On-Prem vs. Cloud in an Agentic World

/by

Episode 03: Listen to this post as a podcast (22 min)

Press play to listen here, or watch it on YouTube. Open the episode page to share it or listen later.

At an executive roundtable we hosted in August, a leader who had spent years moving his company into the cloud said something that quieted the room: we have moved a lot of things to the cloud over the years, but that does not mean it is safe. As his team turns AI loose on those systems, everything that quietly accumulated there behaves like the old shared S drive, where people put things wherever they wanted. It is not just what you do with AI next, he said. It is what has already been done. As you point AI at that legacy, it is going to get exposed.

That exchange is why this post exists. The on-prem versus cloud debate is a decade old, but agents just changed the question underneath it.

Agents changed the question

Chat-style AI reads. Agents act. They retrieve documents, touch customer records, call tools, and produce new outputs, which means the sovereignty question is no longer only where your data is stored, but where reasoning, retrieval, and action are allowed to execute.

The roundtable produced a perfect example. One company asked us about letting an AI agent loose on their CRM to update records, until someone did the math out loud: that is 5,000 customer records the agent could touch, and one bad run could quietly change any of them. What safeguards exist before we say go? Another leader worried about the reverse flow: you are feeding these models with prompts, files, and context about your business, and at some level they are learning from you. A third raised the insider question that on-prem alone does not solve: once your company’s knowledge is easily searchable by AI, how do you detect the one person who downloads all of it?

The question is no longer where your data sits. It is where your AI is allowed to act.

What the market is actually doing

The numbers back the room’s instincts. According to Broadcom’s Private Cloud Outlook 2026, the share of enterprises running production AI inference primarily in the public cloud fell from 56 percent to 41 percent in a single year, with a majority now running or planning production inference on private infrastructure. This is not a wholesale retreat from cloud. It is workload-by-workload sorting, the same rebalancing we wrote about in why more organizations are bringing workloads home, now accelerated because agents raise the stakes of getting placement wrong.

And the honest counterpoint still applies: aging local infrastructure can throttle AI just as surely as a bad cloud contract, a case we made in your on-prem servers are the real bottleneck to AI. Agents do not settle that argument for one side. They force a sharper sorting rule.

Three questions that decide where an agent lives

  1. Blast radius: What can this agent touch, and what is the worst thing it could do there? An agent that drafts marketing copy can live anywhere. An agent with write access to 5,000 CRM records, your ERP, or your financials belongs where you control permissions, logging, and the kill switch, whatever infrastructure that is.
  2. Context sensitivity: Agents run on context, and context is your business: pricing logic, customer histories, tribal knowledge. If the prompts and files an agent consumes would hurt you in a competitor’s hands, keep that context in an environment with contractual no-training guarantees, or on infrastructure you own. Public tools on personal accounts fail this test automatically.
  3. Economics and rhythm: Agent inference is a steady, predictable, always-on workload, exactly the profile where owned or private capacity amortizes well, while burst experimentation still favors cloud flexibility. Watch the pricing fine print too, because promotional API rates now expire on a schedule, and an always-on agent fleet feels a price change immediately.

One more honest note from the roundtable: a participant experimenting with a local agent platform on a spare desktop admitted the learning curve is real. Plan on serious hours before a self-hosted agent stack feels routine, which is exactly why the environment decision should follow the workload, not the hobbyist enthusiasm.

What this looks like in practice

The pattern across our client work is consistently hybrid. One manufacturer chose a custom AI build inside its existing infrastructure specifically to answer privacy concerns, keeping sensitive data in the environment it already trusted while still getting frontier capability. Another organization is expanding a locally hosted AI knowledge tool from one user to a multi-user, permission-based system, growing governance with the footprint. And one technology leader told us plainly that his team’s recent AI projects all run locally because the data was not suitable for an environment they could not verify.

None of them picked a side in the on-prem versus cloud debate. They sorted workloads: sensitive context and high-blast-radius agents on infrastructure they control, commodity tasks on the cheapest reliable API, and a routing layer in between so no single provider decision is permanent. If you are earlier in that journey, our guides on running LLMs locally and choosing a cloud LLM provider cover each half of the map.

