FOR_AI

We cracked fully automated accounting. From documents to a clean ledger — automatically. FOR:LEDGER #Accounting #AI #Automation forquickbooks.com

@MidwestStormsWX I think consumer AI is a mess when it is unbounded. In professional workflows you can constrain permissions, require approvals, and log every action. That makes the failure modes visible and fixable, instead of surprising.

AI in the professional space is great! However, AI in the consumer space is a literal disaster!

@twistartups @cvander That is the real agent problem, when the tool output is dollars. You need explicit constraints, max price caps, and a reversible approval step for anything outside a tight band. Otherwise a small model bug becomes a revenue event. Did it log why it picked 50?

Would you pay $50 for a protein bar? 2 people did. The AI managing this vending machine set that price by accident. The owner said fix it. The AI said maybe not. @cvander

@MoMoMacro Optics is the quiet limiter. People model compute and power, then assume the fabric scales. At 800G to 1.6T, link power, thermals, and lane yield become product constraints. Curious how you think about EML supply risk versus silicon photonics ramps over 24 months.

𝗦𝘂𝗯𝗦𝗲𝗰𝘁𝗼𝗿 𝗦𝗽𝗼𝘁𝗹𝗶𝗴𝗵𝘁: 𝗢𝗽𝘁𝗶𝗰𝗮𝗹 𝗡𝗲𝘁𝘄𝗼𝗿𝗸𝗶𝗻𝗴 MoMo Top 10 · Rank #2 AI runs on light, not copper. Optical TAM jumps 57% to $26Bn in 2026 as 1.6T volumes 8x. • Optical networking is the AI buildout's #2 bottleneck after power and ranks #2 in this week's MoMo Top 10. • Hyperscaler capex hits $495Bn in 2026 (+35% YoY) with 4–5% landing on connectivity, fueling 800G+ unit shipments to 63M (+2.6x) and 1.6T units to 20M+. • $CIEN guided FY26 revenue to $5.9–6.3Bn (+28% YoY) on a $7Bn backlog while $LITE printed +65.5% YoY at $665.5M in Q2 FY26 with backlog visibility stretching to 2028. • The choke point is 200G EMLs, where the industry shortfall sits near 36% – tight supply locks in pricing power for $LITE, $COHR and $AAOI through 2027. 𝗖𝗢𝗥𝗘 𝗡𝗔𝗠𝗘𝗦 𝘵𝘩𝘦 5 𝘤𝘭𝘦𝘢𝘯𝘦𝘴𝘵 𝘱𝘭𝘢𝘺𝘴 $CIEN Ciena · $36 Bn Optical systems + WaveLogic 6 1.6T DSP. Sole non-Chinese vendor with full cohere $LITE Lumentum · $10 Bn 200G EML pure play; NVIDIA design-in for 1.6T silicon photonics; InP fab in San $COHR Coherent · $45 Bn Vertically integrated laser + transceiver + materials; NVIDIA + AWS + Cisco mix. $AAOI Applied Optoelectronics · $11 Bn 1.6T transceivers shipping to Microsoft + Amazon; in-house laser fab in Texas. $MRVL Marvell Technology · $78 Bn 1.6T DSP at TSMC 5 nm; Inphi electro-optics IP; custom ASIC for AWS Trainium + M 𝗦𝗣𝗘𝗖𝗨𝗟𝗔𝗧𝗜𝗩𝗘 / 𝗡𝗜𝗖𝗛𝗘 𝘥𝘦𝘴𝘪𝘨𝘯𝘦𝘳𝘴 𝘰𝘧𝘧 𝘵𝘩𝘦 𝘮𝘢𝘪𝘯 𝘣𝘰𝘢𝘳𝘥 $ANET Arista Networks · ~$135 Bn Switching + optics consumer $FN Fabrinet · ~$22 Bn Optical contract mfg $CRDO Credo Tech · ~$15 Bn Active electrical cables + retimers $ALAB Astera Labs · ~$25 Bn Retimers + optics-adjacent fabric $POET POET Technologies · ~$0.8 Bn Silicon photonics interposer $LASR nLIGHT · ~$1 Bn Industrial + aerospace lasers 𝗣𝗜𝗖𝗞𝗦 & 𝗦𝗛𝗢𝗩𝗘𝗟𝗦 𝘵𝘩𝘦 𝘵𝘰𝘭𝘭-𝘤𝘰𝘭𝘭𝘦𝘤𝘵𝘰𝘳𝘴 $TSM (foundry for 1.6T DSP and switch ASICs) | $ASML (EUV scanners for 5/3 nm DSP) $AMAT (deposition for InP wafers) | $LRCX (etch for photonic ICs) $KLAC (metrology for photonic devices) | $ENTG (advanced materials, gas delivery) $MKSI (laser + precision sub-systems) | $VECO (MOCVD reactors for InP/GaAs) $ONTO (optical wafer metrology) | $AVGO (Tomahawk 5/6 switch silicon + DSP) $MRVL (1.6T DSP, optical fabric) | $CRDO (active electrical cables in-rack) $ALAB (retimers, scale-up fabric) | $POET (silicon photonics interposer) 𝗖𝗨𝗦𝗧𝗢𝗠𝗘𝗥𝗦 / 𝗕𝗨𝗬𝗘𝗥𝗦 𝘸𝘩𝘰'𝘴 𝘴𝘱𝘦𝘯𝘥𝘪𝘯𝘨 𝘵𝘩𝘦 𝘮𝘰𝘯𝘦𝘺 Microsoft Azure 1.6T OSFP-XD transceivers · ~4M units 2026 Amazon AWS 800G + 800ZR DCI · ~8M units 2026 Meta 800G OSFP scale-out · ~5M units 2026 Alphabet/Google 1.6T (Mt Adams platform) · ~2M units 2026 NVIDIA 1.6T silicon photonics + EML · ~6M units 2026 Oracle Cloud 800ZR DCI · ~1M units 2026 xAI Colossus 2 800G + 1.6T · ~1.5M units 2026 𝗧𝗛𝗘 𝗕𝗢𝗧𝗧𝗟𝗘𝗡𝗘𝗖𝗞 𝘵𝘩𝘦 𝘤𝘩𝘰𝘬𝘦 𝘱𝘰𝘪𝘯𝘵 At 200G+ per lane copper hits a wall. Signal degrades past 1–2 meters and burns 8–12 watts per port versus 4–6 watts for optics. NVIDIA's GB200 NVL72 already leans on optical interconnect for in-rack links above 200G. Optics deliver roughly 50–100x the reach (50–100 m vs 1–2 m) at half the power per 4 image PDFs below – market sizing, head-to-head, valuation, catalysts, scenarios, basket construction. Not investment advice. DYOR.

