Zoolatech

7. At @Zoolatech_com, we help companies get the fundamentals right, then build for them. In a five-year partnership with a major North American fashion retailer, our team scaled the app to 10M+ downloads across iOS and Android and doubled development velocity. On Android, the team doubled app revenue in three years and grew engagement time by over 40%. If you're planning a mobile build, reach out. zoolatech.com

6. Pick a partner with a track record. 42.9% of companies rank proven experience as the top factor when choosing a development partner, well ahead of cost. Building at scale is full of decisions that only experience prepares you for.

We took part in @TechBehemoths' 2026 global survey on mobile app development, alongside IT companies from 96 countries. The teams seeing strong returns from mobile mostly do the same unglamorous thing: they get clear on what success looks like and how the app should be built before development starts. Nearly 43% don't, and later they can't really say whether the app worked. Discover the groundwork that puts you in the 11.6%. #Zoolatech #MobileDevelopment #AppDevelopment #ProductStrategy

@DavidLinthicum The hard part is picking the right bet. “Focused over blanket” makes sense, but to know where to focus you need a real read on how your processes run today. Most companies don’t have that. So the focused bet is often just a cheaper guess than the mandate.

PwC Got This Right on AI A lot of my tech clients complain that enterprises are moving too slowly on AI. But at the same time, many of them are pushing an AI-first, enterprise-wide mandate as the answer. That’s exactly where things start to go wrong. PwC makes the point clearly: AI value is uneven. It does not create the same impact in every function, process, or team. When companies try to force AI into everything at once, they dilute investment, spread talent too thin, and create activity that looks impressive but often delivers very little business value. AI should be deployed tactically, in the places where it can do the most good. That means focusing on the areas where it can improve margins, speed decisions, reduce risk, strengthen customer outcomes, or unlock new revenue. Not every workflow needs AI, and not every problem is improved by adding a model to it. That’s why PwC is right to argue for focused bets over blanket mandates. The companies that will win with AI are not the ones launching the most pilots or making the most noise. They’ll be the ones that place concentrated bets in a few high-value areas, while also building the governance, tools, and discipline needed to scale what actually works. The current push to make everything AI-enabled is not strategy. In many cases, it’s just expensive distraction. Done badly, it will create negative business value. Article: pwc.com/us/en/services…

The orchestration/monitoring split is right. Memory is where it gets messy though. When it's just you, memory is one list Toby reads start to finish. In a real company, the same fact lives in three places that don't agree. One system says the deal closed, another still shows it open, a third won't update until overnight. So the watchdog flags drift, and it's not even wrong. The three places it's reading already tell different stories about what happened.

My most valuable AI agent does absolutely no work, and I mean that as the highest compliment. For a while it was just me and my Chief of Staff, Simon. Simon runs six direct reports (Monica, Chandler, Phoebe...), each with their own sub-agents. I'd hand him a goal or task, he'd orchestrate it, work came back. Good, but things slipped, or we'd drift from a goal, or the same problem would resurface three times. No one was watching the work going "is this good?" or "what keeps breaking?" or "what did we learn?" and updating the system. So I gave Simon an assistant named Toby. All Toby does is watch and fix. I love Toby. (Not that way.) He flags when we drift from our goals, catches recurring issues, and manages our memory. The whole system gets smarter instead of just busier. More effective instead of just more powerful. Most people building agents seem to stop at the doer (task completer) or the judge (task verifier). ✔️ AI doers might draft your invoice reminder emails, run competitive research for you, narrow down a big vendor group for you based on criteria, run a forecast on your Brazil market ✔️ AI judges might review your code before you push or check your PPT draft against brand guidelines. Useful, but it only looks when you ask it to or gets triggered after task completion. A watchdog looks all the time, and tells you what you didn't think to check. It's a safety net, a second set of eyes, a bar raiser. All of this came through iteration (and I walked my AI Agent Mastermind students through all four iterations of my AI agent workforce and what I learned), but a massive unlock was splitting the "project management" work in two - orchestration and deep reasoning with one agent + monitoring and memory with another. I keep saying this but the boring bits of AI are some of the most important bits. I'm sure the AI labs will add this functionality into their systems, but for now, Toby is a huge leg up for my AI workforce.

