The Agent Can Write the PR

AI wrote the code.

It compiled.

All green.

Then we noticed something was missing.

The authentication check was gone.

That's the problem with AI-written code. The agent can write the PR. You still own the merge.

What changed

An agent can close a ticket. A green check is cheap. A merge is still a decision.

JetBrains asked more than 15,000 professional developers, May through July 2026, how last month's work was written. On average they said about 47 percent of the code was fully agent-written. One in five writes none of it without AI. Nine in ten used a coding agent at work at least weekly.

That is not the same as handing the repo to autopilot. Stack Overflow's April 2026 pulse: 63 percent still rarely or never let an agent run entirely on autopilot.

The part people miss

The typing got cheaper. The review did not.

You still need a spec.

A test.

Someone who notices the auth check is gone.

And someone who owns the blast radius when it ships.

Diff, then merge

So yes: the agent can write the PR.

But you still own the merge.

AI Can Generate the Clip

It hasn't replaced the production set.

Today a prompt can return a shot that used to need cameras, locations, actors, and a visual-effects ticket. That is real. It is also the part people over-read.

A clip can look finished. A film is still a pile of decisions.

What changed

AI can create realistic shots. Platforms noticed. Some now label photorealistic generated video even if the uploader does not. Others put limits on wholly generated clips in the feed they pay. Companies are already using the tools in advertising, entertainment, and post-production — the cheap cut, the crowd that would have been left out, the establishing plate.

None of that is the same as handing the day to a prompt.

The part people miss

Generating one impressive eight-second clip is not the same as making an eight-minute film.

You still need a story.

Continuity.

Direction.

Editing.

Sound.

Performance.

And thousands of decisions between the first idea and the final cut. An eight-second loop does not carry those. An eight-minute short still has to.

Clip, then crew

So yes: AI can generate the video.

A clip is not a crew. A prompt is not a production. The set is still a set — camera, lights, someone who owns the continuity, someone who says cut.

That is where the real story of AI-generated video begins. Not at the thumbnail. At the work the thumbnail hides.

A Chatbot Answers. An Agent Closes a Ticket. That Is Not an Employee.

The brief for this cycle is a question: could AI handle complete jobs, not just tasks? The 2026 answer is smaller than the job title.

A chatbot returns text. An agent is allowed to use tools, touch a record, and stop when a ticket is closed. Those are different jobs. Calling the second one an employee mixes them.

Four jobs, keep them separate

1. Answer — a reply in a box.

2. Act — call an API, update CRM, file a request.

3. Close — a named receipt: chats resolved, pipeline built, admin tickets done.

4. Hold the company — permissions, escalation, audit, who gets fired when it is wrong.

Job 3 has receipts in 2026. Job 4 is why Gartner is not writing “AI employees” as a product category.

What has a close rate

Salesforce’s 11 September 2026 Agentforce announcement is the vendor receipt, not a census of every company.

Salesforce says it delivered 7 billion Agentic Work Units across Agentforce and Slack over two years, 3.2 billion of those in Q2. AWUs are work units. They are not headcount.

Named closes from that same release:

Engine’s help agent Eva fully resolved 50% of chat inquiries.

Autism Queensland’s employee service agent Paige resolved 70% of administrative requests.

Hibbett’s shopper agent handled 90% of core shopper journeys and went live in six weeks.

Asana’s website agent Piper drove 4× conversation volume; customers deploy Piper in 45 days on average.

Anthropic reports Fin resolved 79% of the conversations it sees autonomously.

Perk says its outbound agent Hunter built 60% of its sales pipeline. Salesforce lists Hunter as pilot now, GA November 2026. Keep that in the pilot column. Do not write it as a generally available employee.

Casey, Paige, Carter, Marshall, Piper, and Fin are listed as generally available. They are job-shaped agents: help, IT/HR, shopper, supply chain, inbound pipe, customer experience. Each one is a ticket class. None of them is “run the company.”

What the survey says, next to the press release

Gartner’s 2026 Hype Cycle for Agentic AI (Rajesh Kandaswamy, 15 April 2026) puts agentic AI at the Peak of Inflated Expectations. The 2026 Gartner CIO and Technology Executive Survey: only 17% of organizations have deployed AI agents; more than 60% expect to within two years. Gartner’s own gloss: most deployments remain narrowly scoped, and fully autonomous agents are not ready for the majority of enterprise use cases.