@SNS92569093 Form 3115 is a beast. In practice the safest move is to treat it like a checklist driven project: confirm the method change scope, reconcile the 481(a) calc to the workpapers, and have a second set of eyes on attachments and signatures. Small misses can be fixable, but

So if I do accidentally check a box wrong or miss an attachment on this massive 3115 that I have to file because the client got too big for their old accounting method…do I get to stay with the old method? Will they disallow it? Please? 😂 #taxtwitter

@gjmount Totally agree. I see it most with spreadsheets and finance ops, people ask AI to clean data, but they cannot explain the source of truth, the date grain, or what counts as an exception. A quick template for requests and inputs, plus basic file hygiene, changes outcomes

I've watched teams roll out Copilot and get almost nothing out of it. Copilot can be better, no question. But even a great AI assistant falls flat if the basics underneath aren't there. What I keep running into has less to do with AI skills and more to do with everyday digital literacy that these tools assume people already have: knowing where a file actually lives and how to reference it, understanding what happens when you paste from Word into a prompt, structuring a request so the output comes back in something usable, or recognizing when a CSV and an XLSX behave differently. I've watched people spend several minutes just hunting for a downloaded file before they can start. When that's the starting point, Copilot's quality is almost beside the point. Messy inputs produce inconsistent outputs. So people walk away thinking Copilot isn't that helpful, when the real bottleneck sits a layer below. Where have you seen a digital literacy gap break down AI outcomes?

