Mark Hendy

Mark’s Musings

  • The CFO Who Took the Business Through the Deal is Often the First Casualty

    The CFO Who Took the Business Through the Deal is Often the First Casualty

    Not the tidy version. The real, uncomfortable one.

    The investment team made representations. They relied on advisors, they wrote the investment plan, they presented it to the IC. Now the cheque is written and their credibility is on the line. Every week of underperformance is a question mark over their judgement. Every green light is validation.

    They project that pressure downward.

    The management team feel it. The CEO feels it. But the CFO feels it first, because the CFO is the one who has to explain why the numbers don’t quite match the investment plan.

    The CFO who took the business through the deal is uniquely exposed. During the process they had to be captain positive. “Here’s how we’ll unlock the value.” “Here’s why the churn is fixable.” “Here’s the evidence behind the margin expansion story.” They were a full partner in selling the deal.

    Now the deal is done. The investment team is nervous. The board is watching. And the numbers — as they always do in the first few months post-close — are telling a more complicated story than the investment plan told.

    Suddenly it’s the CFO’s fault. Not explicitly. But the questions get harder. The calls get more frequent. The patience gets shorter.

    The CFO often doesn’t survive it.

    And here’s the thing — sometimes that’s not even unfair. The CFO who sold the deal is not always the right person to deliver it. Those are different skills. Different temperaments. A different relationship with uncomfortable truths.

    So they leave. Or they’re moved on. Quickly, and quietly, and usually within six months of close.

    The Problem That Creates

    The PE house now has a problem. The CFO is the second most important hire after the CEO. You cannot run a board, manage a lender relationship, or credibly execute a value creation plan without one. The permanent hire — if they’re any good — is on six months’ notice somewhere else. You need time to get this right.

    That’s where the interim CFO comes in.

    The interim CFO isn’t a gap-fill. Done properly, it’s the thing that buys the business the breathing space to make a good permanent hire instead of a rushed one. Someone who can walk in, stabilise the investor relationship, take ownership of the 100-day plan, and leave the business better than they found it — without any expectation of staying.

    The Real Job

    An interim who has been there before — who has stood in that boardroom, managed that investor relationship, built that first management pack from scratch — gives the PE house something they desperately need in that moment: confidence.

    Confidence that the business is in safe hands. Confidence that the reporting will be credible. Confidence that they can take their time and get the permanent hire right.

    Speed kills. Patience wins.

    That’s the job.

    Mark Hendy is an interim CFO specialising in PE-backed businesses. He writes about finance, private equity, and the reality of post-deal life at markhendy.com. Connect on LinkedIn.

  • The SaaSpocalypse Is Real — But the Market Is Panicking About the Wrong Thing

    I spent last week reviewing the tech stack costs across three portfolio companies. The exercise used to be straightforward: count seats, multiply by price, negotiate volume discounts. This time, two of the three CFOs asked me the same question: "Should we be cancelling licences and moving to agents?"

    That question — and the speed at which it has gone from fringe to board-level — tells you everything about where we are in April 2026.

    $285 Billion in 48 Hours

    If you have been anywhere near a Bloomberg terminal this year, you know the term SaaSpocalypse. It started in late January when Anthropic shipped Claude Cowork with industry-specific agent plugins — legal contract review, financial analysis, sales automation — and the market did what markets do: it extrapolated to infinity.

    Bloomberg reported roughly $285 billion wiped from SaaS valuations in a single 48-hour window. Thomson Reuters dropped 15%. LegalZoom cratered nearly 20%. By mid-March, the IGV software ETF was down over 21% year-to-date, and analysts were calling it the largest AI-triggered repricing in software history.

    The logic was brutally simple. If an AI agent can do the work of five humans, why pay for five seats? The per-seat pricing model — the entire economic foundation of B2B SaaS since Salesforce invented it — was suddenly an existential vulnerability.

    What the Market Got Right

    Let me be clear: the structural thesis is correct. Per-seat pricing is dying. I have seen it in our own portfolio.

    One of our companies ran a pilot replacing three junior paralegals' document review work with an AI agent pipeline. The agent does not need a LegalZoom subscription, a DocuSign seat, or a Westlaw login in the traditional sense. It calls APIs, processes documents, and routes exceptions to a human. The annual software cost for those three "seats" — roughly £45,000 — dropped to about £8,000 in LLM API costs.

    That is not a marginal improvement. That is a different business model.

    The survey data backs this up: 40% of IT budgets are reportedly being reallocated from traditional SaaS subscriptions to agentic platforms and token-based usage. CIOs are not asking "how many employees will use this?" anymore. They are asking "how many tasks can this complete?" That is a fundamental shift in procurement psychology, and SaaS companies built on headcount-correlated revenue should be worried.

    What the Market Got Wrong

    Here is where it gets interesting — and where I think the panic has overshot.

    The market treated the SaaSpocalypse as if every SaaS company is equally exposed. They are not. There is a massive difference between a company that sells seats for humans to click buttons and a company that sells the underlying data, workflow engine, or integration layer that agents also need.

    Thomson Reuters does not just sell a UI for lawyers. It sells access to legal databases, case law, and regulatory intelligence. An AI agent doing contract review still needs that data. The delivery mechanism changes, the underlying value does not. Same story with ServiceNow — the Motley Fool piece this week calling it a bargain has a point. Workflow orchestration becomes more valuable when you have agents that need orchestrating, not less.

    The companies that are genuinely toast are the ones that were essentially selling a graphical interface on top of commodity functionality. If your product is a pretty wrapper around CRUD operations and your moat was user habit, then yes, an agent that calls the same APIs without the wrapper is an existential threat. But that is not every SaaS company — it is maybe 30% of them.