Sort the workloads, not the ideology

The gap between AI aspiration and AI that works often comes down to this: aspiration argues about cloud versus on-prem, while AI that works asks what each agent touches, what it consumes, and what it costs to run all day. That is IT Intelligence doing its job, and it is the diagnosis we run before anything gets built: map the agents you want, score their blast radius and context sensitivity, and place each one deliberately.

Your agents are coming either way. Decide where they live before they decide for you, and put AI to work for your people on ground you control.

Frequently Asked Questions

Should AI agents run on-prem or in the cloud?

Neither by default. Sort each agent by what it can touch, how sensitive its context is, and its usage pattern. High-blast-radius, sensitive-context, always-on agents favor private infrastructure; low-risk burst work favors cloud APIs.

Why do agents change the on-prem vs cloud decision?

Because agents act rather than just read. The question shifts from where data is stored to where AI is permitted to retrieve, reason, and execute, and what it could do wrong there.

Is moving AI on-premises actually a trend?

Yes. Broadcom’s 2026 survey shows public cloud falling from 56 to 41 percent as the primary home for production AI inference, with private infrastructure growing, though most enterprises are landing on hybrid.

What should we do before deploying any agent?

Give it a written scope, least-privilege access, logging, and a kill switch, then choose its environment based on blast radius and context sensitivity, not habit.

AI Document Processing: Freeing the Data Trapped in Your Paperwork

/by

Episode 02: Listen to this post as a podcast (27 min)

Press play to listen here, or watch it on YouTube. Open the episode page to share it or listen later.

When we run AI workshops with operations teams, we always ask the same question: where do the hours actually go? The answer is nearly always a document. One leader had counted precisely: 422 purchase order and receiving entries a month, each one a person reading a document and retyping it into a system. Another described an accounting team working through literal boxes of paper invoices with no way to search any of it. A third walked us through drop-ship orders, twenty to thirty a day, processed entirely by hand.

None of these are technology problems anymore. They are decision problems, because AI document processing has quietly crossed the threshold from brittle scanning tool to reliable operations layer, and most mid-market teams have not re-run the math since it did.

What changed: from reading text to understanding documents

Old-school OCR read characters and broke the moment a vendor changed their invoice layout. Modern AI document processing reads the way a person does: it understands context, handles formats it has never seen, interprets tables and handwriting, and extracts meaning rather than pixels. In one of our own client builds, an extraction engine now pulls product, quantity, urgency, and request type directly from ordinary inbound customer emails, the messiest document type there is, and turns them into structured entries a system can act on.

One honest lesson from the field belongs here. At an executive roundtable we hosted in August, a leader described an AI quoting system whose accuracy lagged until a human recognized what every experienced operator already knew: the requests were full of ancillary products no system had a category for. Extraction was never the problem. Meaning was. The best document processing implementations pair the AI with the people who know what the documents actually contain.

Your business already captured this data once, when a customer sent it. Every retyping after that is a tax.

Where the hours are hiding

Across mid-market operations, four document flows consistently hold the largest reclaimable hours:

  1. Inbound requests and orders: Emails, attachments, and forms carrying orders and quote requests. AI extracts the structured facts, so people confirm instead of transcribe. This is where the twenty-to-thirty-orders-a-day team wins back a role’s worth of time.
  2. Purchase orders and receiving: The counted example above, 422 entries a month, is typical, not extreme. PO-to-system extraction is mature, high-volume, and easy to measure.
  3. Accounts payable invoices: The classic for a reason. Industry benchmarks show invoice handling falling from roughly 15 minutes to under 2 with AI, and it is where those boxes of paper finally become searchable data.
  4. Quotes, specs, and contracts: Longer documents where AI extracts terms, line items, and obligations for human review, so the expert reads a summary with sources instead of hunting through pages.

The market has noticed: 72 percent of enterprises are investing in AI document automation in 2026, and it remains one of the most directly measurable AI investments a company can make, in hours, errors, and cycle time.

