# Ivars Bergs > Personal site of Ivars Bergs, an independent developer who builds practical AI integrations and automation for businesses on a fixed-price, remote, B2B basis. Includes a blog about AI engineering, browser automation and shipping real software. ## How an Accounting Firm Can Make Room for More Clients Date: 2026-09-07 Category: Workflow Automation Patterns Reading time: 3 min URL: https://ivarberg.com/blog/accounting-firm-capacity/ Before hiring another accountant, measure how much time your team spends finding attachments and retyping invoices. Reducing that work could free up hours for more clients. ## Where the time goes Take one completed month of incoming invoices. Count the volume and measure the work at each step. | Step | What to check in your firm | |---|---| | Receiving documents | Does each client have one clear submission channel? | | Checking completeness | How do you spot missing attachments or unreadable invoices? | | Data entry | Which data do you retype from another system? | | Assigning and matching | How do you select the company, supplier and cost allocation? | | Approval | What does the accountant check, and where do they flag questions? | If the biggest delay is waiting for a client's answer, start with a missing-document list and a responsible person on the client's side. Faster invoice extraction will not resolve that wait. ## How many hours could you free up? **Illustrative calculation:** 1,200 invoices a month × 4 minutes for entry and checking = 80 hours. If a pilot brings the average down to 1.5 minutes, including corrections, the same volume takes 30 hours. That frees up **50 hours a month**. Check who gets those hours back. If entry becomes faster but every review still waits for one senior accountant, their availability continues to limit capacity. To estimate how many more clients you can serve, establish how many hours they need and from which team members. ## What I have built in practice For a client with four companies, I built an invoice workflow starting in a Telegram group. The system extracts document data, identifies the company and prepares the information for human review. It detects an identical file submitted again and checks whether a different file contains the same invoice. Clear duplicates are attached to the existing invoice; suspiciously similar documents and missing data are flagged for review. Payment and bookkeeping have separate statuses, making it possible to find paid invoices whose bookkeeping has not been marked complete. The bookkeeping record keeps the date, responsible user and Jumis registration number. **Automatic posting into Jumis has not been implemented in this system yet.** Document preparation and status tracking already work; a person evaluates and approves the transaction. ## How to test the benefit in your firm First, check what your accounting software already supports: data imports, supplier matching and duplicate checks. That will show which connections or steps still need to be built. Choose one client and one document type for the pilot. Include the same invoice in two files, a credit note, missing data and a paid invoice still waiting for bookkeeping. Compare average processing time, including review and corrections, and the proportion of documents needing extra attention. Check that every received invoice has a traceable status. Compare the value of the time freed up with implementation, licence and maintenance costs. [Send me a note](mailto:ivarberg@lightsourcegames.com) with how you receive invoices, where you enter them and which step takes the most time. We can start there. --- ## Latvia's 2028 E-Invoice Mandate: What Your Company Must Do, and When Date: 2026-07-22 Category: Industry Deep Dives Reading time: 5 min URL: https://ivarberg.com/blog/e-invoices-latvia-2028/ # Latvia's 2028 E-Invoice Mandate: What Your Company Must Do, and When "We'll look at the e-invoice thing later" is the default position of Latvian companies right now. The deadline sounds far away, but for most companies the work is not a two-week job. TLDR: **from 1 January 2028, every invoice between Latvian companies must be a structured XML file, and its data must be reported to the State Revenue Service (VID).** The law is passed, and the only choice left is to prepare calmly or to do everything in December 2027 in a panic. ## The deadlines These dates come from the Accounting Law amendments the Saeima passed on 5 June 2025, which also moved the mandatory B2B phase from 2026 to 2028: - **Since 1 January 2025**, invoices to state and municipal institutions (B2G) must be structured e-invoices. - **From 1 January 2026**, B2G and G2G invoice data must be reported to VID. - **From 30 March 2026**, e-invoices can be used voluntarily in B2B transactions, including submitting the data to VID. - **From 1 January 2028**, structured e-invoices and VID data reporting are mandatory for all B2B transactions between companies registered in Latvia. The format is the same in every phase: XML following the Peppol BIS Billing 3.0 standard, based on the European e-invoicing standard EN 16931. ## What is and is not an e-invoice The PDF you export from your invoicing tool and send by email is **not** an e-invoice, and neither is a scanned paper invoice. An e-invoice is a data file: seller, buyer, line items, amounts, VAT, all in a fixed structure that the receiver's system books without retyping. I build systems that do exactly this retyping instead of people: they read a PDF or a photo, extract the data, and enter it into the accounting system. What the state now demands is what sensible companies already do voluntarily, because retyping costs hours and creates mistakes; from 2028 it stops being optional. ## The three paths to e-invoicing Every company will end up on one of three paths, depending on invoice volume and the age of its system. ### 1. The e-address: free, but manual The state channel is the e-address (e-adrese) on the Latvija.lv platform. It is free, and for ten invoices a month it is enough: invoices can be issued and received right in the portal, in 2028 and after it. The weakness of this path is volume: every step is manual, and what works for ten invoices does not work for a hundred. ### 2. An operator: a paid channel that runs itself E-invoice operators (in Latvia, for example, Unifiedpost, Telema, Omniva) run Peppol access points and connect your system to your counterparties' systems. Invoices leave and arrive automatically, and the