    What This Means If You Are a CFO

    Here is my practical take, from someone currently navigating this across multiple PE-backed businesses:

    Audit your stack ruthlessly, but intelligently. Do not just cancel licences because agents are trendy. Map each SaaS tool to what it actually provides: is it data, workflow, integration, or just interface? The first three categories will likely survive the transition. The fourth will not.

    Start modelling token-based costs now. The shift from per-seat to per-task pricing is real, but token economics are volatile and opaque. I have seen API costs swing 30% month-on-month as providers adjust pricing. You need a cost model that accounts for this, and you need someone on your team who understands it — not just a vendor's sales estimate.

    Watch the middleware layer. The real winners of the agentic transition might not be the agent builders themselves. Microsoft's Agent Framework 1.0, released last week, unifies Semantic Kernel and AutoGen into a production-ready orchestration layer. That is the plumbing that enterprises will standardise on. If you are making build-vs-buy decisions on agent infrastructure, this is the framework to evaluate first.

    Do not mistake a repricing for a revolution — yet. Most enterprises are still in pilot mode. The 40% budget reallocation figure is aspirational, not actual. In our portfolio, the company furthest along has moved maybe 12% of its SaaS spend to agent-based alternatives. The rest are running proofs of concept. The gap between "we are exploring AI agents" and "we have decommissioned Salesforce" is about three years of integration work and change management.

    The PE Angle

    For anyone in private equity, the SaaSpocalypse is creating a genuinely interesting buying opportunity. High-quality SaaS businesses with real data moats and sticky integration layers are trading at 2022-era multiples. If you believe — as I do — that the best SaaS companies will successfully transition to hybrid pricing models (seats plus tokens plus outcomes), then the current discount is mispriced fear.

    The businesses to avoid are the ones with high seat counts, low switching costs, and functionality that an off-the-shelf agent can replicate. You know the type: they raised a Series B on "AI-powered" features that were really just a ChatGPT wrapper bolted onto a form builder.

    The SaaSpocalypse is real. But like most market panics, it is painting with too broad a brush. The death of per-seat pricing does not mean the death of software businesses. It means the death of lazy software businesses. And frankly, most of those were overdue a correction anyway.

  • Iran Wants Bitcoin for Hormuz Tolls. Here’s Why That’s Not Really a Bitcoin Story.

    Iran Wants Bitcoin for Hormuz Tolls. Here’s Why That’s Not Really a Bitcoin Story.

    Iran has reportedly demanded that ships transiting the Strait of Hormuz pay a $1 per barrel toll — in Bitcoin. Whether this actually happens is almost beside the point. The signal is loud enough on its own.

    The Financial Times reported it. X ran with it. And for anyone paying attention to how money actually moves around the world, this is one of those moments you file away.

    Why Bitcoin? Because Nothing Else Works for This

    Think about the problem Iran is trying to solve. It needs to collect money from ships it doesn’t fully control, in a currency it can actually use, without the transaction being frozen, reversed, or sanctioned before it clears. Try doing that in dollars. Try doing it in euros. SWIFT can be cut off. Bank accounts can be seized. Assets can be frozen mid-transaction.

    Bitcoin can’t be frozen. It can’t be censored. There’s no intermediary to lean on. Final settlement takes minutes, not days. And crucially — no counterparty trust is required. You don’t have to trust Iran. Iran doesn’t have to trust you. You both just have to agree to use the same protocol.

    That’s not ideology. That’s just how the technology works.

    The CFO’s Perspective

    I spend most of my working life thinking about how money moves — how businesses get funded, how transactions settle, where the risks sit in a capital structure. Most of that thinking happens within a framework that assumes the dollar is the world’s operating system. An assumption that has served well for decades but is increasingly worth questioning.

    The weaponisation of the financial system is real and accelerating. SWIFT exclusions, asset freezes, secondary sanctions — these are now routine tools of geopolitics. They’re effective precisely because the global financial system is centralised. Centralised systems have chokepoints. Chokepoints can be controlled.

    When a sanctioned nation-state proposes settling a strategic toll in Bitcoin, it isn’t making an ideological statement about decentralisation. It’s solving an engineering problem. It needs a payment rail that doesn’t have a chokepoint. Bitcoin is the only thing that fits that description at scale.

    The Game Theory Is Already Running

    Jesse Tevelow wrote a long piece on the game theory embedded in this moment, and he’s right about the core dynamic. Once one significant nation-state uses Bitcoin for sovereign settlement — even partially — it changes the calculus for every other state. The competitive pressure to accumulate, or at minimum not fall behind, kicks in.

    The US has already moved. A strategic Bitcoin reserve was announced earlier this year. That wasn’t random. It was a recognition that the game had already started, and that sitting it out entirely carried its own risks.

    We’re now in a world where adversaries — nations that fundamentally distrust each other — can transact without requiring mutual trust. Only mutual adherence to a shared protocol. That’s a genuinely new thing. It has implications that will play out over decades, not months.

    What This Means for Business

    In the near term, not much changes for most businesses. The Strait of Hormuz toll proposal may come to nothing. But the direction of travel is clear, and it’s worth thinking through the second and third-order effects.

    If Bitcoin becomes a meaningful component of sovereign settlement — even for sanctioned or constrained nations — it establishes a precedent. It creates a parallel layer of global financial infrastructure that operates outside traditional banking rails. That layer will grow. It will attract liquidity. It will become harder to ignore.

    For PE-backed businesses with international exposure: the question of which payment rails to support, which currencies to hold, and how to think about counterparty risk in cross-border transactions is going to get more complicated before it gets simpler. That’s a treasury question. It’s also increasingly a strategic one.

    For finance functions more broadly: the era of assuming the dollar-based correspondent banking system is the only game in town is ending. Not quickly. Not completely. But directionally, the trend is unmistakable.

