Console sold for $500M for automating the IT help desk
Palo Alto Networks reportedly paid $500M for Console, a two-year-old startup that automated password resets. The lesson for small teams isn't the price — it's the shape of the problem.
Palo Alto Networks reportedly paid $500 million for Console, a startup whose product resets passwords and grants Figma access. That's per TechCrunch, citing sources, on 2 September 2026.
I want to talk about the second half of that sentence, not the first. Password resets. App access grants. Routine troubleshooting. That is the entire product surface, and it sold for half a billion dollars roughly two years after the company was founded.
💰 The numbers, and what they imply about capital efficiency
Console was founded in 2024 by Andrei Serban. Its full funding history is short enough to fit in a table:
| Event | Amount | Lead / notes |
|---|---|---|
| Seed (June 2025) | $6.2M | Thrive Capital |
| Series A | $23M | Co-led by DST Global and Thrive Capital |
| Total raised | $29M | SV Angel, Abstract Ventures also on the cap table |
| PitchBook valuation, pre-deal | $157M | Third-party estimate, not a priced round |
| Reported exit | $500M | Cash and stock |
Two ratios stand out. $500M against $29M raised is roughly a 17× return on invested capital. And $500M against a $157M third-party mark is about 3.2×, which tells you Palo Alto was not buying a spreadsheet valuation — it was buying a position.
Key takeaway: The most valuable AI products right now are not the impressive ones. They are the ones sitting on top of a high-volume, low-judgement workflow that a company is already paying salaried humans to do.
One detail worth noting for anyone who thinks deals happen at arm's length: Palo Alto CEO Nikesh Arora was an angel investor in Console before the company bought it.
🔍 The product was deliberately unglamorous
Here is what Console reportedly automated, in full:
- Password resets
- Granting app access (Figma, Miro, and similar SaaS seats)
- Routine troubleshooting
Its named customers include Ramp, Flock Safety, and Scale AI — companies with enough headcount that tier-1 IT tickets become a real cost line, and enough engineering culture to trust an agent with them.
Notice what is not in that list. No open-ended reasoning. No "AI that understands your business." Every task has a verifiable end state: either the user can log in or they cannot. That is the property that makes an agent shippable, and it is the property most AI side projects lack.
If you cannot write a one-line test that says "the agent succeeded," you do not have an agent product. You have a demo.
🛠️ Why this shape is buildable from Colombo, and the chatbot shape isn't
I see a lot of Sri Lankan devs building general-purpose assistants. It's the wrong end of the market to enter with no capital. Compare the two shapes honestly:
| Narrow workflow agent | General assistant | |
|---|---|---|
| Success is measurable | Yes — ticket closed or not | No — "was it helpful?" |
| Needs frontier-model quality | Often no | Yes |
| Competes with OpenAI/Google directly | No | Yes |
| Buyer already has a budget line | Yes (IT support headcount) | No |
| Domain knowledge is a moat | Yes | No |
The second column is where an unfunded team gets crushed. The first column is where local knowledge is worth something. Sri Lanka has plenty of workflows that no US startup will ever touch: EPF/ETF filing checks, apparel-sector compliance paperwork, tea auction reconciliation, insurance claim intake in Sinhala and Tamil. Those are small markets globally and defensible markets locally.
A concrete filter I'd apply before starting: is a human currently paid a salary to do this task, and does it happen more than 50 times a week? If both are true, an agent has a price it can be sold at. If not, you are building a toy.
📉 The unit economics nobody models before they start
Console's whole business is the gap between what a ticket costs a human and what it costs an agent. If you're building anything similar, that gap is the model — and it is easy to get wrong, because token cost scales with context, not with how simple the task feels.
A worked example with my own assumed numbers (the source gives none — treat this purely as a sizing exercise):
| Input | Assumption |
|---|---|
| Tickets per month | 2,000 |
| Prompt tokens per ticket | 4,000 (system prompt + user history + tool schemas) |
| Output tokens per ticket | 500 |
| Retry/escalation rate | 15% |
The trap is the system prompt. A 3,000-token tool-and-policy preamble sent on every one of those 2,000 tickets is 6M tokens a month before a single user has typed anything. That's where margin dies, and it's also why prompt caching moves the number so much.
Rather than guessing, run your own numbers through our AI agent cost calculator, check your real prompt size with the token counter, and if you're caching a fixed preamble, model the saving with the prompt caching cost calculator. All three are free and need no signup.
⚠️ If tier-1 support is your job, or your employer's revenue
This is the uncomfortable part, and it lands harder here than in San Francisco.
Sri Lanka's BPO and managed-services sector sells exactly the work Console automated: first-line support, access management, ticket triage, billed per seat or per ticket. A $500M acquisition by a security vendor with seven acquisitions in 2026 is a signal that the buyers of that work are actively building alternatives to buying it.
The competitive picture also just narrowed. TechCrunch notes the deal leaves Sequoia-backed Serval as the de facto startup leader in AI IT service automation — a company valued at $1 billion after a $75M Series B in December 2025. When one player consolidates into a platform and the other is the last independent, pricing pressure follows quickly.
What I would not do is panic. Tier-1 volume shrinking does not remove the need for someone who can debug an actual incident, own a change window, or explain a failure to a non-technical director. Those skills get more scarce as the routine layer thins out.
💡 What this means for you
- Stop building assistants. Start building closers. Pick one task with a binary success condition and own it end to end.
- Pick a workflow with an existing salary attached to it. That salary is your price ceiling and your proof of demand.
- Model token cost before you write the agent, not after. The system prompt is the line item that kills margin.
- If you work in support or BPO, move up the stack now. Incident ownership, change management, and customer-facing judgement are the parts that don't automate cleanly.
- Read the deal correctly. Console raised $29M and exited at a reported $500M in about two years — not because the technology was hard, but because the problem was boring, repetitive, and expensive.
Bottom line: The AI money is in the tickets nobody wants to work, not in the demos everybody wants to watch. Pick something boring, make it verifiable, and price it against the salary it replaces.
Figures here come from TechCrunch's reporting, which attributes the $500M price to unnamed sources; Palo Alto Networks has not publicly confirmed the number. The cost table is my own illustration, not reported data.