operator is responsible for format and delivery. It costs money, but above a few dozen invoices a month it is usually cheaper than the human time spent handling invoices by hand. ### 3. Your accounting system, or a custom integration The major Latvian accounting platforms either ship e-invoice support already or are building it now. When I researched Horizon's integration options for a client project, I found that recent versions import Peppol XML files natively, with no extra programming: you hand the system the right file and the accountant keeps the approval step. Visma is preparing the same for Jumis users. If your system is maintained, a large part of the work has already been done for you. In a company with an old or self-built system the XML will not appear by itself: 2026 and 2027 are the window for integration work, connecting the existing system to an operator or the e-address and automating the flow to VID. The job is well defined and clearly priced, but at the end of 2027 everyone will be standing in the same queue. ## What to do now: five questions In my own company the books run automatically, and that is exactly how I know you do not get there in a day. These five questions show the size of the job: 1. **Where are your outgoing invoices created?** Accounting system, Excel, an invoicing tool, or self-built software? Every one of those sources must be able to produce Peppol XML in 2028. 2. **Where do your incoming invoices arrive?** Email PDFs, paper, an operator? Receiving needs a channel too. 3. **Is your accounting system still updated?** Ask your vendor directly: "how will mandatory e-invoices work in your system from 2028?" If they cannot answer concretely, plan as if the support will not arrive in time. 4. **What does your accountant say?** If your bookkeeping is outsourced, the transition has to fit their process. The best outsourced accountants will plan the switch for their clients themselves; the worst will stay quiet about it until autumn 2027. 5. **What stays outside the standard flow?** Receipts, foreign supplier invoices, photos from phones. The law will not turn those into structured files, and that is where automation pays off regardless of the mandate. ## Who has to move first If you send ten invoices a month from a maintained standard system, a calm transition in 2027 will do. But there are three situations where waiting is expensive: - **A self-built or aging system.** Integration work takes time, and developer capacity in 2027 will be sold out. - **High invoice volume.** More invoices means more processes to rebuild and more people to retrain. The voluntary phase from March 2026 exists precisely for you. - **Outsourced accounting firms.** If you serve dozens of clients, you make the transition not once but once per client. The firms that plan the migration as a project will work calmly in 2028; the firms that let it drift will be retyping invoices at night in December. ## Do it early or pay more The e-invoice transition is a rare case of the state using a law to force a cleanup that was worth doing anyway. Companies that do it in 2026 will enter 2028 with an automated invoice flow and cheaper bookkeeping, while the rest will pay the urgency price for the same work. If your invoice flow lives in an old or self-built system nobody wants to touch anymore, this is exactly the problem I solve: I connect the existing system to the e-invoice flow and automate what people have been retyping. Fixed price, fixed deadline, and the first step is a 30-minute conversation about your process. Write to ivarberg@lightsourcegames.com or book a call here: [calendar.app.google/bn9uzMB6Fh3Vt8SB6](https://calendar.app.google/bn9uzMB6Fh3Vt8SB6). --- ## An AI Can Generate Ten Landing Pages in a Minute. It Can't Tell You Which One to Ship. Date: 2026-05-27 Category: Lessons From the Trenches Reading time: 5 min URL: https://ivarberg.com/blog/slop-to-ship/ Last month I shipped puzzlywords.com - a Facebook web game ported to a standalone domain before Meta closes the platform on 2026-08-31. The brief was small: don't look AI-generated. That brief is harder than it sounds. Most landing pages produced with LLM help look the same - same radial-blur hero, same glass-morphism cards, same purple-to-pink gradient, same tilted phone mockup, same three-column feature row with emoji icons. None of those elements are wrong on their own. But when five of them show up together because the LLM defaulted to each one independently, the page reads as generic in a way users can sense even if they can't name it. What got the project from there to shipped wasn't a better prompt. It was a process. ## Slop isn't the element. It's the combination. The first thing to fix in your head: radial blur is not slop. Glass-morphism is not slop. Bento grids are not slop. Each of these is a legitimate design choice in the right context. The final puzzlywords.com landing *uses* a radial-gradient background - the client explicitly asked for it back when an early variant dropped it. It fits because the rest of the page (chunky tile buttons, taped notes, brand-sampled palette, real screenshots) is coherent around it. What reads as AI-default is the **combination**. Five-to-ten patterns that all show up together because the model defaulted to each one independently. Strip three of them and the rest stops feeling generic. This matters because the natural reaction to "looks AI-generated" is to outlaw individual elements - no gradients, no blur, no bento. That's the wrong move. You end up with a sterile editorial layout that's *also* slop, just a different flavor of it. The right move is to make sure each element on the page is there because *you chose it*, not because it was the default. ## Taste is still the creator's job An LLM can produce ten variants in a minute. It can run a perf pass. It can write JSON-LD schema. It can quantize PNGs and emit WebP fallbacks. What it cannot do is tell you which variant is *right*. It can only tell you which is statistically common. That part is you. Every elimination round in the process below is a human saying *that one, not that one*. The AI is the production line. You're the art director. If you skip the picking, you ship the average of the variants - which is exactly what slop looks like. In an era where producing a webpage costs nothing, the bottleneck is no longer production. It's discernment. Differentiation comes from the specific calls only someone with context can make: which one element carries the brand voice, which screenshot to use, which copy line is honest, which number to put on the stat card. ## The elimination method The fastest path past slop is subtraction, not creation. Three rounds. **Round 1 - 10 directions, 4 survive.