. This is my own framework rather than an established standard, but I find it clarifies most conversations.

Becoming easier to replicate

• Interfaces and standard featuresDashboards, forms, templates and conventional workflows are quick to reproduce, and customers know it.
• PrototypesA working demonstration says little about reliability, scale or support.
• Thin integrationsConnectors that sit on documented interfaces and add no data of their own are easy to rebuild.

Remaining difficult to replace

• DistributionThe relationships, channels and installed base through which a product reaches buyers.
• Proprietary information and the right to use itData that is hard to source, structured over years and licensed in a way that allows it to be used.
• Embedded workflowsProducts customers have configured around their own processes, so that removing them means rebuilding operations, not swapping a tool.
• Trust and regulatory standingA security record, certifications and licences that take years to earn and seconds to lose.
• Switching costsContracts, history and integrations that make leaving expensive and risky.

None of the second list is new. What is new is how rarely buyers priced it separately. When building was slow and expensive, the code carried much of the price and bundled everything else with it. As code gets cheaper, the bundle comes apart and each layer has to justify itself. We made a related argument about data rights and provenance in October, and mapped the assets AI cannot easily buy or build in March.

The best test of that argument is to look at what buyers have paid most for. Two recent transactions stand out, and neither is a payment for code alone.

On 22 October, LSEG agreed to acquire Octus (opens in a new tab) for $3.535 billion in cash from funds advised by Permira and Octus's management. Octus sources hard-to-access credit documentation, including covenants and deal terms, and turns it into structured, searchable data inside workflow tools. LSEG says it serves 750 customers and more than 40,000 users, and expects 2026 revenue of over $300 million with an adjusted EBITDA margin of around 40%. Those are LSEG's expectations, not reported results. The deal is announced rather than complete, with closing expected in the first half of 2027, subject to regulatory approval. Our own analysis of the transaction sets out the detail.

The detail I find most telling is how LSEG describes access to the data. Customers connect to Octus's information “directly via Octus’ platform, an MCP connector or via AI workflows.” My reading is that AI agents are only as useful as the information they can reach, so a source of scarce, structured, rights-cleared data becomes more valuable as more work is done by software that needs it.

Set the price next to Airtable's with care. $3.535 billion against more than $300 million of expected 2026 revenue is a little under 12 times, which is my arithmetic on a forecast. Airtable's 2.7 times is enterprise value to ARR. Those are different measures, and Octus carries a much higher margin. Even so, the gap is wide enough that I think the difference between what the two companies sell, one a flexible tool and the other information that is hard to obtain, explains a meaningful part of it. I cannot prove how much.

The second case is infrastructure and distribution. ServiceNow completed its acquisition of Armis (opens in a new tab) on 20 April for approximately $7.75 billion in cash, funded from cash on hand and debt. ServiceNow says Armis tracks nearly seven billion devices in real time and is trusted by nine of the Fortune 10 and more than 35% of the Fortune 100. It describes the acquired asset intelligence as feeding its Context Engine, the layer intended to ground AI actions in the real state of a company's systems. The buyer paid for a data position, relationships with the largest enterprises and a role in its own AI strategy. See our analysis of ServiceNow and Armis.

Both are reminders that some of this year's largest cheques went to assets that a team with a coding assistant could not assemble over a quarter.

Within a week of Airtable's completion, Bending Spoons agreed to acquire Miro (opens in a new tab) at an enterprise value of $1.355 billion, or about $1.79 billion in equity value including net cash. It reports Miro's ARR at around $600 million, nearly 90% from business and enterprise customers, with more than 750 customers each above $100,000 in ARR and nearly four million paying users. That is roughly 2.3 times ARR on enterprise value, again by my arithmetic. Certain Miro shareholders also agreed to reinvest $295 million of their proceeds in new Bending Spoons equity. We looked at the numbers in our Miro analysis.

Two scaled businesses with real customers, both sold at low single-digit multiples of ARR. Whether that is an opportunity or a warning depends entirely on why the price is low. The test I would apply is whether the thing that has become cheap is the thing that is durable.

The same logic explains why a roll-up such as Bending Spoons can pay these prices. Its model, as it describes it, is to buy digital businesses, transform them and reinvest the earnings. That only works if the customer base is durable. The price it pays is a bet on retention, not on the code. Our study of the Bending Spoons model covers it in more detail.

Disruption will also create attractive entry points. Businesses with strong distribution or hard-to-replicate data can be marked down with the rest of a sector that the market has decided to fear. Strategic buyers who can tell the difference between a feature set and a franchise are the ones likely to benefit.

Revenue quality has always mattered. AI sharpens several of the questions.

Start with retention, and measure it two ways. If AI lets a customer do the same work with fewer people, seat counts can fall even while satisfaction stays high. Ask for retention in dollars and in seats, by cohort, and for the reasons behind each cancellation. Ask specifically whether any customer has replaced a module with something built internally. A single example is a data point. A pattern is a thesis.

Then look at margin. AI features carry real per-use costs, and if usage grows faster than price, gross margin shrinks as the product succeeds. Ask for gross margin before and after AI features, and for inference cost by customer. A fixed per-seat price on top of variable costs is a risk that does not show up in a headline ARR figure.

That brings us to ARR itself. Usage-based and consumption pricing blur the line between recurring and variable revenue. Ask how ARR is defined, how much is committed under contract and how much depends on activity that could fall away. Our guide to ARR versus revenue sets out the distinctions, and the difference between enterprise value and equity value matters in every comparison in this article.

Concentration deserves a fresh look too. A very large customer is also the customer most able to build its own alternative. Ask for the top ten customers' share of revenue, their contract terms, and how deeply the product is wired into their operations. Finally, measure switching costs rather than assuming them: how long implementation takes, how many integrations depend on the product, and how much of the customer's data lives only inside it.

For buyers, I would start by writing down, in one sentence, what the acquisition is for before building the model. If the sentence is mostly about features, the price should be tested hard. If it is about customers, data, distribution or regulatory standing, the work is to establish that those assets are real, transferable and secure. Value each layer separately, and use structure to match risk. A retention-linked earn-out is one way to pay more for durability that has still to be proved. Re-run build-versus-buy at every deal, and assume a competent team can copy the product. What is left is what you are really buying.

For founders preparing to sell, the work begins earlier than most expect. Buyers in 2026 paid most for what could not be reproduced quickly, so the preparation should centre on evidence of exactly that: documented data rights and provenance, retention by cohort, the depth of customer integrations, a clean security history and contracts that show how sticky the product is. Present the economics of your AI features honestly, including what they cost to run. A seller who leads with a feature list invites a buyer to price the product as something it could build itself. A seller who leads with a durable customer base and evidence invites a different conversation. Our guide to AI washing in SaaS valuations covers what buyers discount.

The code is the part of a software company a buyer can most easily imagine replacing. The customers, the data and the trust are not. The next few years of software M&A will be a sorting exercise between businesses that have only the first and businesses that have the rest, and the prices paid will show it.