    The Part I Find Most Interesting

    Beyond the geopolitics, there’s an argument — which Tevelow makes in his original piece — that hard money raises the cost of conflict. When you can’t print your way to war, war gets harder to sustain. The inflationary financing of military adventurism becomes less viable. That’s a long-horizon thesis, and I’d hold it loosely. But it’s not an unreasonable one.

    Historically, the ability to inflate currency has been the hidden subsidy for conflict. Governments rarely raise taxes to fund wars — they borrow and print, and the cost is deferred and diffused. Bitcoin, by design, removes that mechanism. Whether that actually changes behaviour at the nation-state level is an open question. But it’s an interesting structural constraint.

    The Bottom Line

    Iran demanding Bitcoin isn’t a Bitcoin story. It’s a financial infrastructure story. It’s a story about what happens when the tools used to enforce geopolitical compliance — sanctions, payment exclusions, asset freezes — start creating the demand for systems that are immune to them.

    That demand was always going to produce a supply. Bitcoin is the supply.

    The interesting question now isn’t whether this happens — it’s how quickly, and what the incumbent financial system does in response. I’d be surprised if the answer is “nothing”.

    Mark Hendy is an interim CFO working with PE-backed businesses. He writes about finance, AI, and the world at markhendy.com. Follow on LinkedIn.

  • AI Agent Memory: Build the System That Learns. Don’t Be the System Yourself.

    AI Agent Memory: Build the System That Learns. Don’t Be the System Yourself.

    A post went viral today — over 24,000 views in a few hours — claiming that AI agent memory “out of the box sucks” and that you need Obsidian to fix it. It resonated. But I think the conversation is missing something.

    The underlying problem is real. Most AI agents are amnesiac by default. Every session starts fresh. They don’t remember what you told them last week, what decisions you made, what context matters. You end up repeating yourself constantly — which defeats the point of having an assistant at all.

    The Obsidian solution people are sharing works like this: you maintain a structured vault of markdown notes, your agent reads from it at session start, and you manually curate what goes in. It’s better than nothing. But it has a fundamental problem — it still requires you to do the work.

    The Memory Problem, Properly Stated

    The goal isn’t just persistent storage. It’s useful persistent storage. There’s a difference between an agent that can retrieve a file you pointed it at, and one that has genuinely learned from your interactions — that knows what matters to you, what you’ve decided, what patterns recur in your work.

    Manual curation doesn’t scale. If you’re running an AI agent seriously — dozens of interactions a day — you cannot manually decide what gets committed to long-term memory. You’ll either capture too little (and lose signal) or spend as much time curating memory as you save everywhere else.

    What you actually need is a system that does this automatically, with enough intelligence to distinguish noise from signal.

    What I Built Instead

    I run OpenClaw with an AI assistant I’ve named Saul — a PE-facing CFO’s take on the AI agent problem, which I wrote about here. Over the past few months, I’ve built out a three-layer memory architecture that removes the manual curation problem entirely.

    The layers:

    • Daily notes — raw logs of what happened each session. Every interaction, decision, and piece of context gets written here automatically.
    • MEMORY.md — curated long-term memory. The distilled essence: decisions made, preferences established, important context. Think of it as the agent’s actual knowledge of you.
    • Dreaming — a nightly automated process (new in OpenClaw 2026.4.8) that reviews daily notes, scores entries by frequency, relevance, recency and query diversity, and promotes the strongest signals into MEMORY.md automatically. No manual curation.

    The third layer is the one that matters. Every night at 3am, the agent runs what OpenClaw calls a “dreaming” sweep — light phase sorts and stages recent material, REM phase extracts recurring themes, deep phase decides what gets promoted to long-term memory. The thresholds are configurable. The process is auditable. And it happens without me thinking about it.

    The Obsidian Angle

    The Obsidian approach people are excited about is essentially building layer two manually. It works, and if you’re starting from nothing it’s a reasonable place to start. OpenClaw’s memory-wiki plugin (also new in 2026.4.8) is actually Obsidian-compatible — same markdown format, same vault structure — so the two aren’t mutually exclusive.

    But if you’re going to invest time in your agent’s memory architecture, I’d argue the better investment is in automation rather than manual curation. Build the pipeline that decides what matters, rather than deciding manually every time.

    Why This Matters Beyond the Tech

    I’m a CFO. My primary concern with AI agents isn’t whether they’re impressive in a demo — it’s whether they actually reduce friction in the work I do every day. An agent with poor memory creates more friction, not less. You spend time re-explaining context, re-stating preferences, re-establishing where you are on a project.

    The ROI on getting memory right is substantial. An agent that genuinely knows you — your clients, your decisions, your communication style, your priorities — operates at a different level of usefulness. The gap between a well-configured agent and a default one isn’t incremental. It’s categorical.

    If you’re using an AI agent seriously and you haven’t thought about memory architecture, you’re leaving most of the value on the table. Whether you use the Obsidian approach, OpenClaw’s native dreaming, or something else — the manual-entry-only approach isn’t good enough long term.

    Build the system that learns. Don’t be the system yourself.

    Mark Hendy is an interim CFO working with PE-backed businesses. He writes about AI, finance, and the intersection of the two at markhendy.com. Follow on LinkedIn.

  • GLM-5.1: The AI That Works While You Sleep — And Then Some

    GLM-5.1: The AI That Works While You Sleep — And Then Some

    There’s a particular kind of AI announcement that makes me sit up. Not the ones that claim to beat GPT on some benchmark no-one’s heard of. Not the ones with slick demos that quietly ignore the bit where it falls over. The ones that matter are the ones where someone shows you the receipts — actual tasks, actual time, actual results.