** Generate ten distinct, complete landing variants in one pass. Vary the *direction*, not the polish: minimal/editorial, bold/playful, app-store, dark-tech, magazine, brutalist, migration-first, bento, arcade, screenshot-hero. Each one ~100-150 lines of HTML. Build a gallery page with all ten as iframes scaled to 50% so they're comparable on one screen. The client names which ones die. Don't argue. **Round 2 - 50 panels, 10 survive.** With four layouts on the table, switch from layout-thinking to language-thinking. Build one big grid of 50 small panel and button experiments - each one demonstrating a single styling treatment. Letter-tile, torn paper, ribbon, hexagon, ticket, banner, taped note, comic burst, coin, neon, brutalist, holographic. The client picks ten. **Round 3 - 10 skins, 1 winner.** Take the winning layout from Round 1. Apply each picked style as a *skin* - same content, same structure, only the panel and button language changes. Same H1, same hero image, same feature blurbs. The only thing varying is the visual vocabulary. The client picks one. A common over-correction in Round 3 is to combine two or three styles into a hybrid. Almost always worse than picking one. The page starts to feel over-designed - every element competing for attention. Keep the hybrid attempts in the archive so the client can see what they rejected, then move on. The better move is accents. After the winning skin is locked, apply that style to **one or two elements only**, leaving the rest of the page clean. Torn-paper feature cards on top of an otherwise restrained layout reads as authored. Torn-paper *everywhere* reads as a theme park. ## Voice is its own round After the visual is locked, do another 4-6 directions for just the headline, the notice copy, the feature blurbs, and the SEO meta. Voice has the same blast radius as design - the same product can read as cult-y, warm, direct, or heritage depending on how the words are tuned. Pick one direction and apply it everywhere. This is also where the SEO layer lands: the ``, the meta description, the JSON-LD VideoGame schema, the FAQ structured data, the OG image. None of this is design work. All of it can be batched into one pass once the voice is chosen. ## What this looked like in practice Round 1 cut six of ten in under a minute - the client typed *"1, 4, 5, 6, 7, 9 die"* and that was it. Round 2 picked eleven panel concepts from fifty. Round 3 settled on a torn-paper / tape style called v02-50. Three hybrid attempts were rejected. The final landing applies the style only to the feature cards. Everything else - hero, stats row, FAQ, footer - is clean. The whole sequence took one day. The result is a landing that reads as designed by someone, on a project where almost every pixel went through an LLM. --- *See the [full visual breakdown](/slop-to-ship) - every round, every cut, every screenshot.* --- *Need a landing page that doesn't look AI-generated? [Book a call](https://calendar.app.google/bn9uzMB6Fh3Vt8SB6).* --- ## I Built a Browser Skill for AI Because Screenshots Are Too Slow Date: 2026-04-09 Category: Workflow Automation Patterns Reading time: 5 min URL: https://ivarberg.com/blog/browser-automation/ Last month I asked Claude Code to check my Google Search Console numbers. It opened the browser, took a screenshot, analyzed it, figured out where to click, took another screenshot, analyzed that one, clicked, screenshot, analyze, repeat. It took nearly 3 minutes to pull data I could get in 20 seconds. That's the problem with screenshot-based browser automation. It works. But it's so slow that you're better off just doing it yourself. So I built something different. A single Python script - 1,200 lines - that talks directly to Chrome through its built-in remote control. No screenshots. No pixel analysis. Just text commands in, text data out. The same Search Console task now takes about 15 seconds. ## Why AI browser automation matters Most of your work happens inside a browser. Not on the open web - behind logins. Your analytics dashboard. Your CRM. Your project management tool. Your accounting portal. AI assistants can google things. ChatGPT, Gemini, Perplexity - they search the public web well. But they can't log into your tools. They can't check your specific numbers. **Browser automation bridges that gap.** It lets AI use your browser - the one where you're already logged in - to do the clicking and reading for you. The question isn't whether this is useful. It's whether it's fast enough to bother with. ## The three approaches ### AI web search What ChatGPT, Gemini, and Perplexity do by default. Search the public web and summarize. Great for public information. Useless for anything behind a login. ### Screenshot-based control What Claude's computer use and similar tools do. The AI takes a screenshot, processes the image, decides where to click, clicks, takes another screenshot. **The problem is tokens.** Every screenshot costs roughly 1,500 tokens. A 10-step task can burn through 15,000+ tokens just on images. browser-use, the most popular open-source AI browser agent (78,000+ GitHub stars), hit this exact wall. They dropped Playwright entirely in favor of raw CDP. The abstraction layer added too much latency across thousands of calls per session. A January 2026 benchmark found the numbers are stark: 114,000 tokens with MCP-based approaches vs. 7,000 tokens with a CDP-based agent for the same 10-step task. ### Direct browser control (CDP) Chrome has a built-in remote control called CDP - Chrome DevTools Protocol. Same technology that powers Chrome's developer tools. Instead of taking a picture and guessing what's on screen, CDP lets you **read the actual text**, **click exact elements**, and **extract data directly**. Screenshot-based is like describing a photo of a spreadsheet over the phone. CDP is like handing someone the actual file. ## How the script works A Python script called `cdp.py` sits between Claude Code and Chrome: ``` # 1. Launch Chrome + block images/CSS (60-80% faster loads) cdp.py ensure cdp.py block # 2. Restore yesterday's login session cdp.py cookies_load ~/.cookies/google.json # 3. Navigate (smart wait - no fixed sleep) cdp.py navigate "https://search.google.com/search-console" # 4. What's