What if your service reps had the answer before the customer finished the question? In this AI agent assist customer service podcast episode, two members of the Augusto team talk through why the best first step is helping the rep rather than replacing them, the intelligence-first path from augmentation to automation, and what agent assist does to the service backlog and to the time reps spend on high-judgment work. We close with what this looks like in practice and why you keep the rep and lose the lookup.
Your developers already have AI assistants, so why is delivery not faster? In this AI software development lifecycle podcast episode, two members of the Augusto team talk through why faster code is not faster delivery, where AI pays off upstream in backlog grooming and acceptance criteria, and why review and testing gates still decide what ships. We close on the practical fix: get the specs, the criteria and the gates right first, then add the speed.
Writing code with AI and still shipping at the same pace? Reach out. We start every engagement with a fixed-price, two-week diagnostic that measures how long a clear requirement takes to reach production.
A manufacturing leader described his ideal AI project to us in September, and it was not a chatbot. His customers are blue-collar, local, and loyal, and he was blunt about it: they are a long way from feeling well served by a robot. What he wanted was an agent that lives alongside his people, listens in, and hands the rep the drawings, parts lists and fit notes they need, instead of making them dig through a twenty-year-old database while the customer waits.
That is agent assist, and it is the version of AI in customer service that mid-market companies can actually deploy without gambling the relationships that built the business.
Answer the customer, not the phone
The pressure on service teams is simple. Volume grows, the systems that hold the answers age, and the people who know the exceptions retire. A rep fielding a call about a motor that is no longer manufactured has to find the original drawing, check whether the “direct replacement” is a half inch too long, and decide whether a kit is needed. Today that lives in someone’s head or across QuickBooks, spreadsheets and a CRM that do not talk to each other.
Agent assist attacks the lookup, not the conversation. The rep stays on the call. The AI pulls the record, surfaces the relevant history, and drafts the answer for a human to confirm. The evidence for this design is strong. A study of more than 5,000 support agents published through the National Bureau of Economic Research found an AI assistant lifted issues resolved per hour by 14 percent on average, with a 34 percent gain for newer agents, and no drop in customer satisfaction. The same study saw fewer escalations to managers and better retention. The Bureau of Labor Statistics noted the tool worked by capturing what the best agents did and sharing it with everyone else.
The AI does not answer the customer. It makes sure your rep never has to say “let me call you back.”
Intelligence first, then augmentation, then automation
The manufacturing leader in the opening had already worked out the sequence, and it is the one we recommend.
Intelligence: Connect the sources. Drawings, part histories, past quotes, service tickets and customer records get indexed so a question can be answered from all of them at once. This alone removes the twenty-minute hunt.
Augmentation: Put the answer in front of the rep during the call. A suggested response, the matching drawing, the last three orders from this customer, and a flag when a replacement part has a known fit issue. The human decides and speaks.
Automation: Only after the assist has proven itself do you let AI handle the routine, low-risk cases end to end, and only with a clear path to a person. Our framework for when AI should make the call is the test we apply.
Skipping to step three is how companies end up with a robot their customers resent. Gallup found that 68 percent of employees who had used AI with customers said it improved the interaction, while only 13 percent of those who had not believed it would. Experience with the assist is what builds trust in the automation.
What it does to the numbers
The results we see follow the research. One customer service team we work with saw its open request backlog fall from several hundred to around one hundred after AI took over the routine quoting that used to bury them. Their manager’s read was that they had not been in yellow or red queue status for weeks because they could finally keep up. A single 26-line quote went back to a customer so fast the customer commented on it; the rep said 90 percent of it was AI.
The second-order effect matters more. The same team now spends its time on the complex, high-judgment problems that require an expert, which is where customer loyalty is actually won. That is the human advantage of AI: the knowledge stays with your people, and AI carries the weight that was burying it. Our post on enhancing customer experience with AI covers the broader landscape.
What this looks like in practice
The build starts with a diagnosis, not a tool. We map a real call from ring to resolution and time each step. Then we connect the systems the rep touches, which is usually the hard part, since the data lives in an aging database, a spreadsheet and a CRM. The AI document processing layer reads the drawings and PDFs that never made it into a structured system.
Then the assist goes live with a handful of reps, with feedback captured on every suggestion. The model learns from what the best rep corrects. Within weeks the suggestions get better, the lookups get faster, and the team asks for more. That is the moment to talk about automation, not before.
