A Profitable Company With $600M ARR Sold at 2.3x Revenue
On 10 September 2026, Bending Spoons signed a definitive agreement to acquire Miro at an enterprise value of $1.355 billion, about $1.79 billion in equity value including net cash. In January 2022, Miro had raised $400 million at a $17.5 billion valuation. The discount is roughly 90% on equity value and about 92% on enterprise value.
Miro is in good shape: around $600 million in ARR, nearly 90% from business and enterprise customers, close to 4 million paying users, more than 250,000 organisations, profitable, with about $435 million in net cash. The market now prices that business at about 2.3x ARR, a multiple that in earlier years went to stagnating legacy vendors. 24 days before the sale, Miro had moved its AI add-ons into its standard Enterprise licences at no extra cost. No public source quantifies how much of the discount comes from AI and how much from the general reset of software valuations since 2021. This article treats AI as 1 of 4 documented pressures on Miro, and the only one a product owner can act on.
How AI Is Replacing the Product
Miro’s core job is a shared canvas where teams draw diagrams and flowcharts, build mind maps and wireframes, cluster sticky notes and turn a workshop into a summary.
Every one of these outputs is now a prompt. A general assistant produces a flowchart of an onboarding process, a mind map for a product launch or a summary of 60 retro notes in seconds, inside a tool most companies already pay for. Miro’s own MCP page describes its AI functions in the same terms: turn a conversation into a diagram, roadmap, user stories, wireframes or timeline, and summarise a board. Trending Topics describes the path to the sale as running through new AI tools that “handle brainstorming and diagramming almost as a side effect”.
The pricing move confirms it. Miro sold AI Workflows and Prototypes as paid add-ons until 17 August 2026, then folded them into every Enterprise licence for free. When a function is available as a side effect elsewhere, its add-on price goes to 0. What a model cannot reproduce is the shared canvas with 250,000 organisations’ boards on it and the 250-plus integrations around it. That is what Bending Spoons paid 2.3x ARR for, and 2.3x is a legacy-vendor price.
Why This Is a Problem: The Miro Timeline in 4 Signals
Between 2020 and 2022, Miro grew from 5 million to about 30 million users and expanded its paying customer base by 550%. The $17.5 billion valuation priced a future in which the digital whiteboard was infrastructure. Then 4 things happened, in public.
1. Buyers consolidated their tools. From 2022, companies cut duplicate licences and preferred suites from Microsoft, Figma and Canva over stand-alone tools. The return to the office removed part of the use case.
2. Growth slowed and headcount followed. Miro cut 119 roles in February 2023 and reportedly about 275 more in October 2024, roughly 18% of staff. Users kept growing, to more than 100 million, but far below the 2020–2022 pace.
3. AI features stopped being sellable on their own. An “AI-first workspace” positioning, in the CEO’s words, an official MCP server, the acquisition of Reforge in March 2026, and then the free AI add-ons on 17 August 2026.
4. The market repriced the category. The S&P Software & Services Index lost 25% of its value between 12 January and 23 February 2026, after Anthropic introduced Claude Cowork and plugins aimed at professional workflows. Airtable, valued above $11 billion in 2021, sold to the same buyer for $1.285 billion about 5 weeks before Miro.
None of these is churn. Miro kept its customers and lost the premium a market pays for expected growth. The order is the lesson for anyone assessing their own exposure: pricing power went first, growth second, the multiple third. Revenue was still fine on the day the decision was made by someone else. 1 more detail from the acquirer: Bending Spoons states that AI is often central to how it transforms the businesses it buys, and Trending Topics reports that its in-house automation layer runs portfolio brands with a fraction of the previous staff. The technology that turned Miro’s function into a prompt is the buyer’s tool for taking cost out after closing.