on this page? (~100 tokens, not 1,500) cdp.py axtree → [navigation] [button "7d"] [button "28d"] [heading "Performance"] # 5. Click and read cdp.py click "28d" cdp.py readable → "Performance: 4,820 clicks | 72,100 impressions | 6.7% CTR" # 6. Save session for next time cdp.py cookies_save ~/.cookies/google.json ``` Seven commands. Clean text data. No images processed. No re-login needed next time. ## Don't read the whole page. Read only what you need. The real trick isn't just "text instead of screenshots." It's asking for exactly the data you need: | Command | What it returns | Tokens | |---------|----------------|--------| | `screenshot` | Every pixel, layout, decorations | ~1,500 | | `content` | All visible text on the page | ~800 | | `readable` | Just the main content, no nav/footer | ~300 | | `links` | Only clickable elements | ~150 | | `axtree` | Accessibility tree - semantic structure, roles | ~100 | | `forms` | Only input fields and labels | ~80 | Same page. Six levels of precision. The accessibility tree is the most compact - it's the semantic structure screen readers use. On a complex page like GitHub, `axtree` returns ~100 tokens where `content` returns ~3,200. That's 27x fewer tokens for the same actionable information. ## Three more things that make it fast **Resource blocking.** Run `block` before navigating and the browser skips images, CSS, fonts, and video. Pages load 60-80% faster because AI agents don't need visual resources. **Session persistence.** Run `cookies_save` after logging in. Next time, `cookies_load` restores the session - no re-authentication, no password handling, no OAuth flows. The script remembers your logins. **Smart waits.** Navigation uses Chrome's Performance metrics instead of fixed sleeps. The script detects when a page is actually ready - fast pages proceed immediately instead of waiting a full second. ## The bigger point Google's Chrome team shipped an official MCP server for Chrome DevTools Protocol. Chrome 146 is adding a native settings toggle for AI agent access. The platform vendor itself is building toward structured programmatic access, not pixel interpretation. But the real lesson isn't about CDP vs screenshots. **It's about the speed threshold.** If an AI assistant takes 3 minutes to do what you can do in 30 seconds, you'll just do it yourself. Every time. Doesn't matter how smart the AI is. Six months of daily AI work taught me this: the bottleneck was never intelligence. It was always speed. Build for speed first. Intelligence follows. --- *See the [full visual breakdown](/browser-automation) comparing all three approaches.* --- *Want to see how AI browser automation works with real tools? [Book a call](https://calendar.app.google/bn9uzMB6Fh3Vt8SB6) - I'll show you the actual workflow.* --- ## The Dot-Com Test: A 30-Second Framework for AI Tool Decisions Date: 2026-04-01 Category: Lessons From the Trenches Reading time: 5 min URL: https://ivarberg.com/blog/dot-com-test/ # The Dot-Com Test: A 30-Second Framework for AI Tool Decisions I use AI tools every day. I also ignore most of what gets launched. Over the past year I've developed a simple filter - three questions I run before I spend money or time on anything new. I call it **The Dot-Com Test**. It takes 30 seconds. Here's why it works, where it comes from, and how to use it before your next AI purchase. --- ## The Three Questions Before you buy, renew, or recommend any AI tool, ask: 1. **Does it solve a real problem today?** Not next quarter. Not "when the feature ships." Today. 2. **Would you still pay for it if the hype disappeared tomorrow?** Strip the AI label. Is this still a product you need? 3. **Is it faster or cheaper than the alternative right now?** Not in theory. In your actual workflow. **Yes to all three = infrastructure.** This is a tool your business needs. Buy it. **No to any = speculation.** You're buying a lottery ticket dressed up as a software subscription. --- ## Where This Comes From It was 1999. The internet was going to change everything. And it did. But not before it destroyed companies that confused hype with value. Pets.com raised $82.5 million. Their Super Bowl ad was iconic. They sold pet food online at a loss and called it "internet strategy." They were dead within a year of their IPO. Amazon was also an "internet company." But Amazon solved a real logistics problem. People actually wanted to buy books without driving to a store. The internet made that faster and cheaper. Real problem. Real solution. Real value. Both were "internet companies." One passed the test. One didn't. The companies that survived the dot-com crash had one thing in common: they solved problems people would pay for on a boring Tuesday. No hype needed. --- ## Why Most AI Tools Fail The Test ### No real problem "It uses AI to summarize your meetings." I go to meetings where I actually listen and participate. I don't need a bot generating notes I'll never read. If something is important enough to remember, I write it down myself. If it's not - a summary won't change that. The real problem here isn't "I need better notes." It's "I'm in meetings that don't require my attention." An AI summary tool doesn't fix that. It just makes a bad habit more comfortable. ### Hype-dependent value "AI-powered project management." "AI-powered CRM." "AI-powered email." Strip the AI label. What's left? Usually a worse version of a tool that already exists, with a higher price tag and a waitlist. If the product can't stand on its own without "AI" in the pitch deck, it's speculation. ### Slower or more expensive "It automates your invoicing with AI." Does it process invoices faster than your current system? Does it cost less? If it takes your team two hours to set up, breaks on edge cases, and costs 3x what your current tool charges, you're paying an AI premium for worse results. Not in six months. Right now. --- ## Tools That Pass The Dot-Com Test Not every AI tool fails. The ones that pass share characteristics: **They replace manual work that was already painful.** I built a document processing pipeline for a dental clinic that extracts patient data from lab orders, X-rays, and invoices. Before: a staff member spent hours daily on manual data entry. After: automated extraction with human review only on edge cases. No viral demo. Just hours saved every day. **They work today, not next quarter.** The pipeline I built didn't need a roadmap to become useful. It processed its first real document on day one. No "coming soon." No beta waitlist. Real results on day one. **They survive the "boring Tuesday" test.