    GLM-5.1, released last week by Z.ai (formerly Zhipu AI), is one of those.

    Eight Hours. Unattended.

    Let me frame what I mean. Most AI coding tools work in short bursts. You ask them to write a function, review some code, draft a test. Good assistants. But you’re still the loop-closer — the one who notices it’s gone sideways, resets the context, redirects the prompt.

    GLM-5.1 does something materially different. Z.ai ran it for eight hours straight building a Linux-style desktop environment from scratch — file browser, terminal, games. No handholding. It planned, executed, hit blockers, revised its approach, iterated. Hundreds of times. The claim isn’t “it wrote code”. The claim is “it didn’t give up”.

    That’s a different category of capability.

    The Numbers That Matter

    I’m a CFO. I like numbers. Here are the ones worth paying attention to:

    • Vector database optimisation: 600+ iterations, 6,000+ tool calls, 21,500 queries per second — six times the previous best
    • GPU kernel tuning: 1,000+ turns, 3.6× speedup on ML workloads
    • SWE-Bench Pro: 58.4% — ahead of both GPT-5.4 (57.7%) and Claude Opus 4.6 (57.3%)

    That last one is significant. This isn’t some niche Chinese model playing catch-up. On the hardest software engineering benchmarks, it’s beating the models most people consider the gold standard. And it’s open source — MIT licence, weights on Hugging Face, deployable on your own infrastructure.

    Why This Matters to Finance and Business

    I’ve written before about the shift from AI as assistant to AI as agent. GLM-5.1 is the clearest demonstration yet of what agentic AI actually looks like in practice.

    Think about the workflows in a finance function that are genuinely tedious:

    • Building and debugging complex financial models
    • Writing and testing data pipeline logic
    • Iterating on management information templates
    • Automating reconciliation scripts

    These aren’t tasks that fail at step one. They fail at step seven, when the edge case appears. They fail when the data format changes. They fail when the logic that worked last month doesn’t work this month. The human overhead isn’t writing the first version — it’s the iteration.

    If a model can sustain goal-directed effort over hundreds of iterations without losing the thread, that’s not incrementally better. That’s a different class of tool.

    The Open Source Angle

    Z.ai releasing this under MIT licence is genuinely interesting. The dominant models — OpenAI, Anthropic, Google — are all closed. You pay for API access, you accept their terms, you live with their rate limits and pricing changes.

    An open-source model that competes on performance changes the calculus for enterprise deployment. You can run it on-premise. You control the data. You don’t get a pricing change email in March telling you costs are going up 40% in April.

    For PE-backed businesses with sensitive financial data and legitimate concerns about feeding that data into third-party APIs — this matters.

    What I’d Watch

    GLM-5.1 isn’t perfect. It trails on some pure reasoning benchmarks. The long-horizon capability, while impressive, assumes the task is well-defined enough for autonomous execution — genuinely ambiguous strategic questions still need a human in the loop. And “it ran for 8 hours” cuts both ways: great if it’s right, expensive if it’s wrong.

    But the trajectory is clear. Each successive model generation extends the horizon over which AI can operate without human intervention. GLM-5 held position. GLM-5.1 sustained improvement. The question isn’t whether agentic AI is coming to professional services and finance — it’s whether you’re planning for it.

    If you want to know how I’m already using AI agents in day-to-day finance work, read this.

    The Practical Bit

    If you want to experiment: it’s on Hugging Face, runs via vLLM or SGLang, and integrates with standard agentic frameworks. For those less inclined to self-host, API access is through api.z.ai. The coding-focused subscription plan is $10/month — less than a decent lunch.

    I’ll be testing it against some of the financial automation tasks I currently route through Claude. I’ll report back.

    Mark Hendy is an interim CFO working with PE-backed businesses. He writes about AI, finance, and the intersection of the two at markhendy.com. Follow on LinkedIn.

  • GLM-5.1: The Chinese Open-Source Model That Just Beat GPT and Claude at Their Own Game

    GLM-5.1: The Chinese Open-Source Model That Just Beat GPT and Claude at Their Own Game

    Something significant happened in the AI landscape this week, and I suspect it hasn’t got the attention it deserves outside of developer circles. Z.AI — the platform behind the GLM model family, developed by Zhipu AI in China — released GLM-5.1, a 754 billion parameter open-source model that has just topped the SWE-Bench Pro leaderboard with a score of 58.4, beating GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro.

    Let that land for a moment. An open-source, MIT-licensed model, trained entirely on Huawei Ascend 910B chips — no Nvidia, no American silicon — has beaten the flagship closed models from OpenAI, Anthropic, and Google on one of the most respected software engineering benchmarks in existence.

    What Makes GLM-5.1 Different

    The headline number is impressive, but what actually interests me is the architecture of how this model works. GLM-5.1 isn’t just better at answering questions — it’s designed for sustained autonomous execution. In testing, it completed an eight-hour uninterrupted coding session: plan, execute, test, optimise, repeat. 655 iterations. Built a Linux desktop environment from scratch. Increased vector database query throughput by 6.9 times.

    This is a different category of capability. We’re not talking about a better chatbot. We’re talking about an AI that can hold a task in mind, work through it independently, hit dead ends, correct course, and deliver a finished result — the way a competent junior engineer would, but without stopping for the night.

    The technical foundation is a Mixture-of-Experts architecture with 40 billion active parameters per token (not all 754B are active at once, which is what keeps inference costs manageable). It supports a 200,000 token context window with up to 128,000 output tokens. API access is priced at $1.00 per million input tokens and $3.20 per million output tokens — a fraction of what the US frontier models charge.

    Why This Matters Beyond the Benchmarks

    I’ve written before about AI moving from a tool you prompt to a system that acts. GLM-5.1 is a concrete illustration of that shift happening faster than most people expected, and from a direction many in the West weren’t watching closely.