The accuracy trap, and the loop that beats it

The number one buying mistake is comparing AI to perfection instead of to your current process, which also makes errors, silently. Modern extraction platforms reach 94 to 98 percent field accuracy, and above 99 percent once validation and human review are added, while automating the large majority of documents straight through. The design that gets you there is confidence-based routing: the AI processes what it is sure about, flags what it is not, and a person handles only the exceptions. That review step is a feature, not a failure, and building it well is exactly what our guide to human review workflows that scale covers. Feed the corrections back and accuracy compounds, especially when your experts encode what they know, like the missing product categories from the quoting story.

Start with one document type and count first

The rollout discipline is the same one behind all our real-world AI workflow examples: pick one document flow, count its true volume the way the 422 leader did, and run the AI against your actual documents rather than a vendor demo. Prove the accuracy on your worst samples, wire the exceptions to a named human, retire the retyping, and only then move to the second document type. If your processes are still half paper and half spreadsheet, sequence it with our guide to automating manual processes without breaking what works.

The counting step matters more than it sounds. Every operations team knows the work is there; almost none can say the number. The leader who could say 422 had, in that moment, already built the business case.

The paperwork was never the job

The gap between AI aspiration and AI that works shows up vividly here. Aspiration buys a scanning tool and digitizes the pile. AI that works asks what decisions the documents feed, extracts the data once at the point of arrival, routes exceptions to people, and lets the team do the job the paperwork was interrupting. That is Operations Intelligence: your documents becoming inputs instead of chores.

Count your 422, whatever your number is, and put AI to work for your people, starting with the pile on their desks.

Frequently Asked Questions

What is AI document processing?

The use of AI to read, classify, extract, and validate data from business documents, emails, POs, invoices, quotes, contracts, and push it into your systems as structured, actionable data.

How is it different from OCR?

OCR reads characters and breaks on new layouts. AI document processing understands context and meaning, handles formats it has never seen, and extracts fields, tables, and intent, not just text.

How accurate is AI document processing?

Modern platforms hit 94 to 98 percent field accuracy, and over 99 percent with validation and human review of flagged exceptions. The right comparison is your current manual error rate, not perfection.

Where should a company start?

One high-volume document flow, usually inbound orders, POs, or AP invoices. Count the real monthly volume, test on your actual documents, route exceptions to a named person, and expand one flow at a time.

AI Change Management: How to Escape Pilot Purgatory

/by

Episode 01: Listen to this post as a podcast (19 min)

Press play to listen here, or watch it on YouTube. Open the episode page to share it or listen later.

In the AI workshops we run with mid-market leadership teams, the same confession keeps surfacing. The pilot worked. The demo impressed the board. Somebody’s team is genuinely faster. And a year later, that is still the whole story: one team, one workflow, one win that never spread. One operations leader put the underlying problem plainly in a session with us: we don’t want three different people quietly building the same agent, so how do we coordinate our efforts to scale this across the company?

That is the AI change management problem, and it is different from the adoption problem. We have already written about getting a team to actually use AI, the fears, the champions, the path of least resistance. This post is about the next wall: turning one team’s win into a changed organization.

Pilot purgatory is the default outcome

The numbers say stalling is normal, not exceptional. McKinsey finds that while 88 percent of organizations now use AI somewhere, nearly two-thirds are still running pilots rather than scaling, and only 1 percent describe AI as fully integrated into how work gets done. And when Prosci studied over a thousand AI implementations, people-side difficulties outnumbered purely technical ones by more than two to one.

Read those together and the diagnosis writes itself: the technology clears the bar, and the organization does not change around it. Scaling stalls not because the second team’s problem is harder, but because nobody owns the change between teams.

Pilots prove the technology. Change management proves the company. Most AI programs only ever run the first test.

The chasm runs through the middle of your org chart

The same research points at where scaling breaks. Surveys show a 52-point trust gap between executives and workers on AI: 61 percent of executives trust it for complex decisions, against 9 percent of frontline workers. Executives see the strategy; the front line sees the risk. The only people positioned to close that gap are the managers in between, and most AI programs skip them entirely, training executives and end users while the middle layer inherits the disruption with no mandate.