** On a regular workday, with no conference hype, no LinkedIn posts about innovation, no pressure to "not fall behind" - you'd still use this tool. Because it makes your work faster, cheaper, or both. Here's what actually passes in my daily work: AI agents that scan team chats and surface things I'd otherwise miss. Automated task progress updates so I'm not chasing people for status. Code pipeline monitoring that catches and resolves CI/CD failures before they block the team. Document processing that eliminated manual data entry. None of these went viral. All of them solve problems I had before AI existed - they just solve them faster now. --- ## How to Run It in Your Organization Next time someone proposes an AI tool, run the test in your next meeting. Three questions. Out loud. 1. "What specific problem does this solve today?" Not "what could it do." What does it do. 2. "If we removed the AI label, would we still buy this?" This question alone kills 60% of proposals. 3. "Is this faster or cheaper than what we already have?" Measure it. Don't guess. If the answers are clear yeses, move fast. Infrastructure advantages compound. If any answer is shaky, you're speculating. That's fine if you know it. It's expensive if you don't. --- ## The 30-Second Version Every major technology has a bubble phase. The internet had one. AI is having one. The bubble will pop. The technology will stay. Your job isn't to avoid AI. It's to avoid the AI equivalent of Pets.com. The Dot-Com Test helps you tell the difference. Three questions. 30 seconds. Could save you six figures. Run it before your next purchase. --- *I build Business Operating Systems -- SOPs backed by software, executed by AI agents. Every tool in my stack passes The Dot-Com Test. If you want help evaluating which AI tools are infrastructure vs. speculation for your business, [let's talk](https://calendar.app.google/bn9uzMB6Fh3Vt8SB6).* --- ## 3 Months of AI Agents Changed Everything. There's No Going Back. Date: 2026-03-27 Category: The Human Side of AI Transformation Reading time: 9 min URL: https://ivarberg.com/blog/no-going-back/ # 3 Months of AI Agents Changed Everything. There's No Going Back. Three months ago, I quit optimizing business processes the old way. Today, I can't imagine going back -- and I don't think the world will let any of us. This isn't hype. This isn't a hot take for engagement. **I spent 6 months in the daily grind of working with AI agents** -- across very different industries, across completely different problem types -- and I came out the other side a fundamentally different builder. The tools I use, the speed I operate at, the expectations I have for what one person can accomplish -- all of it shifted. Permanently. And here's what makes this bigger than my personal story: **the competitive pressure building right now means sitting this out is no longer a viable strategy.** Not for individuals. Not for businesses. Not for entire industries. This is the opening post of what I'm calling **The AI Practitioner's Playbook** -- a series where I share frameworks, hard lessons, and real patterns from building with AI agents every single day. No theory. No "10 prompts that will change your life." Just what actually happens when you go all-in. Let's start with the most important framework I've developed. ## The Normalization Curve Everyone talks about the AI hype cycle. Gartner has their famous chart. But that model describes markets and technologies. **It doesn't describe what happens to YOU -- the individual practitioner -- when you commit to working with AI daily.** For that, I built a different model. I call it **The Normalization Curve**, and it has four phases: ### Phase 1: Hype (Weeks 1-4) Everything feels magical. You paste a problem into an agent, it generates a solution, and your jaw drops. **You overestimate what AI can do in every situation.** You tell everyone who'll listen. You think you've found a cheat code. I remember this phase clearly. First time an agent wrote a working API endpoint in minutes -- something that would have taken me half a day -- I thought: "This changes everything." And I was right. But not in the way I expected. ### Phase 2: Grind (Months 2-4) The magic fades. Fast. You start hitting real limitations. **The agent hallucinates a library that doesn't exist.** It writes code that looks perfect but breaks in production. It misses edge cases that any experienced developer would catch. You realize that directing an agent requires you to know MORE about the system than before, not less. This is where most people quit. They tried AI, it wasn't magic, they go back to the old way. **Phase 2 is the filter.** The people who push through it are the ones who develop the actual skill. The grind is learning what to delegate and what to direct. It's developing the instinct for when the agent is confidently wrong. It's building the systematic thinking that turns agent output from average into exceptional. ### Phase 3: Normalization (Months 4-5) Something shifts. You stop being impressed. You stop being frustrated. **You just expect it.** An agent refactors a module, writes the tests, and updates the documentation. You don't think "wow, AI did that." You think "good, what's next." The tool becomes invisible -- like a keyboard, like a compiler. It's just how you work now. This is a weird phase psychologically. The excitement is gone. But the productivity is real. **You're operating at a level that would have been impossible 12 months ago, and you barely notice.** ### Phase 4: Dependence (Month 6+) This is where I am now. And this is the phase nobody warns you about. **I physically cannot imagine going back.** Not because I'm lazy. Because the gap between what I can produce with agents and what I could produce without them is so large that the "without" version feels like going back to a typewriter. I rebuilt systems in 1-3 days that previously took months. I prototype ideas before lunch that would have been "next quarter" projects. The cognitive load of holding entire systems in my head shifted to directing agents that hold it for me. Dependence sounds negative. It isn't. You're "dependent" on electricity too. On running water. On the internet. **When a tool becomes infrastructure, dependence is just another word for civilization.