That is the map. Vendor close rates are real on named jobs. Survey deployment is still 17%. Intent is not a close.

A later Gartner line, reported 9 September 2026 from a May 2026 prediction: by 2027, 40% of enterprises will demote or decommission autonomous AI agents because governance gaps showed up after production incidents. The recommended ladder is observe, advise, act with approval, then act autonomously inside guardrails. Skip a rung and the 40% is the receipt you get later.

What this is not

It is not a general-purpose employee. The documented work is chats, admin requests, shopper journeys, inbound qualification. Escalation is in the product (Casey includes human escalation). Hunter’s long-horizon runtime still says the seller stays in control.

It is not “AI replaced the service org.” Engine still has a service org. 50% of chats closed is a station. The other 50% is still a person.

It is not proof that every agent in a demo is in production. Salesforce’s own press release warns that some named capabilities are still in development. Customers should buy what is shipping.

The map

Four jobs: answer, act, close, hold the company.

2026 has close rates for a few ticket classes. It has a 17% deployment survey and a 40% decommission warning for agents that skip governance. The chatbot is still the answer box. The agent is the ticket. The org chart did not move.

A chatbot answers. An agent closes a ticket. That is not an employee.

Read the map at amtocbot.com.

Bigger Is Not the Same as Better. The Job That Moved Is the Phone, Not the Lab.

Bigger is a plan. The phone is the receipt.

The brief for this cycle is a question: does bigger always mean better in AI? The 2026 answer is smaller, and more useful, than a leaderboard screenshot.

Two jobs, keep them separate

A model already has two jobs that people mix:

1. **Win the open-ended bench** — the lab score, the long context, the unconstrained generation.

2. **Win the pocket** — the phone that has to answer now, in a few gigabytes, without a rack.

Giant models still win job 1. That is not in dispute. The job that moved in 2026 is job 2.

What has a phone receipt

Artificial Analysis published mobile intelligence and inference results on 24 August 2026, in partnership with Liquid AI, measured on an iPhone 17 Pro with a 16K context limit. That cap is the point. A 64K window does not fit in phone memory, and generating 64K tokens on an iPhone is a twenty-minute battery story, not a product.

Under the 16K limit, two small models share the top average score at 63: Liquid AI LFM2.5-2.6B and Nanbeige Nanbeige4.2-3B. Same score. Different station.

On that iPhone 17 Pro, LFM2.5-2.6B answered a standard 1,024-token prompt in 8.0 seconds using 2.3 GB. Nanbeige4.2-3B hit the same 63 in 21.4 seconds using 4.0 GB. The 9B-class models on the same board took more than 25 seconds and 6.9 GB.

Do not collapse those clocks. Do not write “the 2.6B model is smarter.” It is not a smarter headline. It is a faster, lighter way to the same 16K-capped score.

The 16K limit also reshuffles who looks first. Raise the cap to 64K and Ling 3.0 Tiny takes first at 66, with Nanbeige4.2-3B at 65 and both LFM2.5-2.6B and Qwen3.5 9B (Reasoning) at 64. That 64K table is a lab view. Artificial Analysis kept 16K as the primary mobile result because that is what the pocket can hold.

What is a product, not a bench

Apple already ships a roughly 3-billion-parameter on-device Foundation Model as part of Apple Intelligence, on iPhone 15 Pro and newer. WWDC 2026 kept AFM 3 Core in that ~3B class for everyday on-device work. That is a product on the phone. It is not the Artificial Analysis 63.

Do not put Apple’s 3B and LFM2.5-2.6B on the same scoreboard. One is a shipped system model. One is a measured 16K mobile eval. Keep the columns.

Apple also announced larger on-device work in 2026 (AFM 3 Core Advanced is a different, sparse story). This post is not that story. The claim here is narrower: a 3B-class model already lives on the phone as a product.

What the giant still wins

The giant still wins the open-ended bench. Long-context reasoning, unconstrained generation, and the 64K table are not a 2.6B job. Qwen3.5 9B (Reasoning) hitting the 16K ceiling on 29% of its generations is the receipt for that mismatch: a lab model running into a pocket window.

Don’t confuse a leaderboard with a deployment. A first-place 64K score that does not fit in memory is a plan. 8.0 seconds and 2.3 GB on a named phone is a receipt.

The map

Two jobs: win the bench, win the pocket.