Which Companies Face the Same Threat
The exposed group shares 1 property: the thing the customer pays for is an output a general model can produce from a prompt, without company-specific data. That includes many products that look nothing like a whiteboard.
| Category | What the customer pays for | Documented signal |
| Visual and document tools: whiteboarding, diagramming, note-taking, presentation builders, template libraries | A diagram, a page of notes, a slide draft | Miro’s AI add-ons went free on 17 August 2026. Miro joins a portfolio that already holds Evernote and Airtable, bought on the same thesis |
| Information and research databases | Search, summarising and drafting over published content | Thomson Reuters lost almost 18% in 1 day in February 2026, RELX 14.4%, Wolters Kluwer around 13% |
| Seat-licensed workflow software for calendar management, document search, sales, financial analysis, data queries and visualisation, legal review, marketing, customer support and product management | A person’s seat in a dashboard | Anthropic released 11 plugins on 30 January 2026, 1 per function; the tools target work that software vendors sell as core products |
| First-draft services: FAQ chatbots on public information, translation, content and copy production, basic design | A draft or an answer from public information | No priced signal yet. The output is identical to what a general assistant returns from the same public information |
A 1-minute test: if a sales engineer’s demo can be reproduced by a general assistant from a 1-paragraph prompt, the product is in the exposed group. The less exposed group is defined by the opposite property. Its output depends on data the model cannot see, such as prices, contracts, claims history and ERP records, or it carries a liability that a model cannot hold. That distinction runs through the 6 options below.
Solutions: 6 Options When AI Can Replace Your Product
1. Classify Your Revenue by Replaceability Level
Tag every feature or service line at 1 of 3 levels: prompt-replaceable, where a model produces the output without company data; context-bound, where the output depends on verified company data such as prices, specs, contracts or claims history; and accountability-bound, where someone is liable for the output and needs an audit trail. Then attach revenue and paid seats to each tag.
Applied to Miro, the split is visible from the outside: diagramming, flowcharts, mind maps and templates sit at level 1. Board history, 250-plus integrations and agent access sit at level 2. Enterprise security and compliance controls sit at level 3. Only the last 2 groups carry pricing power in 2026.
2 inputs make the audit concrete. First, a churn rate analysis by feature usage: which functions do customers stop using once they have a general assistant, even while the contract continues. Second, a seat-count check: if 1 agent does the work 3 people did, per-seat revenue at level 1 falls while the logo stays on the customer list. The audit costs about 1 workshop per business unit. Skipping it leads to the most common failure: an AI roadmap that builds level-1 features, the exact functions general models already do.
2. Move the Value From the Interface to the Data
When the interface becomes a prompt, the interface stops being the product. What Miro still owns is 250,000 organisations’ boards, plus the integrations and MCP server that make those boards readable by other systems. For a regulated business the equivalent asset is verified data: product specs, prices, policy wordings, contract terms, claims history. A general model either cannot see this data or gets it wrong.
A practical test: write 20–50 questions a customer would ask about your products or services, with the correct answers, and run them through a general assistant. Every wrong or invented answer marks a level-2 asset you own and the model does not. The next step is structuring that data into 1 schema, 1 update cycle and 1 retrieval layer. This is a data-engineering project measured in weeks, and it is the foundation every later option depends on.
3. Sell the Audit Trail and the Liability
Where a wrong answer has legal or financial cost, buyers keep paying for who is accountable. That means a documented audit trail, data residency, obligations under revDSG/FADP and the EU AI Act, and a named provider that a FINMA-supervised firm can put into its outsourcing register. This is level 3, and it is the part of an offer a general model cannot absorb.
The February 2026 sell-off shows the split inside 1 industry. The market repriced legal and data information providers overnight on the expectation that the drafting and summarising part of their offer becomes a prompt. The verified-source, liability-bearing part of the same offer is what procurement still signs for. In enterprise sales into DACH regulated accounts, the documentation is part of the product: DPIA support, hosting location, retention rules, model provenance and a written escalation path to a human. Vendors that can hand over that file in the first meeting shorten the compliance review where AI pilots in regulated firms often stall.
4. Make Your Product Callable by Agents
Miro’s answer to the model layer was to become a source for it: an official MCP server lets assistants and coding agents read boards as context and write results back. The commercial reason is on the buyer side. 94% of B2B buyers used LLMs during their purchase process in 2025. If your product or data cannot be reached inside an agent’s workflow, you are outside the workflow where decisions are prepared.