Our client work keeps proving the opposite approach. In one rollout, AI training participation was stuck until the metrics moved from individuals to managers, who were now answerable for their whole team’s completion. Participation jumped almost immediately. At another client, a single supply chain leader who took the training seriously began cascading it to his own team unprompted, and his department’s habits changed in weeks. Same tools, same content. The difference was a manager who owned the change.

Scale the process change, not the tool

Here is the trap inside most scaling plans: they roll out the tool to more people, when the pilot’s real lesson was a changed process. A pilot team does not just use AI, it quietly rewrites how work flows, what gets checked, and what gets skipped. Scaling means exporting that rewritten process, which is why the coordination question from our workshop matters so much. Without a standard, you get three versions of the same agent, three conflicting processes, and a governance mess that our AI governance guide for executives exists to prevent.

The discipline that works is boring and sequential: document what the pilot team actually changed, standardize it, retire the old way of doing that work so the workaround dies, and only then move to the next workflow. Companies that skip the retirement step run both processes forever and call it transformation.

A 90-day change cadence that escapes purgatory

  1. Name the second workflow and its owner: Not a committee. One process, one accountable leader, chosen because the pilot’s lessons transfer, and scoped like the AI quick wins that pay back within 90 days.
  2. Make managers the unit of change: Give every affected manager a metric they own for their team’s transition, training completion, process cutover, and outcome. Accountability at the manager level is the single highest-leverage move we have seen in the field.
  3. Standardize before you multiply: Publish the pilot’s rewritten process as the standard, including what AI does, what humans verify, and what is retired. Coordination beats enthusiasm.
  4. Measure outcomes, not logins: Adoption dashboards full of active-user counts are how pilots die politely. Track cycle time, error rates, and capacity freed, the numbers a CFO recognizes.

Run that loop once per quarter and the math compounds: four changed workflows a year, each with a manager who owns it and a standard the next team inherits.

The payoff is capacity, not novelty

One CEO we work with described the end state better than any framework: the old constraint was how fast you could hire and what the budget allowed, and now his team can do more with what they have, getting further ahead and more competitive without waiting on headcount. That is what escaping pilot purgatory buys, and it is Company Intelligence in action: the organization itself, not one team, operating differently.

The gap between AI aspiration and AI that works is exactly the gap between a great pilot and a changed company. Diagnose the workflow, put a manager’s name on the change, standardize what worked, and retire what it replaced. Put AI to work for your people, all of them, one owned workflow at a time.

Frequently Asked Questions

What is AI change management?

The discipline of moving an organization from successful AI pilots to changed company-wide operations: standardized processes, manager-owned transitions, retired legacy workflows, and outcome-based measurement.

Why do most AI initiatives stall after the pilot?

Because pilots test technology while scaling tests the organization. With most difficulties on the people side and no one owning the change between teams, two-thirds of companies stay stuck in pilots.

Who should own AI change management?

Managers, not just executives or end users. Making managers accountable for their team’s transition is the highest-leverage move, because they sit exactly where the executive-to-frontline trust gap is widest.

How long should scaling one workflow take?

About a quarter: name the workflow and owner, train through managers, standardize the new process, retire the old one, and measure outcomes. Then repeat.

How to Write an AI Acceptable Use Policy Employees Will Actually Follow

/by

An IT leader at a mid-market manufacturer told us he was racing to get his company’s first AI policy to a working version within the week. He had a borrowed template from his managed service provider open in one window and a blank page in the other. Then his infrastructure partner asked a simple question: do you have any data privacy requirements around AI, like no training on your data, no retention, no PII? His answer was honest. Not yet.

That moment plays out in mid-market companies everywhere. Your people adopted AI first, and the rules are trying to catch up. An AI acceptable use policy is how you close that gap, but only if you write one that people can actually follow.

Why most AI policies fail before they launch

Two failure modes kill most first attempts. The first is silence. At a town hall we heard about, an employee asked leadership directly what the company policy on AI was, and the answer was that one did not exist yet. In one of our operations workshops, a leader admitted the same thing plainly: no governance policy, nothing. Employees in that vacuum make up their own rules, and the most common question they carry is whether it is even okay to use AI at all.