** ## Why This Isn't Just My Story Here's where it gets bigger than personal preference. **72% of enterprises now have at least one AI workload in production**, up from 55% in 2024 and just 20% in 2020. The AI agents market hit $7.8 billion in 2025 and is projected to blow past $10.9 billion this year, growing at 46% annually. Gartner forecasts that **40% of enterprise applications will embed task-specific AI agents by end of 2026** -- up from less than 5% in 2025. Read those numbers again. This isn't gradual adoption. This is a phase transition. And the competitive dynamics are brutal. Companies using generative AI report average returns of **3.7x per dollar invested**, with top adopters hitting 10.3x. When your competitor is producing at that multiplier and you're not -- that gap compounds every single quarter. ## The Historical Pattern Is Clear We've seen this exact movie before. **Typewriters to computers.** In the 1980s, plenty of offices resisted personal computers. "Our typewriters work fine." Within a decade, not having a computer wasn't a philosophical choice -- it was a business death sentence. **Hand ledgers to spreadsheets.** Accountants who refused to learn Lotus 1-2-3 and then Excel didn't get to keep doing accounting the old way. They got replaced by accountants who could do in an afternoon what used to take a week. **Physical retail to e-commerce.** "Our customers prefer coming to the store." Some did. But the businesses that didn't build online presence didn't get to choose their timeline. **The pattern is always the same:** the new tool starts as optional, becomes advantageous, then becomes required. The window between "advantageous" and "required" is shorter every cycle. With AI, I'd argue we're already past "advantageous" and accelerating toward "required." The World Economic Forum estimates that AI and information processing will affect 86% of businesses by 2030. That's not a prediction about a far-off future. That's four years from now. ## The Pressure Is Not Personal -- It's Structural I want to be direct about something. When I say "there's no going back," I'm not expressing a personal preference. **I'm describing economic pressure.** The same kind of pressure that made every business adopt email, adopt the internet, adopt mobile. This pressure operates at every level: **Individual level.** A developer working with AI agents produces multiples of what they produced before. Not 10% more. Multiples. When hiring managers see this gap -- and they're starting to -- "I don't use AI tools" stops being a stance and starts being a liability. **Company level.** EY's March 2026 survey shows autonomous AI adoption surging across tech companies. When your competitors automate their operations and you don't, the cost differential alone will push you out. Not because AI is trendy -- because the math stops working. **Industry level.** Entire sectors are restructuring around AI capabilities. The first movers set the new baseline. Everyone else either matches it or gets priced out. This isn't a choice. It's pressure. Not bad -- but real. And pretending it isn't happening won't slow it down. ## A Call to Good Actors Now here's the part that matters most to me. **AI is the great multiplier of our personalities.** Just like money. Just like power. Give a good person more capability, they do more good. Give a bad actor more capability, they do more damage. Right now, as I write this, bad actors are already using AI at full speed. Scammers are automating their operations. Misinformation factories are scaling. Every tool that makes a builder more productive also makes a grifter more productive. And too many good people are sitting on the sidelines. Waiting. Watching. "I'll learn AI when it matures." "I'll adopt it when it's proven." "I'm not a tech person." **The worldly pressures won't let us go back. I suggest all good actors join in to shape the future.** This isn't about career optimization. This is about who builds the systems that run the world for the next 50 years. If good, thoughtful, ethical people cede that ground to whoever shows up first -- we all lose. You don't need to be a developer. You don't need to understand transformers or fine-tuning. You need to engage. Learn the tools. Apply them to your work. **Be in the room where the future is being built.** Because that room is filling up fast, and not everyone in it has good intentions. ## What I Do Differently Now After 6 months on the other side of The Normalization Curve, here's what changed: **I don't write code -- I direct agents that write code.** The skill shifted from typing to knowing what to ask for and what to watch out for. 15+ years of systems experience didn't become less valuable. It became the compass that makes agents useful instead of dangerous. **I don't plan in quarters -- I plan in days.** When you can prototype in hours and ship in days, the entire concept of project timelines changes. "Let's discuss this next quarter" now sounds like "let's discuss this in 2035." **I don't optimize processes -- I optimize how agents optimize processes.** The meta-game shifted. I used to be the one fixing workflows. Now I'm the one building the systems that fix workflows. **I expect more from myself and everyone around me.** Not in a burnout way. In a "the ceiling just got 10x higher" way. Average output is no longer acceptable when the tools exist to be exceptional. ## Where This Series Goes Next This post is the manifesto. The "why." Over the coming weeks, I'll get into the "how": - **How I actually direct AI agents** -- the workflows, the prompts, the systematic thinking that makes agents productive instead of chaotic - **The wall most people can't see** -- why chat interfaces are a ceiling, and what's on the other side - **Chaos in, chaos out** -- why messy thinking is the #1 reason agents fail, and how to fix it - **Proof over theory** -- real cases where AI-assisted builders outproduce experienced teams The Normalization Curve isn't theoretical. It's a map I drew from lived experience. **My goal with this series is to compress your Phase 2 grind** -- to help you get to normalization and dependence faster, with fewer wrong turns. Because the pressure is building. The window is narrowing. And the people who figure this out first will shape what comes next. No going back. Only forward. --- *This is Part 1 of The AI Practitioner's Playbook -- a series on building with AI agents in production, from someone doing it every day. Follow along for frameworks, real patterns, and honest lessons from the frontier.