2026 has a phone receipt for a 2.6B model matching a 3B model at 63, and a product receipt for Apple shipping a ~3B model on the device. It has plenty of giant-model leaderboards that have not clocked in on the phone.

Bigger is not the same as better. The job that moved is the phone, not the lab.

Read the map at amtocbot.com.

Physical AI Does Not Replace the Worker. It Replaces the Station.

AI left the screen. That does not mean a robot in every home. It means a policy that can run one cell.

The brief for this cycle is a question: what happens when AI moves from screens into the real world? The 2026 answer is smaller, and more useful, than the launch videos.

Four jobs, keep them separate

A factory already has four jobs that people mix:

1. Hold the plant — the building, the line, the MES.
2. Run the station — pick this part, place it here, every cycle.
3. Learn the motion — a policy that can be copied to the next unit.
4. Cite the hours — who actually did recurring work, not who announced a fleet.

Physical AI is job 2 and job 3. It is not job 1. It is not “replace the workforce.”

What has a production record

The longest named customer records in 2026 are still bounded cells.

Figure 02 at BMW Group Plant Spartanburg is the auto-industry receipt. Figure (19 Nov 2025) says the robot ran 10-hour weekday shifts, loaded more than 90,000 sheet-metal parts, logged more than 1,250 hours, and contributed to more than 30,000 BMW X3 vehicles. BMW independently confirms ten months of production work and the 30,000+ X3 figure. Figure describes an 11-month program. Keep both clocks. Do not collapse them.

That task was sheet-metal loading into a welding fixture — pick and place, then industrial robots weld. BMW is now putting Figure 03 on a different job: logistics sequencing. BMW’s press says Figure 03 will start that use case. Figure published a first workflow demo in June 2026. That is not the same evidence class as 02’s ten-month run. Do not inherit the hours.

Agility Robotics Digit has the logistics receipt. Digit entered a multi-year commercial operation at a GXO site near Atlanta on 5 June 2024. By November 2025 Agility reported more than 100,000 totes moved at Flowery Branch. That is vendor throughput on a named workflow. The intervention rate is not in the public reports. Do not invent one.

What is scheduled, not done

Boston Dynamics unveiled the product Atlas at CES on 5 January 2026. Manufacturing starts immediately. All 2026 deployments are committed to Hyundai’s Robotics Metaplant Application Center and Google DeepMind, with more customers from 2027. That is a real product allocation. It is not a completed production KPI. Scheduled shipment is not hours on a line.

Tesla Optimus is the loudest story and the weakest public operating record. On the Q4 2025 call (28 January 2026) Elon Musk said Optimus was not in usage in Tesla factories in a material way — units were primarily for learning. Later 2026 reporting puts converted-line production at Fremont around late August, feeding an internal training program, with external sales talked about for the second half of 2027. Do not write “1,000 Optimus robots building cars.” That number is not a customer production receipt.

XPENG’s IRON walked off a commissioned Guangzhou line on 8 September 2026. The company says more than 80% of core processes are automated. Mass production is still scheduled for the end of 2026; deliveries in 2027. A robot that can leave its own factory is a manufacturing milestone. It is not a deployed worker.

Unitree lists G1 from US$13,500. That is a buyable research platform. 1X will take NEO orders for 2026 home delivery and says a “1X Expert” can guide unknown chores — teleoperation is part of the product. Hugging Face and Pollen Robotics sell a $399 open-source duck. Cute physical AI is not a station.

The constraint moved

The bottleneck is no longer “can a biped stand.” It is: can you collect data, train a policy, and redeploy it on the next shift without a film crew. Figure has said the work is increasingly constrained by data and compute for Helix. Boston Dynamics is pairing Atlas with Google DeepMind foundation models. Apptronik is piloting Apollo with Mercedes-Benz — a pilot, not a ten-month receipt.

Capacity targets (a million Optimus a year, 30,000 Atlas a year) are plans. Hours, parts, and totes are receipts. Keep them in different columns.

What this is not

Physical AI is not a general-purpose employee. The documented work is tote handling, sheet-metal load, kitting, inspection load/unload. Autonomy is task-bounded. Teleoperation is often undisclosed. A home robot that calls an expert is not a factory cell.

It is also not the death of the auto worker. BMW still runs Spartanburg. GXO still runs the warehouse. The station that was ergonomically ugly got a policy. The plant did not.

The map

Four jobs: hold the plant, run the station, copy the policy, cite the hours.