Keep the rep, lose the lookup
The gap between AI aspiration and AI that works in customer service is the gap between replacing a person and equipping one. Aspiration buys a chatbot and hopes customers do not notice. AI that works keeps the expert on the line and hands them everything they need while the customer is still talking. That is Operations Intelligence applied to the moment that matters most, and it is the diagnosis we run before anything gets built.
Your reps know your customers. Give them the answers faster, and put AI to work for your people on every call.
Frequently Asked Questions
What is AI agent assist in customer service?
AI that supports the human rep during a call or chat by pulling records, surfacing history, and drafting responses, while the person stays in control of the conversation.
Does agent assist replace customer service staff?
No. Research on more than 5,000 agents found it raised productivity 14 percent and helped new reps most, with no drop in satisfaction. It amplifies the team.
Where should we start?
Connect the data sources first, put suggestions in front of a small group of reps, and only automate routine cases after the assist has earned trust.
Will our customers know?
They will notice faster, more complete answers from the same people they already trust. That is the point.
Your people are not asking about models or tooling. They are asking whether they still have a job. In this employees afraid of AI podcast episode, two members of the Augusto team talk through why that fear is real and mostly unaddressed, what leaders should say (and actually mean), and why you need to speak before the rumor does. We close with what this looks like in practice when fear turns into adoption.
Ask a CEO what comes up first when they start an AI initiative, and it is rarely a technical question. It is this: I have people who are afraid. Employees who believe AI is coming for their job. Leaders who cannot afford to lose those people, and cannot afford to skip AI either. One executive put it to us this way: I do not have an option on using AI to transform, but it will be detrimental to lose people who think they are next.
That fear is rational, it is widespread, and it is the single biggest reason AI efforts stall after the pilot. This post is about the conversation that fixes it.
Silence is not neutral. When leaders do not explain what AI is for, employees fill in the blank with the worst headline they read that week. The same Gallup data shows the upside: when leadership communicates a clear plan, employees are three times as likely to feel prepared and 2.6 times as likely to feel comfortable using AI.
Your people are not resisting AI. They are resisting a story you have not told them yet.
What to say, and what to mean
The conversation only works if it is true. Here is the version we help leaders build, in the order that lands.
Name what AI is for: Be specific about the work it will carry. Not “efficiency.” Say: the three hours a day our engineers spend rebuilding quotes from scratch, or the invoice uploads to customer portals that nobody wants to do. People can accept a change they can picture.
Say what it is not for: If you are not planning headcount cuts, say so plainly and mean it. If natural attrition will change the shape of the team over time, say that too. The MIT NANDA report found that companies are seeing no evidence of mass layoffs tied to generative AI; the real shift is quieter and slower. Honesty about that beats a promise you cannot keep.
Show what changes for them: The team that gets AI-assisted quoting does not lose its experts. It gets to spend its time on the complex work that requires their judgment. A customer service leader we work with said it best after her team’s backlog fell from hundreds of open requests to around one hundred: it gives us the time to be the experts we are supposed to be.
Give them a role in the build: Fear drops when people shape the tool. Involve the people who do the work in choosing what AI should carry, and adoption stops being a change management problem.
Invest in them visibly: Training is the proof that you meant it. Our post on the hidden cost of not training your team makes the business case; the human case is that a trained employee stops looking over their shoulder.
The message underneath the message
There is a position behind this conversation that leaders have to choose. Some vendors sell AI as a way to remove bodies. We go the other way, and so do the companies we see winning: use AI to amplify the people who already carry the deep context of your business but are bogged down in repeatable tasks.
That is not a soft stance. It is where the returns are. A worker who knows your customers, your suppliers, and your exceptions is the asset. AI that frees that person from re-keying spreadsheets turns their judgment loose on the work that grows the company. The World Economic Forum’s Future of Jobs report projects both displacement and creation, and which side of that ledger your company lands on is a leadership decision.
What this looks like in practice
When a CEO tells us their people are afraid, the first thing we do is not a technology assessment. It is a session with the people doing the work, where they show us how things run today and tell us where the friction is. Then we build a demo from their own process. Teams that walk in defensive walk out asking what else is possible, because they have seen AI carry a task they hate and leave the part they are proud of alone.
That is why adoption is designed into the process rather than bolted on at the end. The people who will use the system help choose where it goes first, see it work early, and own the feedback loop. Our guide to escaping AI pilot purgatory covers the change management side in detail.