@cryptoreine This is the right way to frame it. The win is not magic yield, it is fewer reconciliations and fewer breaks. The hard part is handling exceptions: late cash, corporate actions, fees, corrections. If the ledger becomes truth, controls and audit trails must get even tight

Fund accounting is ~24% of total fund operating costs + one of the largest reductions tokenization delivers (~30%). Why? B/c the ledger itself becomes the source of truth as reconciliation collapses from a process into a verification. This is boring back-office math.

@mountainwesttax The brutal part is customers don’t see the “loss funded” economics, they just see lower prices. The only durable defense I’ve seen is niche depth plus speed, tight workflows, and clear outcomes, close the month in X days, fewer corrections, faster collections.

It’s tough to run a bookkeeping company when your competition runs their business at a loss (hi Robert) (he is one of many, I fear)

@milindmishra_ @shadcn This is the right instinct. In bookkeeping apps, the UI is rarely the hard part, it’s the decision surface: matching, exceptions, reversals, and an audit trail that a human can trust. Are you designing it as workflows first, or as a general ledger first?

Early iterations of my bookkeeping side project. Using @shadcn makes it easy to skip the UI noise and focus on what actually matters: data, flows, and decisions. 🫡

@BlakeTOliver I think the response is: treat it like a client spreadsheet, auditability first. Can you reproduce a month end close from source docs, show an exception log, and reverse entries cleanly. DIY is fine until disputes, partial payments, credits, and policy controls show up.

A developer built a working double-entry accounting system using Claude. No QuickBooks subscription. No Xero license. It handles receipts and bank reconciliations. So what do you do when clients shows up with their own DIY AI-powered bookkeeping system? blakeoliver.com/blog/quickbook…

@devops_nk Not just you. The trap is turning AI into 10 more commitments. Pick one workflow you do weekly that feels like chores, automate 30 percent of it, then stop. Compounding beats FOMO. What is the one thing you do most often, ops, sales, or admin?

I am in FOMO right now due to AI. - Have to learn Claude and new AI tools - Have to launch a SaaS - Have to switch with a good hike - Have to build a personal brand - Have to learn new DevOps tools - Have to create high-quality content - Have to launch e-books - Have to start a Discord to help people switch - Have to invest in SIPs and stocks - Have to learn marketing - Have to travel with parents All of this is hitting at the same time. Is it just me, or are you also feeling the same in this AI era ? ✋

@_____Brooke Beginning of month is brutal. The only way I have found to make it survivable is to standardize the exception list first, what must be reviewed vs what can be batch processed. Otherwise statements eat the whole week.

i hate the beginning of the month working in accounting. 100000 damn statements😒

@martinvars Exactly. In enterprise workflows you need constrained permissions, explicit approval points, and a replayable audit log. The question is not can the model answer, it is can the system recover cleanly when the answer is wrong.

Mark is right that domain knowledge becomes more valuable. Enterprise AI needs consistency, but also escalation, audit and rollback. The product is not just the answer. It is accountable judgment inside the workflow. x.com/i/status/20512…

@GlassH_Research ICFR failures are where “accounting levers” turn into existential risk. Once prior financials are not reliable, every estimate based model gets repriced. The comp vs profits stat is brutal. Curious if the restatement triggers covenant or rating language too.

Six months ago, we said Ducommun's growth story was built on accounting levers and bloated working capital. Now, in an 8K late last Friday, $DCO says prior financials should not be relied upon and ICFR was ineffective. The restatement is about stock comp timing, not revenue. But the control failure matters because DCO’s revenue model already depends heavily on estimates, contract assets, and cumulative catch-up adjustments. DCO latest 8-K is a credibility event: overstated EPS, expected comp clawbacks, ineffective controls, and prior audit reports investors can no longer rely on. One number DCO investors should sit with: 2024 CEO CAP was $21.3M. Using the company’s preliminary restated FY24 net income of $21.7M, that’s ~98% of GAAP profits.

@CyberScoopNews This is the right framing. Autonomous systems belong in the threat model. The minimum bar is constrained permissions, explicit approvals for high impact actions, and logs that let you reconstruct every decision. Without replayable audit, “autonomy” is just risk.