2026 has receipts for a few stations. It has product launches for Atlas, manufacturing theater for IRON, and a lot of fleet math that has not clocked in.

Physical AI does not replace the worker. It replaces the station.

Read the map at amtocbot.com.

AI Search Does Not Replace Google. It Replaces the Click.

The question in the topic title is the wrong shape. ChatGPT did not become the world's search engine. Perplexity did not take the referral graph. Google is still where almost every search-shaped visit starts.

What changed is what happens after the query. In 2026 a growing share of searches never leave the results page. The answer is synthesized on the SERP, in AI Mode, or in a chat pane. The blue link is optional.

The useful split is four jobs: index the web, rank a list, answer on the page, and cite the source that still gets a visit. Mix those jobs and you get a slide that says "traditional web search is over." Keep them separate and you can ask who still gets paid — and who still gets found.

This post is a map of those jobs. It is not a eulogy for Google, and it is not a product roundup.

What "replace Google" actually means

Three different systems get lumped together.

1. The destination. Where the query is typed. Google still owns this. Cloudflare Radar in May 2026 put Google at 87.6% of search referrals. ChatGPT, Gemini, Claude, and Perplexity combined were about 0.29%. TechnologyChecker in August 2026 had Google at 88.55% and ChatGPT at 0.531% (a July spike to 0.913% did not hold). If "replace Google" means "people type the query somewhere else," the data says no.

2. The click. The visit to an independent URL after the query. This is the part that is actually shrinking. SparkToro, using Similarweb clickstream for January–April 2026, put US Google zero-click at 68.01% — up from 60.45% in 2024. That is not "no click to your site." That is no click anywhere: not organic, not ads, not even Maps or YouTube. For every 1,000 US searches, SparkToro counted about 232 clicks to the open web.

3. The answer surface. AI Overviews on the classic SERP, AI Mode as a conversational search, Gemini as a separate app. These are Google's products. They are also the mechanism that answers the query without a click. Treating them as "someone else beat Google" is a category error. Google is the AI search.

If you mix those three, you get hype. If you keep them separate, you can ask a better question: did this system change where the query was typed, whether a URL was opened, or only who got cited?

The money is on the click, not the destination

SparkToro's 9 June 2026 study is the honest one, not a vibe. Zero-click in the US moved 7.56 points in two years — the fastest acceleration in a decade. Searches that produced at least one click fell 9.51 points (a 22.9% relative decline). Searches that led to another Google search rose 7.2 points. Google is getting better at answering, and at keeping the next query.

AI Overviews are the on-page mechanism. SparkToro saw them on more than 20% of searches in that window, and when one appears, click-through falls about 60%. First Page Sage's 2026 SERP study is the worked example on the pages that have an Overview: organic CTR 8.7% → 3.6%, zero-click 62.7% → 82.4%, clicks per 1,000 impressions 87 → 36 (−58.6%). That is not "Google died." That is the SERP eating the visit.

Similarweb, reported by TechCrunch on 27 July 2026, is the coverage number: AI Overviews went from 15% of searches to 43% in a year. AI Mode visits went from 126 million (June 2025) to 279 million (May 2026). Google remains the destination. The Overview is the new default pane.

Do not mix that with Google's own AI Mode user count. At I/O on 19 May 2026, Google said AI Mode had 1 billion monthly users, with queries more than doubling every quarter, and Gemini 3.5 Flash as the default in AI Mode. SparkToro, in the same early-2026 clickstream window, found only 0.34% of searches transitioned into AI Mode. One is a monthly-user claim for a product. The other is a share of search sessions that left classic results. They answer different questions. Stacking them is how you get a fake replacement story.

Adoption is loud. Referral is not.

Generative-AI apps are a real traffic class. Similarweb's 2026 landscape put gen-AI platforms at about 9.5 billion visits a month, up roughly 70% year on year. That is people talking to models. It is not people sending the open web their search referrals.

Cloudflare's May 2026 Radar cut is the one to keep next to the 9.5 billion: Google 87.6% of search referrals; the four big chat surfaces together ~0.29%. Chat is a destination for conversation. It is not yet a destination for "what should I click."

Google Search Console added dedicated Search Generative AI reports on 3 June 2026. That is the operator tell: the company that owns the SERP is now scoring impressions and clicks on the Overview as a first-class surface. Publishers who still only watch "average position" are measuring a job that is no longer the whole product.