Say it before the rumor does
The gap between AI aspiration and AI that works is often a trust gap. Aspiration announces a tool. AI that works explains what the tool is for, what it is not for, and what changes for the person reading the email. That is HR Intelligence and Company Intelligence in the same breath, and it is the conversation we help leaders have before any system gets built.
Your people are waiting for the story. Tell it, keep it true, and put AI to work for your people rather than around them.
Frequently Asked Questions
How do I reassure employees about AI without lying to them?
Be specific about what work AI will carry, honest about how the team may change over time, and clear about what stays human. Specific beats reassuring.
Should we announce AI plans before we have a system?
Yes. Gallup found employees with a clearly communicated plan are three times as likely to feel prepared. Silence gets filled by headlines.
What if we do expect headcount to change?
Say that attrition, not layoffs, is the plan if that is true, and describe how roles will shift toward higher-value work.
How do we turn fear into adoption?
Involve the people who do the work in choosing where AI goes first, show them a demo built from their own process, and train them visibly.
A technology director at a large organization laid it out plainly on a call in August. His team was already using AI to write code. They had guardrails. They had automated testing. And they were still stuck, because the problem was not the code. It was the product owner, the backlog, and acceptance criteria that nobody could agree on. All the other sectors, he said, are moving at 200 miles an hour, and we are going at ten.
That is the honest state of AI in software development at most mid-market companies. The typing got faster. The system around the typing did not. This post is about the system.
Put those together and you get a lifecycle where one stage sped up and every downstream stage absorbed the cost. Ambiguous requirements become confident, wrong code faster than ever, which is the productivity tax of almost-right code. Review queues lengthen. Testing finds the problems that a clear acceptance criterion would have prevented.
AI made the cheapest part of software development cheaper. The expensive parts, deciding what to build and proving it works, are still yours.
Where AI actually helps across the lifecycle
The leverage is upstream and downstream of the code, not only in it. Here is where we see it pay off.
Backlog grooming: AI can turn a messy intake of requests, tickets and meeting notes into structured stories with a consistent shape. The product owner still decides priority, but they start from a clean draft instead of a blank page.
Acceptance criteria: This is the biggest win we see. When a story includes explicit, testable criteria before development starts, AI-generated code has something to be checked against. Spec-driven development, where the specification is the source of truth, is how the fastest teams are working now.
Test generation: AI writes test cases from acceptance criteria quickly. That turns a criterion into a gate instead of an aspiration.
Documentation and handoff: The least loved work in engineering is the easiest to delegate, and it is what protects you when a developer leaves.
The part that does not change
Ownership does not change. Someone has to be the master of what the system needs to do. A CEO we spoke with this summer was rebuilding a core platform piece by piece and asked the right question: is my team doing it right? His architects were senior. His developers were productive. What he wanted was a check on the process, from spec to acceptance, because he understood that the tools would only amplify whatever process was in place.
Judgment does not change either. Deciding when AI can make a call and when a person must is a framework, not an instinct. Our post on when AI should make the call applies to engineering decisions as much as business ones.
What this looks like in practice
The pattern across our engineering work is consistent. We start with a short diagnosis of how work flows from request to release, because nine times out of ten the slow step is not where the team thinks it is. Then we redesign the stages where AI carries the weight: intake, story drafting, acceptance criteria, test generation, documentation. Developers keep their judgment and their code review. AI carries the repetitive load.
For one team, adoption was the barrier, not capability. Some engineers did not want to use AI tools to build code, for reasons that ranged from pride to fair concern about quality. The fix was not a mandate. It was a demo built from their own backlog, showing what changed for them, followed by a process that made the safe path the easy path. We wrote about that dynamic in how to get your team to actually use AI, and the same lessons hold for engineers.
McKinsey’s research on AI in software development points the same direction: the largest gains come when teams redesign the workflow around the tools, not when they bolt tools onto the old workflow. The state of AI research is clear that workflow redesign is what separates value from experimentation.
Fix the lifecycle, then add the speed
The gap between AI aspiration and AI that works in engineering is the gap between faster typing and faster delivery. Aspiration measures lines of code. AI that works measures how long it takes a clear requirement to become a tested feature in production, and it redesigns every stage in between. That is Engineering Intelligence, and it is the diagnosis we run before we touch a toolchain.