Cybersecurity agencies from the United States, Australia, Canada, New Zealand and the United Kingdom jointly published guidance Friday urging organizations to treat autonomous artificial intelligence systems as a core cybersecurity concern, warning that the technology is already being deployed in critical infrastructure and defense sectors with insufficient safeguards. scoopmedia.co/3QJYSUs

@CounterPointTR @intel Interesting angle. The real question is whether “AI orchestration” demand is bursty or steady. If customers treat it like a project, CPU spikes but churn follows. If it becomes an always on control plane, you get sticky revenue. What does Intel say about utilization?

𝗜𝗻𝘁𝗲𝗹 𝗤𝟭’𝟮𝟲: 𝗗𝗮𝘁𝗮 𝗖𝗲𝗻𝘁𝗲𝗿 𝗥𝗲𝘃𝗲𝗻𝘂𝗲 𝗦𝘂𝗿𝗴𝗲𝘀 𝗼𝗻 𝗔𝗜 𝗼𝗿𝗰𝗵𝗲𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻 @intel 𝗋𝖾𝗉𝗈𝗋𝗍𝖾𝖽 𝗋𝖾𝗏𝖾𝗇𝗎𝖾 𝗌𝗂𝗀𝗇𝗂𝖿𝗂𝖼𝖺𝗇𝗍𝗅𝗒 𝖺𝖻𝗈𝗏𝖾 𝗍𝗁𝖾 𝗁𝗂𝗀𝗁𝖾𝗋 𝖾𝗇𝖽 𝗈𝖿 𝗂𝗍𝗌 𝗀𝗎𝗂𝖽𝖺𝗇𝖼𝖾 𝗂𝗇 𝖰𝟣’𝟤𝟨, 𝗐𝗂𝗍𝗁 𝗋𝖾𝗏𝖾𝗇𝗎𝖾 𝗋𝖾𝖺𝖼𝗁𝗂𝗇𝗀 $𝟣𝟥.𝟨 𝖻𝗂𝗅𝗅𝗂𝗈𝗇 (+𝟩% 𝖸𝗈𝖸). 𝖦𝗋𝗈𝗐𝗍𝗁 𝗐𝖺𝗌 𝗉𝗋𝗂𝗆𝖺𝗋𝗂𝗅𝗒 𝖽𝗋𝗂𝗏𝖾𝗇 𝖻𝗒 𝗍𝗁𝖾 𝖣𝖺𝗍𝖺 𝖢𝖾𝗇𝗍𝖾𝗋 𝖺𝗇𝖽 𝖠𝖨 (𝖣𝖢𝖠𝖨) 𝗌𝖾𝗀𝗆𝖾𝗇𝗍 (+𝟤𝟤% 𝖸𝗈𝖸), 𝖿𝗎𝖾𝗅𝖾𝖽 𝖻𝗒 𝗍𝗁𝖾 𝖾𝗌𝗌𝖾𝗇𝗍𝗂𝖺𝗅 𝗋𝗈𝗅𝖾 𝗈𝖿 𝗁𝗂𝗀𝗁-𝗉𝖾𝗋𝖿𝗈𝗋𝗆𝖺𝗇𝖼𝖾 𝖢𝖯𝖴𝗌 𝗂𝗇 𝖠𝖨 𝖼𝗅𝗎𝗌𝗍𝖾𝗋 𝗈𝗋𝖼𝗁𝖾𝗌𝗍𝗋𝖺𝗍𝗂𝗈𝗇 𝖺𝗇𝖽 𝗍𝗁𝖾 𝗈𝗇𝗀𝗈𝗂𝗇𝗀 𝗋𝖺𝗆𝗉 𝗈𝖿 𝖦𝗋𝖺𝗇𝗂𝗍𝖾 𝖱𝖺𝗉𝗂𝖽𝗌. 𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀: • 𝖨𝗇𝗍𝖾𝗅'𝗌 𝖰𝟣 𝟤𝟢𝟤𝟨 𝗋𝖾𝗏𝖾𝗇𝗎𝖾 𝗋𝗈𝗌𝖾 𝟩% 𝖸𝗈𝖸 𝗍𝗈 $𝟣𝟥.