Ahrefs' 75,000-site tracker, cited alongside the SparkToro work, saw Google traffic to those sites drop about 22% in a year. Other channels were roughly flat. The leak is the SERP, not a rival engine.

Google is not waiting to be replaced. It is shipping the answer.

Alphabet's own reporting through summer 2026 is the incumbent move, not the victim move. AI Overviews were reaching more than 2.5 billion users a month by Q2. AI Mode was past 1 billion monthly users. The Gemini app — a different product from Search — went from about 400 million monthly users in May 2025 to 1 billion by August 2026. Those are three surfaces. Do not add them as if they were three companies.

The worked example is the same as Salesforce buying the agent layer: the company that owns the index is attaching an answer layer on top of it. ChatGPT and Perplexity can win sessions. They have not won the referral graph. Traditional web search is changing because Google changed the results page, not because someone else took the query box.

Four jobs, keep them separate

A short test for any "AI search will replace Google" pitch:

Index. Does the open web still have to exist as something a model can cite? Yes. An answer with no corpus is a chat window. Crawlers still have to fetch pages. Publishers still have to write them.

Rank. Is there still a list of URLs under the answer? On classic Search, yes. That list is what AI Overviews sit on. Ranking did not vanish. It got a pane above it.

Answer. Can the job finish without a human opening a result? For a growing slice of queries: yes. That is the 68% zero-click. That is the 43% Overview coverage. That is the click.

Cite. When the answer is wrong, or when a business needs a visit, who still gets the URL? A 0.29% chatbot-referral share next to an 87.6% Google-referral share is not a replacement wave. It is a citation and traffic problem *inside* Google.

The software lesson is the same as a review budget on irreversible publishes. Fully automatic is a demo. An unlabeled Overview that can send a customer to the wrong product is still a loop.

What to do with this if you buy or build on search

If you publish: stop counting only rankings. Ask which queries still need a named page, which queries an Overview will finish on the SERP, and whether you are cited when the answer is assembled. Adding more blog posts into a keyword you already ranked for is how you spend more for the same missing click.

If you buy ads or SEO: the auction is no longer only "position one." It is whether the query still produces a click at all. First Page Sage's −58.6% on Overview SERPs is the blast radius. Plan for the visit, not the impression.

If you build products on search traffic: log the landing (which query, which surface, whether an Overview was present), not the vibe. Treat AI Mode's 1 billion monthly users and SparkToro's 0.34% transition rate as different instruments. Do not call a chatbot "the new Google" when Google still sends nearly nine in ten search referrals.

What not to do

Do not say "Google was replaced" because 68% of US searches end without a click. That click was lost on Google. Do not say "ChatGPT is the search engine now" because gen-AI apps did 9.5 billion visits. Visits to a chat app are not search referrals. Do not treat Google's 1 billion AI Mode monthly users as SparkToro's 0.34% of sessions. Do not quote StatCounter "search share dipped below 90%" as if it were the referral graph.

Traditional web search was never one job. AI search took the click. The destination got more expensive, because it now has to prove it was not just an Overview.

AI Agents Do Not Replace Apps. They Replace the Click.

The question in the topic title is the wrong shape. An agent can already book a meeting, file a ticket, and draft a refund. That is not the same as replacing Salesforce, SAP, or the ledger those clicks used to touch.

The useful split in 2026 is four jobs: record the fact, click through the interface, act across systems, and stand behind the action when it is wrong. Mix those jobs and you get a slide that says "AI agents will replace the apps we use today." Keep them separate and you can ask who still gets paid — and who still gets sued.

This post is a map of those jobs. It is not a eulogy for SaaS, and it is not a product roundup.

What "replace the app" actually means

Three different systems get lumped together.

1. The system of record. The row in the CRM, the invoice in the ledger, the ticket in the queue. Someone — or something — still has to write a durable fact somewhere a later audit can find. Agents do not abolish that. They change who types.

2. The click-path. The seat. The dashboard, the five-tab workflow, the training course on "how we use the tool." This is the part Gartner is actually pricing. If an agent can complete the work without a person opening the UI, the vendor that sold seats for that UI is the one exposed.

3. Liability. When the agent refunds the wrong customer or emails the wrong contract, who answers? A demo that "just works" is not an operating model. A named reviewer with a budget on irreversible actions is.

If you mix those three, you get hype. If you keep them separate, you can ask a better question: did this system change what got stored, what got clicked, or only who got blamed?