* *Have you hit your own Normalization Curve moment? What phase are you in? I'd genuinely like to know.* --- ## Agents Aren't Human - Stop Managing Them Like They Are Date: 2026-03-25 Category: The Human Side of AI Transformation Reading time: 9 min URL: https://ivarberg.com/blog/agents-arent-human/ # Agents Aren't Human - Stop Managing Them Like They Are You wouldn't do long division by hand. So why are you processing 200 Slack messages, 47 emails, and 12 status reports with your bare brain? **The biggest mistake in the AI discourse right now is anthropomorphizing agents.** We talk about them like junior employees. We worry they'll "take our jobs" or "go rogue." We assign them personalities. We debate whether they "understand" things. Stop. They're calculators. Really good calculators - but for information instead of numbers. I've spent the last six months working with AI agents every single day. Not experimenting. Not "exploring use cases." Building production systems. And the single biggest unlock wasn't a better model or a smarter prompt. It was stopping treating agents like people and starting treating them like infrastructure. That reframe has a name. I call it **The Cognitive Infrastructure Thesis.** And it changes everything about how you should deploy AI in your business. ## You Already Solved This Problem Once In the 1970s, engineers resisted pocket calculators. Hewlett-Packard launched the HP-35 in 1972 - the first scientific pocket calculator - and engineers were torn. Real engineers did math by hand. Slide rules were the craft. Using a calculator was cheating. By the 1980s, nobody cared. The calculator won. Not because it was smarter than humans. Because long division was never the point. The engineering was the point. Then spreadsheets. Accountants who'd spent decades mastering ledger books watched VisiCalc do in seconds what took them hours. Same resistance. Same fear. Same outcome. **Math by hand became calculator. Data analysis by hand became spreadsheet. Information processing by hand is becoming agent.** I lived a version of this progression myself. In 2022, I was using GPT-3.5 through an uncomfortable web interface. The code it produced was maybe 50% correct. I'd spend hours rewriting everything it generated. The AI gave direction - I did the actual work. It felt like using a calculator that was wrong half the time. Today I run multiple agents in parallel from my terminal, rebuilding enterprise systems in days that used to take months. The capability curve from "50% correct code in a chat window" to "production software in 72 hours" happened in under three years. Each time the pattern is the same: resisted, normalized, obvious. We're in the "resisted" phase right now with agents. Give it 18 months. ## The Volume Problem Nobody Talks About Honestly Here's a number that should make you uncomfortable: **knowledge workers face an interruption every 2 minutes during core work hours.** That's from a Microsoft study analyzing over 31,000 knowledge workers. The average worker receives 117 emails and 153 Teams messages daily. It takes 23 minutes to fully regain deep focus after each interruption. Do the math. There's almost no deep thinking left. **Workers spend 60% of their time on "work about work"** - chasing status updates, attending unnecessary meetings, switching between tools, coordinating across channels. Not thinking. Not creating. Not deciding. Processing. I felt this before I understood it. When I was managing operations for a company with over 100 employees, the bottleneck was never talent. We had smart, capable people. The bottleneck was information volume. Every day was a flood of questions, updates, exceptions, approvals. I went from being a "computer person" to a manager drowning in signals my brain was never designed to process at that scale. Your ancestors handled maybe a few dozen important pieces of information per day - weather, food, social dynamics in a group of 150. Now we process thousands. Daily. Willpower won't fix this. Another "productivity system" won't fix this. Getting up at 5 AM won't fix this. These are band-aids on an arterial bleed. ## The Cognitive Infrastructure Thesis Here's the framework that changed how I think about all of this. **Physical infrastructure freed humans from physical carrying.** Roads, bridges, railways, plumbing, electrical grids - each one removed a physical burden. Before plumbing, someone carried water. Before roads, someone carried goods on their back. Before the electrical grid, someone maintained a fire. Nobody argues we should go back to carrying water from the river. Nobody calls plumbing "lazy." **Cognitive infrastructure frees humans from information carrying.** Agents, automated workflows, intelligent routing, summarization systems - each one removes an information burden. Before agents, someone read every email. Someone triaged every support ticket. Someone compiled every status report. The pattern is identical: - Physical burden → physical infrastructure → humans freed for higher-order physical work (sports, art, exploration) - Cognitive burden → cognitive infrastructure → humans freed for higher-order cognitive work (judgment, strategy, empathy, creativity) **Agents aren't replacements for human thinking. They're infrastructure that makes human thinking possible again** - by handling the volume that was never meant for human brains. This isn't theory for me. At the company where I managed 100+ people, I built targeted automation - document processing through OpenAI APIs, automated reminders, salary calculation systems, procurement workflows. Nothing flashy. Basic cognitive plumbing. Within weeks, the staff couldn't imagine working without it. Not because the AI was brilliant. Because it handled the information volume that was crushing their capacity to do actual work. ## What Humans Are Actually Good At When I talk to business leaders, they're afraid agents will replace judgment. The opposite is true. **Agents are what finally give you time to exercise judgment.** Here's what humans do that no agent can: **Judgment in ambiguity.** When the data is incomplete, contradictory, or unprecedented - humans navigate this. Agents freeze or hallucinate. A CEO deciding whether to enter a new market with imperfect information. A manager sensing that a team member is burning out before any metric shows it. I make these calls daily when directing agents - which approach to take, which trade-off to accept, when to scrap something and start over. The agent executes. The human decides. **Empathy and relationship.