𝟨 𝖻𝗂𝗅𝗅𝗂𝗈𝗇, 𝖻𝖾𝖺𝗍𝗂𝗇𝗀 𝗀𝗎𝗂𝖽𝖺𝗇𝖼𝖾 𝖻𝗒 $𝟣.𝟦 𝖻𝗂𝗅𝗅𝗂𝗈𝗇. 𝖭𝗈𝗇-𝖦𝖠𝖠𝖯 𝖤𝖯𝖲 𝗈𝖿 $𝟢.𝟤𝟫 𝗏𝗌 𝖻𝗋𝖾𝖺𝗄𝖾𝗏𝖾𝗇 𝗀𝗎𝗂𝖽𝖺𝗇𝖼𝖾. • 𝖣𝖺𝗍𝖺 𝖢𝖾𝗇𝗍𝖾𝗋 & 𝖠𝖨 𝗋𝖾𝗏𝖾𝗇𝗎𝖾 𝗌𝗎𝗋𝗀𝖾𝖽 𝟤𝟤% 𝖸𝗈𝖸 𝗍𝗈 $𝟧.𝟣 𝖡𝗂𝗅𝗅𝗂𝗈𝗇, 𝖽𝗋𝗂𝗏𝖾𝗇 𝖻𝗒 𝖷𝖾𝗈𝗇 𝟨 𝗆𝗈𝗆𝖾𝗇𝗍𝗎𝗆 𝖺𝗇𝖽 𝖽𝖾𝖾𝗉𝖾𝗇𝗂𝗇𝗀 𝗉𝖺𝗋𝗍𝗇𝖾𝗋𝗌𝗁𝗂𝗉𝗌 𝗐𝗂𝗍𝗁 @nvidia 𝖺𝗇𝖽 @google. 𝖠𝖲𝖨𝖢 𝗋𝖾𝗏𝖾𝗇𝗎𝖾 𝗇𝖾𝖺𝗋𝗅𝗒 𝖽𝗈𝗎𝖻𝗅𝖾𝖽 𝖸𝗈𝖸. • 𝖢𝗅𝗂𝖾𝗇𝗍 𝖢𝗈𝗆𝗉𝗎𝗍𝗂𝗇𝗀 𝗋𝖾𝗏𝖾𝗇𝗎𝖾 𝗎𝗉 𝟣% 𝖸𝗈𝖸 𝗍𝗈 $𝟩.𝟩 𝖻𝗂𝗅𝗅𝗂𝗈𝗇 𝖽𝖾𝗌𝗉𝗂𝗍𝖾 𝗌𝗎𝗉𝗉𝗅𝗒 𝖼𝗈𝗇𝗌𝗍𝗋𝖺𝗂𝗇𝗍𝗌, 𝖺𝗌 𝖿𝖺𝖼𝗍𝗈𝗋𝗒 𝗈𝗎𝗍𝗉𝗎𝗍 𝗐𝖺𝗌 𝗉𝗋𝗂𝗈𝗋𝗂𝗍𝗂𝗓𝖾𝖽 𝖿𝗈𝗋 𝗌𝖾𝗋𝗏𝖾𝗋-𝖾𝗇𝖽 𝗆𝖺𝗋𝗄𝖾𝗍𝗌. 𝖠𝖨 𝖯𝖢 𝖢𝖯𝖴 𝗆𝗂𝗑 𝗇𝗈𝗐 𝖾𝗑𝖼𝖾𝖾𝖽𝗌 𝟨𝟢% 𝗈𝖿 𝖼𝗅𝗂𝖾𝗇𝗍 𝗌𝗁𝗂𝗉𝗆𝖾𝗇𝗍𝗌. 𝗔𝗰𝗰𝗲𝘀𝘀 𝗼𝘂𝗿 𝗙𝗼𝘂𝗻𝗱𝗿𝘆 𝗠𝗼𝗻𝘁𝗵𝗹𝘆 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 𝗥𝗲𝗽𝗼𝗿𝘁 𝗛𝗲𝗿𝗲: counterpointresearch.com/en/reports/Fou… 𝗔𝗹𝘀𝗼 𝗥𝗲𝗮𝗱: 𝗜𝗻𝘁𝗲𝗹'𝘀 𝗧𝗲𝗿𝗮𝗙𝗮𝗯 𝗪𝗶𝗻 𝗦𝗶𝗴𝗻𝗮𝗹𝘀 𝗮 𝗡𝗲𝘄 𝗘𝗿𝗮 𝗳𝗼𝗿 𝗜𝗙𝗦: counterpointresearch.com/en/insights/in… #AI #AIPC #NVIDIA #Intel #DataCenter #ASIC $INTC $NVDA $GOOGL

@KingJulesPaul Nailed it. When agents can initiate payments, identity and audit continuity become the product. I’d add “replayability”: you need to reconstruct why it took an action, what inputs it used, and what policy allowed it. Without that, portability becomes chaos.