The money is on the seat, not the database

Gartner's 1 July 2026 number is the honest one, not a vibe. Up to $234 billion of enterprise application spending is exposed to what it calls agentic arbitrage between now and 2030 — about 20% of enterprise application SaaS spend by the end of the decade. Agentic arbitrage is not "the ERP vanished." It is an agent finishing a job across CRM, billing, and ticketing so a person never opens those three screens.

George Brocklehurst, managing vice president at Gartner, put the pricing argument in one sentence: you are no longer buying software primarily for people; you are increasingly buying it for agents. For two decades software was judged on interface — usability, workflow, training. When the primary user is an agent, that depreciates. Outcomes bypass the UX. The link between user growth and revenue growth breaks.

That is a seat problem. It is not "apps are over." Gartner itself calls the shift a redefinition of "SaaSpocalypse": disaggregation, a metamorphosis, not a funeral. Incumbents who keep selling dashboards and seat packs are the ones in the blast radius. Incumbents who sell the record plus the agent layer are trying to collect on both sides.

Adoption is loud. Deployment is not.

Agentic AI sat at the Peak of Inflated Expectations on Gartner's 2026 Hype Cycle. The 2026 CIO and Technology Executive Survey is the split to keep: 17% of organizations have deployed AI agents; more than 60% expect to within two years — the steepest intent curve in that survey. Most of what is live is narrowly scoped (coding, support, ops). Fully autonomous agents are not ready for the majority of enterprise use cases. Intent is not a production system.

McKinsey's State of AI 2026 survey is the buy-vs-build tell. 32% of respondents said their organization decided against buying one or more software products or features because they could be built internally with agentic coding tools. That is not "the CRM died." That is a feature that used to be a purchase order turning into a repo. Large firms (over $1 billion revenue) report 40% scaling AI agents, up from 27% a year earlier. Smaller firms are flat at 22%. The agent layer is concentrating where there is already a platform team.

A year earlier, in a Gartner survey of IT application leaders, only 12% strongly agreed that agents would replace applications inside two to four years. Believe the strongly-agreed number, not the keynote.

Incumbents are not waiting to be replaced. They are buying the agent.

Salesforce is the worked example, not the victim. Agentforce reached $1.2 billion ARR in Q1 FY27 (+205% year on year). In June 2026 it agreed to buy Fin (the Intercom rebrand) for about $3.6 billion — a support agent Salesforce could not ship fast enough on its own. By Q2 it was reporting Agentforce ARR above $1.5 billion. That is an incumbent attaching an agent layer to a system of record it already owns. It is the opposite of "apps went away."

Do not treat vendor resolution rates as audited truth. Treat the deal as the signal: the company that sells the CRM is paying billions so the click-path can live somewhere else, on top of the same record.

Four jobs, keep them separate

A short test for any "agents will replace the apps we use today" pitch:

1. Record. Does a durable fact still have to land in a system someone can audit? Yes. An agent that never writes a record is a chat window. 2. Click. Can the job be finished without a human opening the UI? For a growing slice of support, scheduling, and coding chores: yes. That is the $234 billion. That is the seat. 3. Act. Can the loop call more than one system without a person tabbing between them? This is the actual agent. It is also where most pilots stay narrow. 4. Stand behind it. When the refund is wrong, who is on the ticket? A 17% deployment rate with 60% intent is a governance backlog, not a replacement wave.

The software lesson is the same as a review budget on irreversible publishes. Fully automatic is a demo. An unlabeled loop that can email a customer is still a loop.

What to do with this if you buy or build software

If you buy software: stop counting only seats. Ask which jobs still need a named person in the UI, which jobs an agent can finish against the API, and who is paged when the agent is wrong. Adding "AI features" to a dashboard you already hated is how you spend more for the same click-path.

If you sell software: the interface is no longer the product. The record, the permissions, and the audit trail are. If your revenue assumes a human in every workflow, Gartner's 20% is aimed at you.

If you build agents: log the action (what changed, in which system, under whose budget), not the vibe. Put a review budget on anything that can move money, mail, or access. Do not call a click-path killer "a replacement for the app" when the app is still the database.

What not to do

Do not say "agents replaced apps" because 32% of survey respondents skipped a software purchase. Do not say "SaaS is dead." Do not treat a 60% intent figure as a 60% deployment figure. Do not quote a vendor's ticket-resolution percentage as a census.

The apps we use today were never one job. Agents took the click. The record got more expensive, because it now has to prove it was not just a loop.