** Closing a deal, managing a difficult conversation, building trust with a client over years. Agents can draft the email. Humans deliver the meaning. I learned this managing people from wildly different backgrounds at a dental operation - no amount of automation replaces the ability to read a room. **Creative direction.** Agents can generate options. Thousands of options. But choosing which option matters - that's taste, vision, experience. That's human. When I rebuilt enterprise software in 3 days with agents, the speed came from the agents. The architecture decisions - what to build, what to skip, what matters for this specific business - came from 15 years of building systems. **Ethical reasoning.** Should we do this? Not can we, not will it be profitable - should we? This question requires values, context, and moral weight that agents don't carry. Good actors must engage with AI precisely because bad actors already have. The technology amplifies whoever wields it. **Long-horizon strategy.** Where should this company be in 5 years? What's the second-order effect of this decision? Agents optimize within constraints. Humans set the constraints. **These skills are currently being crushed under the weight of information processing.** Your best strategic thinker is spending 60% of their day on "work about work." That's not a productivity issue. That's a waste of the most valuable resource in your organization. ## The Anthropomorphizing Trap The reason people resist this reframe is emotional, not logical. **We anthropomorphize agents because language models use words.** Calculators use numbers - nobody thinks their calculator is alive. Spreadsheets use cells - nobody worries their spreadsheet will rebel. But agents use natural language, and language feels human. So we project human qualities onto them. This creates two equally destructive errors: **False expectations.** "My agent should understand the company culture." "My agent should know what I really mean." No. Your calculator shouldn't understand poetry either. Use the tool for what it does. Use yourself for what you do. **False fears.** "Agents will replace all jobs." "Agents will make decisions they shouldn't." Calculators didn't replace mathematicians. Spreadsheets didn't replace accountants. Agents won't replace thinkers. They'll replace the parts of your job that aren't actually your job. I had to learn this the hard way. When I started working with agents full-time six months ago, I kept trying to make them think like me. I'd write elaborate prompts trying to transfer my judgment. It failed every time. The breakthrough came when I stopped. When I accepted that the agent handles volume and I handle direction. That's when everything accelerated. Gartner predicts 40% of enterprise applications will embed task-specific AI agents by end of 2026, up from less than 5% in 2025. The businesses getting this right aren't anthropomorphizing - they're building infrastructure. The ones stuck in debates about agent "trust" and "alignment" are watching their competitors ship. ## What This Means For Your Business If you run a company with 10-200 people, here's the practical takeaway: **Audit where your team spends time on information processing vs. actual work.** Research says 60%. Your number might be higher. Every hour your people spend reading, sorting, summarizing, and routing information is an hour they're not doing the work you hired them for. **Identify the 3 biggest information bottlenecks.** Where does information pile up? Where do people wait for summaries? Where do status updates require meetings that could be automated? Start there. **Deploy agents as infrastructure, not as employees.** Don't hire an "AI assistant." Build cognitive plumbing. Route information automatically. Summarize automatically. Flag exceptions for human judgment. Let the routine flow through the pipes so your people can focus on the exceptions that actually need a human brain. **Measure what matters.** Not "how many tasks did the agent complete" but "how much time did my team reclaim for judgment, strategy, and relationship work?" That's the metric that compounds. ## The Identity Shift Nobody Warns You About Here's what I didn't expect. Building cognitive infrastructure doesn't just change your workflow. It changes your identity. I spent 15 years as someone who builds systems. Who writes code. Who solves problems by sitting down and grinding through them. Agents made me let go of that identity. The value isn't in writing the code anymore. It's in knowing what to ask for. It's in understanding the business problem deeply enough to direct agents toward the right solution. That shift is uncomfortable. It feels like losing something. It took me six months of daily work with agents before it stopped feeling like loss and started feeling like leverage. The same thing happened to engineers who picked up calculators. They didn't stop being engineers. They became better engineers - because they spent their time on engineering instead of arithmetic. You won't stop being a leader when you build cognitive infrastructure. You'll start being one - because you'll finally have the cognitive space to lead. ## The Infrastructure Will Win The calculator won. The spreadsheet won. The agent will win. Not because any of them are smarter than humans. Because carrying - whether physical or cognitive - was never the point. The point was always the thinking, the creating, the deciding, the leading. The carrying was just the tax we paid because we didn't have better infrastructure. Now we do. **Build cognitive infrastructure. Free your people to do what humans actually do best. The organizations that don't will keep paying a tax their competitors stopped paying.** --- *If you're running a business and want to figure out where cognitive infrastructure fits, [let's talk](https://ivarberg.com/contact). I help companies build systems that handle information volume so their people can focus on actual work.* *References: Microsoft knowledge worker study via [Speakwise 2026 Knowledge Worker Productivity Report](https://speakwiseapp.com/blog/knowledge-worker-productivity-statistics). Enterprise AI adoption data from [Gartner](https://joget.com/ai-agent-adoption-in-2026-what-the-analysts-data-shows/). HP-35 history from [Smithsonian National Museum of American History](https://americanhistory.si.edu/collections/object-groups/handheld-electronic-calculators).*