One thing we'd add from the delivery side: under the operating model sits the data itself. A company can reorganize teams around shared AI skills and still stall, because those skills only compound when they can reach clean, connected company data. And in most enterprises that data lives across systems that were never built to share it. So the reorg opens the door, but the data-access work is usually what determines whether anyone walks through it.

That last line is the whole thing. Identity on the rails proves who transacted, not whether the agent should have bought it at all. You can encode the obvious limits. The trouble is always the case nobody wrote a rule for, and attribution doesn't help there. Visa closed the identity gap. The judgment one stays open.

.@Visa just gave AI agents their own card credentials. @OpenAI is partnering. Agentic commerce has promised a lot, but delivered very little. Is this finally agentic commerce's mainstream moment? Announced yesterday at the Visa Payments Forum. The headline reads like "AI can shop for you now." The mechanic underneath is the interesting bit. Start with tokenization. It's the technology that takes your 16 digit card number and swaps it for a secure cryptographic token, so your real number never touches the merchant. That's what made Apple Pay work. Visa Intelligent Commerce extends those tokens to agents. Once an agent holds a token bound to it specifically, the agent has an identity on the Visa network. The same rails that already clear 300 billion transactions a year. So when your agent goes to pay: - Visa knows which agent is transacting - The merchant knows it's a recognized, trusted agent operating under your rules - Your bank can authorize in real time, against spending limits and approvals you set Say your agent buys something and you tell your bank you never authorized it. With agent identity on the network, there's a record that it really was your agent, acting inside the guardrails you gave it. That changes agents from being a hostile "bot" to a trusted customer. It also rhymes with every agent problem I keep running into. Autonomy only works when the system can prove what the agent did.

This matches what we see on the modernization side. The "no legacy" edge is real but it ages out. Today's fast fintech is writing the systems it'll have to modernize itself in a few years, and the incumbents bolting AI onto old cores were clean-stack startups once. Everyone ends up in the same cycle, just at different points in it.

🚨 JUST IN: Fintech just had its biggest year on record – $504 billion in revenue. New Global Fintech Report from BCG + FT Partners: the sector is now growing ~4x faster than traditional incumbents. The headline numbers: → $504B revenue, +22% YoY → Trading & investments +38%, deposits +30% → Equity funding $58B in 2025 (+53%) → 42 fintech IPOs last year Fintech is now 4% of all global financial-services revenue. What's driving it looks different this time. The report ties much of the growth to AI changing the economics: fintechs deeply integrating AI are seeing ~5x developer productivity. And because most fintechs aren't tied to legacy core-banking systems, they can adopt those tools far faster than incumbents bolting AI onto decades-old infrastructure. That's the story behind the number. The last cycle was growth at any cost. This one is being driven by speed and efficiency – leaner stacks, faster shipping, better margins. The open question is whether the pace holds as the sector scales and regulators look closer. Record revenue and record scrutiny tend to arrive together.

The duplicates and missing invoices are the part nobody warns you about going in. People picture migration as moving data from A to B, when most of the work is sitting with a year of messy transactions and deciding what's real. Getting the bank balance to reconcile on the far side is the hard part, and you earned that one. Nice work.

Few months back , my manger asked me if I can do migration from one ERP to another, I said yes but sincerely I have never done it before. I love challenging tasks so I told her I will do the migration, i thought it was something that will be so simple but no, it wasn’t. At some points I paused the migration and didn’t touch the client account. Last week, I was told to complete the migration and it became more complex as there were duplicate transactions,invoices and missing invoices. I thought asking AI to put me through the process was going to be easy but Claude AI sometimes carry me go where I no know 😂😂😂. Then I had to sit down and mapped out how to go about the whole process and at the end I successful migrated the client to Xero. Bank balance in Xero matches the actual bank balance Payables and receivable matches Vat returned prepared and submitted to HMRC💪 Moral lesson: don’t run away from hard tasks it, most times it helps and improve your technical skills

The irony is that those 11 weeks are where the real risk sits. Security review and data mapping take that long because that’s the stage where things break in production. Compress the 3 weeks of coding and all you’ve done is reach the 11-week wall sooner, with more code piled up behind it.