    The geopolitical dimension is real. This model was trained on Huawei hardware using Huawei’s MindSpore framework — a deliberate demonstration that China’s AI development pipeline is no longer dependent on US export-controlled chips. The export restrictions that were supposed to slow Chinese AI development have instead accelerated domestic alternatives. That is a significant strategic development, regardless of where you sit on the AI competition question.

    The open-source dimension is equally significant. With weights published under an MIT licence, GLM-5.1 can be downloaded, fine-tuned, and deployed by anyone. The closed-model advantage that OpenAI and Anthropic have built commercial moats around is being systematically eroded — not just by each other, but by well-resourced open-source releases like this one.

    What I Take From This

    I use AI heavily in my work — for financial analysis, document preparation, research, and increasingly for autonomous background tasks. The pace at which these systems are improving is not slowing down. If anything, GLM-5.1 suggests the competitive field is widening: more players, more approaches, more open options.

    For anyone running a business or advising one, the practical implication is straightforward: the cost of access to frontier-level AI capability is falling rapidly, and the choice of provider is expanding. The question is no longer whether to use these tools — it’s which ones, for what, and how to build processes around them that compound over time.

    GLM-5.1 is worth watching. Not because it’s the final word, but because it’s a clear signal that the race is genuinely global, the open-source movement is closing the gap faster than expected, and the next twelve months are going to be interesting.

    GLM-5.1 is available via z.ai on the GLM Coding Plan, with weights on Hugging Face under MIT licence.

  • Conway and the End of the Chat Window

    Last week, someone at Anthropic made a packaging error. A build of something called Conway — an internal project for an always-on AI agent — leaked out into the wild. Within hours, screenshots were circulating on Twitter showing an extensions system, webhook endpoints, and a Chrome integration that looked nothing like the Claude chat interface we all know.

    Anthropic played it down. A release packaging issue caused by human error, they said. Not a security breach. Sure. But the cat is out of the bag, and what it reveals is far more interesting than the leak itself.

    From Prompt-Response to Always-On

    Every AI product you use today works the same way: you type something, it responds, you go back and forth until you get what you need. It is fundamentally a conversation. Conway is something different. It is an agent environment that stays running. It has webhook endpoints — public URLs that external services can call to wake the agent up when something happens. It has an extensions system where you can install custom tools, UI tabs, and context handlers. It uses Chrome autonomously to handle multi-step tasks on the web.

    This is not a chatbot that got fancier. This is the architectural blueprint for AI that operates like a team member — one that never logs off, never forgets its context, and responds to events in the real world without waiting for you to type a prompt.

    I find this significant because I have been building exactly this kind of system for myself over the past few months. My own AI assistant, Saul, already runs scheduled tasks, publishes content, monitors data, and sends me summaries — all without me being in the loop for every step. It works. But it is held together with cron jobs, Python scripts, and API calls that I have wired up manually. Conway suggests Anthropic wants to make this kind of continuous AI operation a first-class product, not a weekend project for the technically inclined.

    The Enterprise Is Already There (Sort Of)

    A Belitsoft report published this week, drawing on Salesforce’s 2026 Connectivity Benchmark data, says the average enterprise now runs 12 AI agents. Twelve. Expected to hit 20 by 2027. But here is the kicker — half of those agents operate in complete isolation. They do not talk to each other. They do not share context. They are twelve separate hammers looking for nails.

    This is exactly the problem Conway appears to be solving. The extensions architecture, the webhook system, the persistent state — it is all about creating a single agent environment that can integrate with everything, rather than deploying a dozen disconnected point solutions. The shift is from “we have AI tools” to “we have an AI operating system.”

    What This Means If You Run a Finance Function

    I spend most of my professional life inside PE-backed finance teams, and the implications here are not abstract. Think about what a CFO’s week actually looks like: cash flow monitoring, covenant compliance, board pack preparation, variance analysis, vendor negotiations, investor reporting. Every single one of those workflows involves gathering data from multiple systems, applying judgment, producing an output, and sending it somewhere.

    An always-on agent does not replace the judgment part — not yet, anyway — but it can collapse the gathering, formatting, and distribution steps into something that just happens. The board pack data is already pulled and formatted before you sit down on Monday morning. The covenant calculations are running continuously, not quarterly. The cash position is reconciled and summarised before the daily stand-up.

    This is not speculative. I have automated parts of this already. But Conway-style infrastructure would make it dramatically easier to set up and maintain, which means it stops being something only a CFO who can write Python does and starts being something any competent finance team can deploy.

    The Uncomfortable Bit

    There is a tension here that I think most people in the AI space are not being honest about. If you make agents always-on, event-driven, and capable of taking action autonomously, you have fundamentally changed the trust model. A chatbot that gives bad advice is annoying. An always-on agent that takes bad action at 3am is a different category of problem entirely.

    Conway’s architecture seems to acknowledge this — the extensions and webhook systems suggest granular control over what the agent can and cannot do. But the history of enterprise software tells us that permissions and guardrails are only as good as the people configuring them. And in most mid-market PE-backed businesses, the people configuring them will be a mix of finance staff, IT generalists, and maybe one overstretched CTO. The governance question is not solved by better architecture alone. It requires new operational disciplines that most organisations have not even started thinking about.

    Where This Is Going

    Here is what I think happens next. Anthropic ships Conway — or something very like it — within the next quarter. Google already has its own agent infrastructure play with Gemma 4 and Vertex agents. OpenAI is pushing GPT-5.4 with desktop task automation that scored 75% on real-world benchmarks. Microsoft just shipped Agent Framework 1.0. The always-on agent is not a research project anymore — it is a product category that every major AI company is racing to own.