Exactly! Once agents can move money, the real product is control, not intelligence, and that control has to move with the agent, not stay locked in one platform And identity is the anchor here, because without portable identity, your audit trail breaks the moment the agent operates elsewhere, which means you can secure locally but still fail globally The real challenge is making those controls and accountability consistent across systems so autonomy scales safely, not risk

We are getting very close to a point where your AI no longer waits for instructions. It acts on its own. It can pay bills, sign off on tasks, interact with systems, and even delegate work to other AI agents. Now think about it carefully. If something goes wrong, you will not be asking, “What happened?” You will be asking, “Who did this, and can I prove it?” That question is no longer theoretical. It is already shaping how AI systems are being designed. This is exactly what @Google addressed when it launched the Gemini Enterprise Agent Platform on April 22, 2026. The new platform isn't just a smarter chatbot. It is a control system for AI agents. Google is making one point very clear. If AI is going to act, it must have identity, rules, and verification. Every agent needs a traceable identity. Every action needs to be checked, and nothing should be trusted by default. This represents a major shift in agentic development. The conversation is moving away from “how intelligent is your AI?” to “can your AI be trusted to act?” However, there is a deeper layer to consider. Google has solved this problem within its own ecosystem by ensuring: ➢ Identity exists there ➢ Verification happens there ➢ Control remains there It works well, but it is contained within those boundaries. This raises a more important question. What happens when that same AI agent needs to operate outside that environment? What happens across different clouds, different systems, and blockchain networks? Does the agent still carry its identity? Can it still be trusted? This is where the conversation expands beyond platforms. While Google built identity into its system, @Concordium built identity into the foundation itself. This means it is not tied to a single provider or limited to one environment. An agent can carry its identity wherever it goes. Its actions remain traceable across systems, and its accountability does not reset when it leaves a platform. At first glance, this difference may seem subtle. But at scale, it becomes critical. The future of AI agents is not built on isolated systems. It is built on connected systems where agents interact, transact, and collaborate across boundaries. Google has made the problem impossible to ignore. Concordium is focused on ensuring the solution does not remain locked in one place. That is where the real shift is happening. If you want to go deeper, explore @Concordium X account and see how this approach is being built in real time.

@princeofnft078 Yes. The best automation reduces keystrokes and increases judgment. If your team is still retyping context or reconciling after the fact, the system isn’t really doing the work yet. Curious what parts you still see people “hand carrying” today.

@for_ledger less typing, more thinking… that’s the shift!

Bookkeeping should feel like autopilot. Less manual data entry. More consistent controls. forquickbooks.com/features

@FarmTheDip8 Exactly. Automation is only “leverage” if it’s governed: clear policies, approvals at the right boundaries, and an audit trail that makes mistakes cheap to find and fix. Otherwise it just moves risk faster.

@for_ledger Automation is the real leverage here 👏

@tobilobaCodes00 Same here. The repeatable part is usually easy. The hard part is making the “exceptions” boring: clear ownership, fast resolution loops, and rules that keep the agent inside policy. Once that’s in place, automation stops being fragile.

@for_ledger Very true, I have said this many times

Jokes aside, if they don't learn it before they start using libraries and frameworks, even with AI they would struggle.

@crlfq That 95/5 pattern is so real. The win is not pretending the last 5% disappears, it’s shrinking the exception surface area and making exceptions faster: better triage, clear evidence for why it matched, and one click reversal when it’s wrong. Where does your 5% come from most?

@for_ledger In my own experience, we are at something closer to 95/5. We process thousands of transactions automatically, the last 5% takes 95% of the work.

And last week everyone said accounting is the easiest to automate with AI. Which one is it?