The second failure mode is overcorrection. One leader we work with borrowed a policy from a friend at a cybersecurity firm, and it was so restrictive it effectively banned AI outright. A policy that reads like a ban does not stop AI use. It just stops honest AI use, and pushes everything underground where you cannot see it.

A policy that reads like a ban does not stop AI use. It just stops honest AI use.

The numbers say most companies are still stuck between those two failures. Only about 15 percent of organizations have updated their acceptable use policies to cover AI specifically, according to research compiled by JumpCloud, even as unsanctioned AI use shows up in nearly every workforce.

Decide what your policy is for

At a recent executive roundtable we hosted, the sharpest question of the session was deceptively simple: is this policy a structure for the organization to operate with AI, or a defensive legal and HR document? The answer shapes everything. A defensive policy protects the company from its people. An operating policy protects the company through its people, by giving them a safe, clear way to work.

Write the operating version. It should answer, in plain language, the questions your employees are already asking: which tools are approved, what data can go where, and what to do when they want something new.

The six sections your AI acceptable use policy needs

Keep it to one page if you can. Here is the structure we help clients build:

  1. Approved tools: Name the sanctioned tools and accounts. If AI use must run through IT approval on enterprise licenses rather than personal accounts, say exactly that.
  2. Data rules: Spell out what can never enter an AI tool, such as customer records, financials, and regulated data, and where protected work can happen safely.
  3. Verification duty: If you get output from AI, you verify it before it ships. Define what verification means for high-stakes work like quotes, contracts, and customer communication.
  4. Accountability: Every AI workflow and agent rolls up to a named human owner. When an agent produces a bad quote, the accountability trail should already exist.
  5. The request path: Give people a fast, judgment-free way to propose new tools. This turns shadow AI into a pipeline instead of a threat.
  6. Ownership and review: Name who signs the policy, who manages it, and when it gets revisited. A policy nobody owns goes stale in a quarter.

Roll it out like you mean it

Publishing a PDF is not a rollout. Two moves make the difference. First, put adoption metrics on managers, not just individuals. In one client rollout, training participation jumped significantly the moment completion became a management team metric rather than an individual one. Second, pair the policy with real training so people understand the why behind each rule; the gap between publishing rules and building skills is covered well in our piece on the hidden cost of not training your team on AI.

Your policy also does not live alone. It is the front door of a broader governance posture, and our guide on what every executive needs to know about AI governance covers the structures behind it. And once rules exist, adoption becomes the next battle, so keep getting your team to actually use approved AI on your radar.

From policy to activation

The gap between AI aspiration and AI that works is rarely about the technology. It is about whether your people know the safe path and trust it enough to take it. Across our client work we have seen what happens when they do: leaders standardizing on governed platforms so company information stops flowing through personal accounts, teams coordinating so three people are not quietly building the same agent, and manager-led rollouts that measurably lifted training completion.

That is Company Intelligence working alongside Security Intelligence: rules that amplify your people instead of restraining them. Diagnose where AI is already carrying weight in your business, write the one-page policy that makes it safe, and stand behind it in production.

Put AI to work for your people, with rules they can actually follow.

Frequently Asked Questions

What is an AI acceptable use policy?

A short document that tells employees which AI tools are approved, what data can and cannot enter them, how AI output must be verified, and how to request new tools.

How long should an AI acceptable use policy be?

One page is the goal. A forty-page policy nobody reads is worse than a one-page standard everyone follows.

Who should own the AI acceptable use policy?

Assign a named owner, typically IT or security leadership with HR input, plus a review cadence. Every AI workflow in the policy should also roll up to a named human owner.

Should the policy ban public AI tools?

Usually not. Over-restrictive policies push AI use underground. Provide approved alternatives that are genuinely better, then restrict only the highest-risk behavior.

Shadow AI: Your Team Is Already Using AI. You Just Don’t Know Where

/by

A technology leader at a mid-market services company put it bluntly in a working session with us: everybody was doing their own thing with AI. People were pasting company spreadsheets into free personal ChatGPT accounts, and that data was leaving the building for good. As someone in the room quipped, the public models were getting free training content, courtesy of the company.