    For CFOs and finance leaders, the practical question is not whether to adopt this technology, but how to build the internal capability to govern it. That means understanding what your agents are doing, why, and with what authority. It means having someone on the team — or on retainer — who can configure, monitor, and audit these systems. And it means accepting that the competitive advantage will not go to the company that deploys the most agents, but to the one that gets them working together coherently.

    The Belitsoft numbers tell us enterprises are already halfway there on deployment. Conway tells us the infrastructure for always-on operation is coming. The missing piece is the operational maturity to make it work safely and effectively. That, more than any model benchmark, is where the real work is.

  • You Don’t Deploy AI Agents Anymore — You Hire Them

    Yesterday, monday.com launched something called Agentalent.ai — a managed marketplace where enterprises can discover, evaluate, and “hire” AI agents for defined business roles. Not install. Not deploy. Hire.

    You post a role. You review qualified agents. You select based on task fit, budget, and operational readiness. Pricing starts around $2,000 a month per agent. They launch with 17 agents available. Built in collaboration with AWS, Anthropic, and Wix.

    If you’re a CFO and that doesn’t make your headcount model twitch, you’re not paying attention.

    The Language Shift That Matters

    I’ve been building with AI agents for the best part of two years now — wiring up Claude to handle research tasks, automating financial reporting pipelines, getting agents to do the kind of grunt work that used to eat a junior analyst’s entire Tuesday. But the framing has always been tooling. You set up an agent like you’d set up a spreadsheet macro. It’s a thing on your computer.

    What monday.com has done — deliberately, with their HR-style language — is shift the frame from tools to workers. And that’s not just marketing fluff. It’s the conceptual bridge that will get the rest of the C-suite to finally understand what’s happening.

    A Belitsoft report published this weekend puts numbers on it: the average enterprise now runs 12 AI agents. Twelve. And that’s expected to hit 20 by 2027. But here’s the kicker — half of those agents operate completely alone, unconnected to any other agent or system. They’re doing their little jobs in their little silos, and nobody’s orchestrating the whole thing.

    Sound familiar? It should. That’s exactly what happens when a company hires people without a coherent operating model. You end up with twelve contractors, half of whom don’t talk to each other, doing overlapping work with no shared context. I’ve walked into PE portfolio companies that look exactly like this — except with humans.

    The CFO’s New Headcount Problem

    Here’s where it gets interesting for anyone sitting in a finance seat. When an AI agent costs $2,000 a month and can do the work of a task that previously required a $6,000/month contractor, that’s a straightforward business case. Any CFO can model that. The ROI practically draws itself.

    But the real question isn’t “should we hire the agent?” It’s “how do we account for a workforce that’s now 30% software?”

    Think about what sits in your headcount model today. Salaries, employer NI, pension contributions, benefits, training costs, recruitment fees. Now think about what sits in your AI agent budget. SaaS subscriptions, API usage fees, compute costs, maybe some integration consulting. These two things live in completely different cost categories, get approved through different processes, and are managed by different people. But they’re increasingly doing the same work.

    In the PE world I operate in, headcount is one of the first things a new investor scrutinises. “What’s your revenue per head?” “What’s your fully-loaded cost per FTE?” These metrics are foundational to how value creation plans get built. But nobody’s asking “what’s your revenue per agent?” yet. And they should be, because if you’re running 12 agents and growing, that’s a material line in your operating model that isn’t being tracked like one.

    The Coordination Tax

    The Belitsoft finding that half of deployed agents work alone is, I think, the most important data point in their entire report. It mirrors what I’ve seen first-hand. Companies get excited, they spin up agents for customer support, for code review, for data entry, for reporting — and each one works reasonably well in isolation. But the value compounds when agents talk to each other, and almost nobody has figured that part out yet.

    This is an orchestration problem, and it’s fundamentally a management problem. You need someone — or something — deciding which agent handles which task, what context gets shared, where the human review gates sit. NVIDIA’s new Agent Toolkit, announced with partners including Salesforce, SAP, and ServiceNow, is trying to solve the infrastructure side of this. Okta’s new “secure agentic enterprise” framework, going GA at the end of this month, is tackling identity and access. But the management layer — the actual decision-making about how to deploy and coordinate these things — that’s still a gap.

    And it’s a gap that, in most companies, probably falls to the CFO. Not the CTO. Not the CISO. The CFO. Because ultimately this is a resource allocation problem. You have a pool of human and non-human workers. You have tasks that need doing. You need to figure out the optimal mix, track the cost, measure the output, and report on it to a board that still thinks in FTEs.

    What I’m Actually Doing About It

    In my own setup, I’ve started treating agent costs the way I treat contractor costs — as a blended workforce line, not a software line. My AI assistant Saul runs daily tasks for me: research, publishing, monitoring. I track what he does, what it costs, and what it would cost if a human did it instead. Not because I’m obsessive about it (okay, partly because I’m obsessive about it), but because I think this is the accounting framework that PE firms will expect within 18 months.

    The $600 billion flowing into AI agent ecosystems in 2026 isn’t going into chatbots. It’s going into digital workers — things that take tasks, complete them, and cost money every month. If your chart of accounts still treats all of that as “IT software subscriptions,” you’re going to have a very confusing board pack by Christmas.

    Where This Goes

    monday.com’s marketplace is clunky right now — 17 agents isn’t exactly a deep talent pool. But the model is right. Within a year, I’d expect to see the big consulting firms offering “blended workforce planning” as a service line. Within two, PE due diligence will include an AI agent audit alongside the usual people and tech reviews.

    For CFOs, the action item is boringly practical: start tracking your agents like you track your people. Give them cost centres. Measure their output. Build the reporting now, while it’s still simple, because it won’t be simple for long.

