monday.com spent a year learning an awkward lesson: sprinkling AI on existing workflows looks like powdered sugar on a cake—pretty, and easy to wipe off. The work platform, used by more than 250,000 companies, rebuilt its product around humans and agents sharing the same boards, launching in May 2026. By the time Anthropic published the case study on August 20, customers had logged more than five million agent interactions. The number is loud. The sharper question is why 'add an AI feature' failed—and what moat SaaS still has once Claude lives inside the board.
From “AI dust” to a full rebuild
According to Anthropic’s Claude blog case study (Aleksandra Todorova, 2026-08-20), monday.com’s shift came in three phases.
First came the industry default: embed summaries, classifiers, and automations into the original platform. In May 2025 the company ran an internal “AI month.” Adoption spiked, then hit a ceiling. Orly Stern Izhaki, VP of Product for the AI Works Platform, called the result “AI dust”—automations sprinkled on old workflows without changing the product’s core value. People tried features; sustained usage did not follow.
Second, leadership admitted that adopting AI features is not the same as becoming an AI company, and ordered a rebuild: agents as first-class teammates inside existing permissions, governance, context, boards, and triggers.
Third came post-launch proof. Since the May 2026 release, customers have logged more than five million agent interactions. monday says it serves roughly 250,000 companies from SMBs to the Fortune 500—figures presented as company statements, without an independent audit in the piece.
Chief product and technology officer Daniel Lereya called the agent-first shift one of the company’s most significant decisions: not bolting AI onto old flows, but reimagining what the platform should do.
Agents designed as coworkers, not a sidebar
monday noticed a pattern: enterprises want AI, yet stall with a chat window that runs beside real work. Their fix is concrete—every agent gets a name and avatar; colleagues assign work through triggers and mentions, as they would with a person.
The case maps four job families:
- IT: intake and triage, knowledge-base gaps, incident war rooms and postmortems
- HR: resume screening, interview scheduling, hiring coordination, feedback collection
- Marketing: competitive-signal monitoring and battlecard updates
- Executive office: meeting prep, turning decisions into tracked tasks, org-health scans
Claude shows up four ways: monday Agents (prompt-built), Bring Your Own Agent (Claude Managed Agents joining the board), pre-built Agents Store teammates (Claude plugins as specialists), and a Claude Coding integration (tasks assigned from the board, Managed Agents executing in the customer’s environment, results written back).
One marketing assembly line is the end-to-end sample: a Strategist Agent turns messy input into a structured brief; a Landing Page Builder on Claude Managed Agents produces a page variant; a Brand Reviewer checks brand and legal bars; a human manager makes a single publish-or-revise call.
Product value: the scarce asset is the worksite
On raw model skill, little here is novel—Claude already writes copy, reviews brand, and edits code. monday’s bet is welding agents into a worksite that already has permissions, data, and approval paths.
That is harder than shipping another enterprise chatbot, and stickier. When an agent can see boards, history, and role boundaries, it finishes accountable jobs rather than one-off chats. The fifth lesson in the piece lands cleanly: capability needs matching infrastructure; monday invested in monday DB so agents can run at organizational scale.
For developers, BYOA and the coding integration mean Managed Agents must accept board dispatch, status write-back, and human handoff. For enterprises, the pitch shifts from “another AI toggle” to “virtual coworkers you can govern.” For end users, the learning curve is @-mentioning an agent—if governance is truly transparent.
A cold note: five million interactions is a cumulative total, not retention. Productivity ranges like 30%–50% also appear in other Anthropic partner stories; treat them as joint narrative, not reproducible trials.
Competition: SaaS fighting the chat box for the entry point
For two years, horizontal assistants have stolen attention from vertical SaaS. monday’s answer is not a smarter chat; it is demoting the agent to a cell on the board—humble-looking, strategically sharp. Whoever owns task state and the permission graph owns agent scheduling.
Versus OpenAI’s more common “plugin / GPT into the app” story, this case stresses Claude Managed Agents, customer-owned execution environments, and store-packaged specialist teammates. Roughly: one side keeps building a universal dialogue layer; the other helps vertical platforms put agents on the org chart.
Salesforce, ServiceNow, and Atlassian tell adjacent stories. monday’s differentiating narrative is the failure first—“we hit AI dust, then rebuilt”—which reads more like a postmortem than a victory lap. The hard competition fact: once users @-mention agents on a board, swapping models is easier than swapping platforms. Anthropic wins inference share; monday wins the system of work. Allies for now; not guaranteed forever.
Risks, limits, and open questions
First, the piece is almost entirely a joint Anthropic–monday narrative: no independent evals, controls, or churn data. Five million interactions, 250,000 customers, and productivity ranges should be labeled party statements.
Second, coworker-shaped agents enlarge the blast radius. An agent that can open war rooms, reject candidates, or rewrite landing pages hurts more than a sidebar chat when prompt injection or config drift hits. Governance is named as a prerequisite; mis-operation rates and audit sampling are not.
Third, “agent-first” can become org-politics theater. monday admits the mental model is harder than the tech. Readers exporting the method should remember monday controls its own product; most companies are customers who cannot rebuild the vendor.
Fourth, deep Claude welding raises switching costs and transmits model-vendor safety incidents into business systems. That is commercial lock-in, not neutral plumbing.
Critic’s take
The most valuable line is not the five million figure. It is Izhaki’s: adopting AI features is not the same as becoming an AI company.
Through 2024–2025, B2B software paid tuition on AI dust. The industry is now correcting with costlier moves—information architecture, permission models, and who is allowed to finish work. Claude is the engine; monday is selling the cockpit. Read this only as another Anthropic customer-success note and you miss the sharper claim: work platforms are rewriting the org chart, and agents are a staffing reform, not a feature list.
Direction gets a pass; outcomes do not—not until we see completion rates, human rollback rates, and whether customers will pay for agent seats as their own SKU.
Next 6–12 months
More vertical SaaS will announce similar rebuilds; some AI-dust products will lose budget at renewal. Watch three signals: whether agent seats become a separate SKU; whether Managed Agents / BYOA land on default enterprise procurement lists; and whether incident disclosure keeps pace with capability marketing. When agents write battlecards, reject candidates, and open war rooms, silence is no longer a virtue.
For Anthropic, cases like this argue Claude can be a worker inside an enterprise OS. For monday, the wager is that whoever first makes agents into coworkers keeps the entry point in the next SaaS reshuffle. Win, and it is rich. Lose, and the dust is merely more expensive.