For a brand this means exposing structured data through MCP or an API with proper permissions, and being the source the agent calls rather than the screen the user opens. 2 trade-offs are real: access scopes and logging must be designed before exposure, and content reachable by agents needs the same negative list as any public channel, so the agent never surfaces what should stay internal.
5. Reprice Seats Before the Market Reprices Them
Seat-based revenue compresses when 1 agent handles work that per-user licences were priced for. Miro has close to 4 million paying users, and the separate price for its AI capabilities went to 0 on 17 August 2026. The question for any level-1 product is which comes first: your own pricing change or the customer’s seat reduction at renewal.
The revenue operations task is to model the current book under 3 seat-compression scenarios, 10%, 25% and 50%, and decide the pricing change while the customer retention rate is still high. Options with different risk profiles: a platform fee plus usage, outcome-based pricing per resolved case or processed document, or tiers defined by data access instead of headcount. Usage pricing adds revenue volatility. Waiting adds a repricing you do not control.
6. Own the Agent Layer on Your Own Data — the Lab51 Approach
For companies whose product or service is information work at level 2, the decision is buy vs make. Rent general tools per seat and accept that the differentiation sits in someone else’s model. Or build the agent layer on your own verified data and own it. Lab51 builds the second option for regulated DACH clients, on Swiss or on-premise infrastructure, with the model underneath treated as interchangeable.
The method, as applied in current client scopes:
- Knowledge audit and source mapping. The company’s own material is crawled and ingested: product data, FAQs, technical specs, policies, PDFs, ERP data. A negative list defines what must never appear in an answer.
- Normalised knowledge base with hybrid search. All data is structured into 1 schema and stored for semantic retrieval, with keyword plus semantic search so the agent finds exact item numbers and specs, not just similar text.
- Curated answer matrices. For high-stakes questions, for example comparisons against 3–5 named competitors, the agent retrieves pre-verified tables instead of generating freely.
- Benchmark validation. 20–50 must-get-right questions are agreed with the client and tested before launch. Accuracy is documented in a validation report.
- 1 knowledge base across channels via MCP. The same data layer serves the website, WhatsApp, Facebook Messenger, TikTok and Xiaohongshu, so there is 1 source of truth instead of 5 diverging bots.
- Timeline and operation. About 8 weeks for the knowledge base and comparison engine. Channel integrations take 1–6 weeks each, depending on platform approvals. Monitoring and monthly reporting continue after go-live.
The outcome is a durable asset the client owns: a structured, updatable data layer and a presence on the channels where customers ask questions. A level-1 product loses pricing power to the model. A level-2 data layer gains it, because the model needs it to answer correctly. A caveat for accuracy: this describes project design. Outcome metrics are measured per project after launch, and the benchmark exists so that performance is a documented number rather than an impression. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, mostly over unclear value and weak risk controls. The validation step is the answer to both.
Why Now: 3 Dated Signals
First, the exit math is public. 2 established SaaS companies, Airtable and Miro, sold to 1 buyer about 5 weeks apart, both at a low single-digit multiple of ARR on enterprise value. Any product at level 1 now has a reference price, and boards will use it.
Second, AI features stop being sellable within months. Miro’s AI add-ons became free inclusions on 17 August 2026. A roadmap that monetises AI features on top of a replaceable core has a short window before customers expect those features at no cost.
Third, buyers are already running the test on you. 94% of B2B buyers use LLMs in their purchase process, and 41% of German companies now use AI in their business. Your customers are checking which parts of your offer a model can do, whether or not you have checked.
What to do this quarter: run the classification from option 1, 1 workshop per unit. Pick the level-2 data asset with the most revenue attached. Start the data layer now. A validated knowledge base takes about 8 weeks, and the next planning cycle starts in January.
Miro did not fail. Its core function became a prompt, and the market has now put a number on what that costs: about 2.3x ARR for a profitable category leader. The brands that keep a growth multiple through this cycle are the ones whose product runs on data a model cannot reproduce and carries accountability a model cannot hold. Deciding which parts of your revenue are which takes about 1 day. Building the layer that survives takes about 8 weeks. Both are cheaper than finding out the price from a buyer.
Wondering if AI can replace your brand? Contact us for the guidelines tailored for your company.