    We spent decades building HR systems to manage human workers. We’re about to need something equivalent for the digital ones. And the CFO who figures that out first is going to look very clever at the next board meeting.

  • The Panic is the Point: Bitcoin’s Worst Q1 Since 2018 and What the Smart Money Is Actually Doing

    The Panic is the Point: Bitcoin’s Worst Q1 Since 2018 and What the Smart Money Is Actually Doing

    The Fear and Greed Index hit 8 last week. Eight. Out of a hundred. I’ve been watching that index for years and I’ve seen it touch double digits maybe a handful of times. Every one of them felt like the end of the world. Every one of them, in hindsight, looked like a buying opportunity.

    Bitcoin just closed its worst first quarter since 2018, down 23.8% from $87,500 at the start of January to around $66,600 by quarter end. The narrative that greeted April was bleak: ETF outflows, macro headwinds, geopolitical noise, retail capitulation. The index has been below 15 for 47 consecutive days — the longest such streak since the Terra-Luna collapse in 2022. Social media sentiment is, apparently, at its most negative since late February. Everyone is miserable.

    And yet something interesting is happening underneath the surface. Something that I think most of the commentary is missing.

    The Divergence Nobody Is Talking About

    Here is the part that caught my attention. While the Fear and Greed Index was screaming extreme panic, spot Bitcoin ETFs snapped a four-month outflow streak in March, pulling in $1.32 billion in a single month. Corporate Bitcoin treasuries hit record levels in early 2026, with public companies collectively holding over 1.1 million BTC — somewhere north of 5% of total supply. And the largest asset managers have not moved their macro targets: $150,000 to $200,000 by year end is still the institutional consensus.

    So you have a situation where retail sentiment is at historically depressed levels, and institutions are quietly filling their bags. That divergence is not new — it happens in every asset class, every cycle. But in Bitcoin it tends to be particularly pronounced because the retail holder base is so emotionally reactive, and because the on-chain data makes the institutional accumulation visible in a way that equity markets don’t.

    I am not making a price prediction here. I’ve been around long enough to know that timing markets is mostly a story you tell yourself after the fact. But I do think there is something analytically interesting in the gap between what the sentiment data says and what the flow data says. When those two things diverge this sharply, it is usually worth paying attention.

    Fear and Greed as a Contrarian Instrument

    The Crypto Fear and Greed Index is a blunt instrument. It aggregates volatility, momentum, social media volume, surveys, dominance, and trends into a single number. It is not sophisticated. But its very simplicity is what makes it useful as a contrarian signal — it tells you how the crowd is feeling, and the crowd is famously wrong at extremes.

    The historical data on sub-10 readings is striking. According to analysis of prior cycles, readings below 10 have occurred on fewer than 20 trading days since the index’s inception, clustered around the March 2020 COVID crash, the May 2021 China mining ban, and the June 2022 Terra-Luna contagion. The median 90-day return from sub-15 readings has historically been around +38%. Sub-10 readings have averaged +43% over the following 90 days. The caveat — and it is an important one — is that during the post-Terra contagion in 2022, the subsequent 90 days produced only a modest +4% as cascading liquidations kept a lid on recovery. Context matters.

    The current context feels more like 2020 than 2022 to me. The fear is driven by macro uncertainty and sentiment exhaustion, not by a structural collapse in the ecosystem. There is no Three Arrows Capital moment lurking. The ETF infrastructure is intact. Corporate treasury demand is structural, not speculative.

    What the Institutional Behaviour Actually Tells Us

    I spent some time this weekend reading through the Q1 flow data. The picture is messy but directionally clear. January and February saw $1.8 billion in ETF outflows as the price fell from $87K and macro risk-off sentiment hit. Then March happened: $1.32 billion back in, suggesting institutional re-entry at levels they consider attractive. Meanwhile, CoinDesk noted that Bitcoin is entering April at its most hated sentiment level since the Ukraine war began — a data point that is simultaneously depressing and, for a contrarian, quietly exciting.

    There’s a Morgan Stanley Bitcoin ETF that was recently approved with a notably low fee structure — another piece of institutional infrastructure quietly being laid while retail stares at the Fear and Greed number and panics. Infrastructure gets built in bear markets. That’s always been true.

    I hold Bitcoin. I have held it through worse than this. My view hasn’t changed: the long-term thesis — fixed supply, increasing institutional legitimacy, ETF-driven structural demand — is intact. A 24% Q1 drawdown is uncomfortable but it is not abnormal for an asset that is still, by any traditional measure, in an early adoption phase.

    The Noise vs. The Signal

    The thing that strikes me about the current moment is how clean the signal actually is, once you cut through the noise. Retail fear at historic extremes. Institutional accumulation quietly continuing. Corporate treasuries at record levels. ETF infrastructure expanding. The narrative is all doom, but the flows tell a different story.

    I am not saying it cannot go lower. Some analysts think there’s room for another leg down if macro conditions deteriorate further. Maybe. But I’ve found that the best time to think clearly about Bitcoin is when everyone else has stopped thinking clearly about it — and right now, a Fear and Greed reading of 8 suggests that the crowd has well and truly checked out.

    The panic, as far as I can tell, is the point. It is the mechanism by which assets transfer from weak hands to strong ones. It is not comfortable to watch in real time. But the data, as best as I can read it, suggests the strong hands are doing exactly what they always do: accumulating quietly while the timeline argues about whether it’s over.

    It’s probably not over.

  • IKEA’s Chatbot Accidentally Made €1.3 Billion. Here’s What CFOs Are Missing.

    IKEA’s Chatbot Accidentally Made €1.3 Billion. Here’s What CFOs Are Missing.