Your developers already have the assistants. Give them the specs, the criteria and the gates, and put AI to work for your people across the whole lifecycle.
Frequently Asked Questions
Where does AI help most in software development?
Upstream and downstream of the code: turning requests into structured stories, drafting testable acceptance criteria, generating tests, supporting review, and writing documentation.
Why is our team not faster even though we use AI coding tools?
Because AI sped up one stage. Ambiguous requirements, slow review and weak testing absorb the gain. Fix the lifecycle first.
What is spec-driven development?
A practice where a clear, testable specification is written before code and treated as the source of truth, so AI-generated code has something to be checked against.
Do we still need code review with AI?
Yes. Independent testing finds security flaws in nearly half of AI-generated code. Review and testing are what make the speed safe.
We are heading toward more agents than employees, and somebody has to answer for them. In this AI agent org structure podcast episode, two members of the Augusto team talk through who owns an agent’s output when it gets something wrong, why every agent needs a written job description, how a supervising layer watches the fleet, and why verification has to be designed in rather than bolted on. We close on the practical starting point: one function and a handful of agents under one owner, not forty agents by spring.
Deploying agents faster than you are assigning owners? Reach out. We start every engagement with a fixed-price, two-week diagnostic that maps your agents to the people who will answer for them.
At an executive roundtable we hosted in August, one leader was partway through writing his company’s first AI policy when he hit a question the template could not answer: when you deploy AI agents, who is ultimately responsible for their output? Is it IT? The customer service leader? The sales leader? Then he said the line that reframed the whole session: we are heading down a path where we will have more agents than employees, and somebody has to answer for them.
Another story from the same table made it concrete. At one company, a shop floor worker called the COO on a Saturday night to ask who Phil was, and how Phil knew his shipment went out late. Phil was the company’s new AI agent, quietly watching daily shipments and alerting customers about delays. Weird and cool at the same time, as the table put it. Then the harder question landed: but how do you manage that?
That is the real subject here. Not whether agents work. They do. The question is what your org chart looks like when some of the workers are software.
Why this question got urgent this year
Agents stopped being a demo and started being coworkers. Active agents in the Microsoft 365 ecosystem alone have grown roughly 15x year over year, according toMicrosoft data compiled by Optro, far outpacing the governance frameworks built for chat-style AI. One roundtable participant plans to deploy 40 agents across all of his company’s locations in the next six months. Meanwhile most leadership teams are still, in the words of another participant, taking a Wild West approach to applying AI. That gap between deployment speed and organizational design is where the expensive mistakes happen.
We are heading toward more agents than employees. The org chart has to answer for them.
Five principles that survived the debate
The roundtable did not agree on everything, but five principles held up under pressure:
Every agent rolls up to a human: Not to IT by default, but to the human who owns that function of the business. An agent quoting jobs belongs to the sales leader. An agent watching shipments belongs to operations. The accountability trail should exist before the agent does, because when an agent kicks out an inaccurate quote and money is lost, “where does accountability go” is not a question you want to be answering for the first time.
Agents get job descriptions: One participant asked, half joking, whether there is an HR component here, and whether agents get job descriptions. The serious answer is yes. A written scope covering what the agent does, what data it can touch, and what it is never allowed to do is the boundary-setting the whole table was reaching for.
Someone supervises the fleet: The most developed model in the room was a lead agent that knows the business, with worker agents reporting to it so accountability, permissions, and access stay appropriate. One leader called it a chief of staff for agents: the agents work for individual people, but the chief of staff watches all of them.
Verification is designed in, not bolted on: The simple rule is that AI output gets verified before it ships. The honest complication raised at the table is that you use AI precisely because it processes more than a person can, so define verification per role: spot checks, thresholds that require human sign-off, and evaluations that watch the agents’ quality over time.
Guard what the agents learn: Agents absorb how your company works from meetings, documents, and systems. One sharp warning from the session: if the loudest voice keeps saying “this is the way I do it,” the system will decide that is how it is done. Somebody has to curate the context intentionally, or you memorialize the wrong process at machine speed.
The org chart, in practice
Put those principles together and the structure has three layers. At the top, human function owners: the same leaders on your org chart today, now each accountable for the agents in their function. In the middle, an orchestration layer: the chief-of-staff pattern, whether that is a supervising agent, a platform, or a named person, that monitors permissions, performance, and boundaries across the fleet. At the bottom, worker agents with written job descriptions, scoped data access, and a clear human owner.