Exactly. I spent years at Accenture watching 6 month ERP migrations where maybe 3 weeks was actual development. the rest was change advisory boards, data mapping workshops, waiting for a security review that took 11 weeks. AI coding tools compress the 3 weeks. nobody's compressing the 11.

Things 25 years of enterprise tech taught me about AI coding tools: - open source is always a go-to-market strategy - the bottleneck is never the code - the vendor accelerating fastest always calls for a pause - every standalone tool gets absorbed into the platform - security gets filed under "later" until the first breach - flat-rate pricing is a temporary gift. usage-based is the real plan

@a16z Once every call is recorded, you’ve got a thousand hours of conversation where maybe three sentences changed a decision. The recording was always the easy part. Finding those three without someone who sat in the room is the problem nobody’s solved yet.

Most work conversations are now being recorded by default. You should probably assume that everything you say at work is getting recorded from here on out. What’s emerging is a new category of enterprise software, organized around voice instead of text. The system of record today is structured data: CRM entries, tickets, docs. But the highest-value context lives in conversation: the nuance on a customer call, the real argument in a product review, the offhand comment in a leadership meeting that quietly changes the roadmap. LLMs are uniquely good at taking that unstructured voice data and making it structured, searchable, and queryable. That’s a large enterprise opportunity, and we’re still early in understanding what the software layer looks like and who owns it. a16z GP David Haber on what AI recording means for the future of work: a16z.news/p/everything-i…

@alliekmiller Pulling the judgment out is the part that breaks teams. Data and processes you can extract. Judgment lives in the senior person who sees an exception and knows it’s wrong but can’t say why. Watch them a month and you’d still miss half of it.

Years ago someone asked me what comes after the AI-first enterprise. I said the autonomous enterprise. The core of it is a flywheel: four parts (well, five including feedback) that feed each other, and every loop makes the next one smarter: → Goals - what is the company prioritizing, what does each team own, what does good look like → Context - everything that's happened before, plus current day goings-on, plus your resources (budget, tokens, people) → Action - where work actually gets done, SOPs and skills so AI can execute against Gmail, Calendar, Salesforce, other data, tradeoff analysis against goals and context, simulations → Decisions - what actions were taken, the memory layer, where the system reflects on what worked and feeds it back in → Feedback (duh) You'll hear it called closed loops. Satya Nadella called it hill climbing to a small room a few weeks. But I worked at Amazon for four years, so it's going to be a flywheel in my head until I die. The hard part of all of this is the plumbing (ex: connecting AI into your tools safely, pulling the judgment out of your company, extracting and reviewing your processes, codifying your qualities and value, governance and documentation to be able to interpret and review what is happening) - and managing it so the system acts the way you actually want. Then monitoring the whole thing constantly, with humans and with AI. And trust me when I say that single exec workshop I'm doing right now is on building a self-learning org and carving out this flywheel.

That 40% is a snapshot, not a constant. It holds while the GPUs stay busy with work that matters. Six months in, the idle batch jobs creep back, someone leaves a debug run going over the weekend, half the calls retry without anyone noticing. The hardware cost stays fixed. The return slides unless someone owns keeping utilization tight.

I talked to a founder of an AI startup generating about a 40% annual return on the cost of the GPUs he was using. I.e. he could make $400 in annual revenue for every $1000 worth of hardware he used.

Substitution looks free on a pricing page and rarely is. Swapping a frontier model for an open one that's "good enough" usually means rewriting prompts and re-tuning guardrails, because the cheaper model fails in different places than the one it replaced. The token savings are easy to see. What's harder to count is the work of re-verifying everything the old model already got right.

Three forces are reshaping the AI cost structure : 1. Foundation labs are moving up the stack into applications 2. Frontier model prices keep rising for the smartest models 3. Open-source models have crossed the good enough threshold for most use cases The natural response from AI buyers is substitution. 🧵