    Most companies deploy AI to cut costs. IKEA deployed AI to cut costs and accidentally discovered a billion-euro revenue stream hiding in the data their chatbot was collecting.

    This is the story every CFO in every PE portfolio company should be reading right now. Not because of IKEA. Because of what it reveals about how most finance leaders think about AI — and how badly they’re getting it wrong.

    The Setup

    IKEA — or more precisely, Ingka Group, the largest IKEA retailer — built an AI chatbot called Billie. The brief was simple: handle level-one customer service enquiries. Reduce call volumes. Cut costs. The standard playbook.

    Billie did its job. From 2021 to 2023, it resolved roughly 47% of customer enquiries it received — around 3.2 million interactions handled without a human, saving an estimated €13 million.

    If you’re a CFO, that’s a clean win. Cost out, efficiency up, ROI positive. You’d put it in the board pack and move on.

    IKEA didn’t move on.

    The Signal Nobody Was Looking For

    The interesting number wasn’t the 47% that Billie resolved. It was the other 53%.

    When IKEA’s team analysed the enquiries Billie couldn’t handle, they found something unexpected. A huge proportion weren’t complaints or order issues. They were customers asking for help designing their homes.

    People were calling IKEA — a furniture shop — and saying: I’ve got this room. What should I do with it?

    This wasn’t in anyone’s business case. No strategy deck had “launch a design consultancy” on the roadmap. It was a signal buried in the noise of customer service data, and it would have stayed buried if someone hadn’t been paying attention.

    The Pivot

    Here’s where it gets good.

    Instead of just improving Billie’s resolution rate — the obvious move, the one every consulting firm would have recommended — IKEA did something much smarter. They took 8,500 call centre workers and reskilled them as remote interior design consultants.

    Read that again. Eight and a half thousand people. Not made redundant. Reskilled.

    The AI handled the routine queries. The humans handled the high-value, creative, relationship-driven work that customers were already asking for. IKEA didn’t replace their workforce with AI. They promoted their workforce because of AI.

    The result? Remote interior design sales hit €1.3 billion by the end of their 2022 financial year — 3.3% of Ingka Group’s total revenue. A brand new service line, created from a signal that existed in their customer service data all along. Their target is 10% of total sales in the coming years.

    Why CFOs Get This Wrong

    I’ve sat in enough board meetings to know how this story usually goes.

    A CFO sees the AI chatbot business case. It says: deploy chatbot, save €13 million in customer service costs, payback in 18 months. They approve it. They monitor the cost savings. They report the efficiency gains. Job done.

    That’s not wrong. But it’s incomplete.

    The €13 million in cost savings is a rounding error compared to the €1.3 billion in new revenue. The chatbot wasn’t the product. The chatbot was a listening device.

    Most AI business cases are framed as cost reduction exercises. Automate this process. Eliminate these headcount. Reduce that cycle time. And they work — the savings are real. But they’re also the least interesting thing AI can do.

    The interesting thing is what AI reveals about your customers when you stop looking at it as a cost tool and start looking at it as an intelligence tool.

    The PE Angle

    If you’re a PE operating partner reading this, think about your portfolio.

    Every portfolio company has customer service data. Most of it sits in a ticketing system that nobody reads except the support team. Some of it gets summarised in a monthly report that the board glances at between the P&L and the cash flow forecast.

    What if that data contains the same signal IKEA found? What if there’s a billion-euro service line hiding in your Zendesk tickets?

    The companies that will win the next decade aren’t the ones that use AI to do the same things cheaper. They’re the ones that use AI to discover things they didn’t know their customers wanted. That’s a fundamentally different value proposition — and it requires a fundamentally different kind of CFO.

    The Kind of CFO That Catches This

    The old-school CFO sees AI as a line item. A cost to manage, an efficiency to capture, an ROI to calculate.

    The new-school CFO sees AI as an intelligence layer. Every automated interaction is a data point. Every pattern in that data is a potential business model. Every customer service complaint is a market signal.

    IKEA didn’t need a McKinsey engagement to discover the design consultancy opportunity. They needed someone who looked at the chatbot’s failure cases and asked: why are these people calling us?

    That’s not a technology question. It’s a business question. And it’s the kind of question that CFOs — with their bird’s-eye view of costs, revenues, and customer patterns — are uniquely positioned to ask.

    The Uncomfortable Truth

    Here’s what makes this story uncomfortable for a lot of finance professionals.

    The €13 million saving was predictable. You could model it in advance, put it in a business case, and track it against plan. That’s the kind of AI outcome that finance teams are comfortable with.

    The €1.3 billion revenue stream was unpredictable. It emerged from the data. Nobody forecast it. Nobody budgeted for it. It required curiosity, not spreadsheets.

    If your AI strategy only captures the predictable value, you’re leaving the transformative value on the table. And in a competitive market, someone else will find it first.

    What To Do About It

    Three things, starting tomorrow:

    1. Audit your AI for signals, not just savings. Every AI tool in your business is generating data about customer behaviour. Who’s reading it? What patterns are emerging? If the answer is “nobody” and “we don’t know,” you have a blind spot the size of IKEA’s design consultancy.

    2. Look at the failures, not just the successes. IKEA’s breakthrough came from what Billie couldn’t do. The 53% failure rate wasn’t a problem to fix — it was a market to serve. What are your AI tools failing at? Those failures might be your biggest opportunities.

    3. Stop framing AI as a cost play. If every AI business case in your portfolio starts with “reduce headcount” or “automate process,” you’re optimising for efficiency while your competitors are optimising for discovery. The cost savings are table stakes. The revenue signals are the game.

    Mark Hendy is a PE-facing CFO and founder of Tanous Limited. He writes about the intersection of AI, finance, and business transformation at [markhendy.com](https://markhendy.com).