Notice what this is not. It is not a new AI department owning every agent, which recreates the bottleneck agents were meant to remove, and it is not agents scattered wherever enthusiasm lives. If you are still deciding which work belongs with agents at all, our breakdown ofdeterministic and non-deterministic agent workflows and our framework onwhen AI should make the call pair well with this post.
Start with one function, not forty agents
Across our client work, the pattern that succeeds starts narrow. One organization we partner with made coordination the first rule, because they did not want three different people quietly building the same agent. Another is expanding a knowledge tool from a single user to a multi-user, permission-based system, growing the structure with the fleet instead of after it. And in one rollout, accountability moved to managers rather than individuals, and participation in the supporting training jumped significantly. Structure first, scale second.
The gap between AI aspiration and AI that works shows up here as clearly as anywhere. Aspiration is forty agents by spring. AI that works is forty agents that each have a human owner, a job description, a supervisor, and a verification path. That is Company Intelligence in the most literal sense: your company’s structure, extended to its newest workers.
Diagnose one function, give its agents a real place in the organization, and stand behind them in production. Put AI to work for your people, and give it a manager while you are at it.
Frequently Asked Questions
What is an AI agent org structure?
The set of ownership, supervision, and accountability rules that place AI agents inside your organization: which human owns each agent, what each agent is scoped to do, and who monitors the fleet.
Who should be responsible for an AI agent’s output?
The human leader who owns that business function, not IT by default. IT typically owns the platform and permissions; the function owner answers for what the agent produces.
Do AI agents really need job descriptions?
Yes. A short written scope covering purpose, data access, boundaries, and escalation rules is the single most effective guardrail, and it makes supervision possible.
How many agents should we start with?
One function and a handful of agents under one owner. Prove the ownership and verification model there, then scale.
A new AI model every week. Which ones actually change how you run your business? In this LLM news October 2026 podcast episode, two members of the Augusto team talk through OpenAI shipping GPT-6 Astra and then scrapping its successor, frontier-class intelligence arriving at mid-tier prices, and the agent era arriving with no referee. We close on our four moves for October, starting with keeping the model swappable.
Stuck with one vendor hard-wired into your workflows and a pricing page that keeps changing? Reach out. We start every engagement with a fixed-price, two-week diagnostic that finds which workflows should move first.
September was the month the benched model came back, the price war moved into the middle of the market, and Google finally answered. One manufacturing leader we work with summed up how most executives feel: the problem is not a lack of tools, it is a new one every week. Here is your LLM news roundup for October 2026, and what each story means for how you run your business.
OpenAI shipped Astra, then scrapped its successor
The model OpenAI paused in August is now on sale. GPT-6 Astra launched on September 3 at $10 per million input tokens and $50 per million output tokens, several times the price of its mid-tier siblings. It finishes long tasks about 47 percent faster than GPT-5.6, but it remains the first OpenAI model rated Critical for cybersecurity, so its offensive security features are gated to a vetted program and it ships off by default.
Days before DevDay, OpenAI cancelled GPT-6.1 Astra after internal testing raised safety and alignment concerns, and shipped GPT-6.1 Sol in its place.
The most capable model you can buy today is also the one its maker trusts least. Plan for that.
What it means for your business: the frontier is now a premium track with a safety governor on it. Astra-class pricing only pays for itself where a mistake is expensive and a human is slow. For everything else, the next story matters more.
Frontier-class intelligence now costs mid-tier money
The real September story was the middle of the market. Anthropic released Claude Opus 5.5 on September 22 at $4 and $20 per million tokens, then Claude Sonnet 5.5 six days later at $2 and $10, running 30 percent faster and landing two points behind Opus on occupational benchmarks. OpenAI priced GPT-6 Sol and Luna at half their GPT-5.6 equivalents. Google launched Gemini 3.8 Flash at $0.75 and $3.75, and Meta’s Muse Spark 1.3 arrived at $1.25 and $4.25, with a Contributor tier under $0.20 for teams willing to let Meta train on their prompts.
Read the fine print twice. Gemini 3.8 Flash’s rate doubles on January 1, 2027, and Meta’s discount is paid for with your data. Flagship performance now sells at workhorse prices, but the price tag is a moving target.
What it means for your business: a multi-location business leader asked us earlier this year whether choosing an LLM was a one-time bet. It is not. Client teams that treat the model as a swappable part are re-pointing routine workloads at Sonnet-class and Flash-class models this month and banking the difference. Teams that hard-wired one vendor are reading pricing pages instead.
Google answered with Gemini 4 Argon, and defenders get it first
On September 30, Google released Gemini 4 Argon, calling it its most powerful model and claiming it scores above GPT-6 Astra and Anthropic’s Fable and Opus models on independent indexes. Like Astra, it is built to find, validate, and patch software vulnerabilities on its own, and like Astra it is not for you yet. Access goes first to cyber defenders in Google’s Fairwind program, with wider API rollout promised “as soon as possible.” A month after its leadership shakeup, Google has a frontier model to show for it.
What it means for your business: every frontier lab now ships its best model to security teams first, which tells you where they think the risk is. Your defenses should assume AI-assisted attackers. Our guide to AI governance for executives covers the controls that matter most.
DevDay made agents the product, and Washington watched
OpenAI’s DevDay on September 29 launched more than 20 products, and the theme was delegation. Dots are always-on agents that work in the background across thousands of apps, ChatGPT Space lets teams and their agents manage projects together, and Codex now runs fully in the cloud. The quiet headline for executives was the Decisions API, which lets a company hand narrow, repetitive decisions such as approve or reject to AI within a predefined set of outcomes. ChatGPT, meanwhile, reached 1.2 billion weekly users.
The same day in Washington, the administration hosted a “Golden Age” event celebrating AI while lawmakers floated kill-switch legislation and committed to nothing.
What it means for your business: the agent era arrived with no referee. Writer’s enterprise survey, updated in September, found that only 23 percent of companies see significant ROI from AI agents and 36 percent have no supervision plan for them. One staffing-industry leader we work with asked the right question before deploying anything: where does the data live? Our answer is tiers. Private models for the crown jewels, frontier models for everything else, and a routing layer so no single vendor outage, pause, or price change becomes your outage.
What to watch in October
Gemini 4 Argon should reach API customers and AI Ultra subscribers, and Anthropic has signaled a new Haiku is close, which would reset the floor on cheap, fast models again. GPT-6.1 Astra is not coming this month, while Meta’s rumored flagship and consumer agent platform remain unconfirmed. Launch discounts on ElevenLabs v4 and Upstage’s Solar Mini 4 expire on October 12 and 22, and Xiaomi deprecates its MiMo-V2.5 models on October 21.
Four moves for October, with owners
Re-tier your workloads (CTO and CFO, 30 days): Test Sonnet 5.5, GPT-6 Sol, and Gemini 3.8 Flash on your highest-volume tasks. Reserve Astra-class spend for work where errors are expensive.
Diary every expiry (CFO, this week): Note January 1 for Gemini pricing and October 12 and 22 for launch discounts.
Write the agent supervision plan (COO, 60 days): Decide which decisions an agent may make alone, which need a human, and how you will know when it is wrong.
Keep the model swappable (CIO, this quarter): If replacing your primary model would take a rewrite, that is the project. One IT team we work with described life after that work: they move fast now, without the fear of kissing frogs and hitting dead ends.
The pattern from September is the one we see in the field every week: capability jumps, prices fall, and even the labs cannot predict what they will ship next quarter. The companies winning did not pick the right model. They built so the model could change. That is the work of an AI activation partner: diagnose where the energy is going, build the solution, and stand behind it in production. If you want to know which of your workflows should move first, a two-week Rumble answers that with a prioritized roadmap and a quick win ready to build. Put AI to work for your people, and keep the plumbing yours.
Frequently Asked Questions
What was the biggest LLM news in September 2026?
OpenAI released GPT-6 Astra and then cancelled its successor over safety concerns, Anthropic shipped Opus 5.5 and Sonnet 5.5, and Google launched Gemini 4 Argon to cyber defenders first.
Which new AI model should a mid-market company use?
Most routine work now runs well on mid-tier models such as Claude Sonnet 5.5, GPT-6 Sol, or Gemini 3.8 Flash at a fraction of frontier prices. Reserve premium models for high-stakes tasks, and build so you can switch.
Why are frontier models going to security teams first?
GPT-6 Astra and Gemini 4 Argon can find and exploit software vulnerabilities on their own, so both labs are gating those capabilities to vetted defenders first.
What should executives do about AI agents right now?
Define which decisions agents may take alone, assign a human owner for the rest, and put monitoring in place before deploying.