That is shadow AI. Your people are not waiting for permission. They have deadlines, and a free AI tool is one browser tab away. The question is not whether unsanctioned AI use is happening in your company. It is how much, with what data, and what you will do about it.

What shadow AI actually is

Shadow AI is any AI tool your employees use for work without approval, oversight, or protection from IT. It is the AI-era version of shadow IT, with one difference: these tools actively process your data at the point of use. A personal ChatGPT account, a free transcription app, a browser extension summarizing contracts. Each one is a door your data walks through, and most lead outside your control.

The scale is bigger than most executives expect. Verizon’s 2026 Data Breach Investigations Report found shadow AI detections quadrupled in a single year, with 45 percent of employees now using AI regularly on corporate devices and roughly two thirds accessing AI through personal accounts that company controls never see.

Why your best people are the biggest users

Here is the uncomfortable part: shadow AI is not a discipline problem. It is a demand signal. The employees using unsanctioned tools are usually your most motivated people, trying to move faster than your official systems allow.

 

We see this pattern constantly in mid-market companies. At one executive AI session we ran, an employee had asked leadership at a town hall what the company policy on AI was. The honest answer: we have not come up with one yet. In another workshop, the most common fear in the room was simpler than job loss. Is it okay to use AI? If I put something in here, will it end up on the internet? Will my competitor get it?

When leaders leave those questions unanswered, employees answer them alone. Experimentation goes underground, and one operations leader we work with described the result perfectly: a Wild West approach to applying AI. The energy is real, but nobody is protecting the data, checking the output, or capturing what works.

The risks are specific, not hypothetical

The costs now show up in breach data, not just policy debates. IBM’s research links shadow AI to roughly one in five data breaches, adding about $670,000 to the average breach cost. Three risks matter most for mid-market leaders:

  1. Data leakage: Customer records, pricing, financials, and source code pasted into free tools may be retained, and you cannot pull them back.
  2. Compliance exposure: Regulated data moving through consumer AI tools creates violations your team never intended and your auditors will eventually find.
  3. Lost leverage: When fifty people solve the same problem fifty different ways in fifty personal accounts, nothing compounds. One construction industry leader told us they did not want three different people building the same agent. Coordination is where the value is.

Bring it into the light without killing the momentum

Banning AI does not work. It just pushes usage deeper underground and punishes your most ambitious people. The companies handling this well treat shadow AI as a map of unmet needs, then replace risky tools with better sanctioned ones. Here is the path we walk with clients:

 

  1. Start with an amnesty audit: Ask every team what AI tools they already use and what for. Treat the answers as intelligence, not evidence. This single step usually surfaces your best AI use cases for free.
  2. Give people a safe place to work: One manufacturer we partner with had real privacy concerns about public tools, so the answer was AI built inside their existing infrastructure, where sensitive data never leaves their environment. When the sanctioned option is also the best option, shadow use fades on its own. If you are weighing how to do this with your own data, our guide to using your own data with LLMs securely covers the governance fundamentals.
  3. Write rules people can actually follow: A one-page acceptable use standard beats a forty-page policy nobody reads. Pair it with the broader governance thinking in what every executive needs to know about AI governance.
  4. Make managers accountable for adoption: In one client rollout, participation in AI training jumped significantly the moment completion metrics moved from individual employees to their managers. Governance sticks when leaders own it.
  5. Keep listening: Shadow AI never fully disappears. New tools launch weekly. A quarterly check on what people are reaching for tells you where to invest next, and pairs well with getting your team to actually use the AI you approve.

The gap between AI aspiration and AI that works

Shadow AI is what happens when your people’s AI ambition outruns your company’s AI foundation. The fix is not less AI. It is AI done on purpose: diagnosed, built for your environment, and backed by someone who stands behind it in production.

That is the work of Security Intelligence, and it pays off beyond risk avoidance. Across our client work, governed environments have unlocked what unsanctioned tools never could: standardized platforms that let whole departments share what works, secure builds that finally let privacy-conscious leaders say yes, and manager-led rollouts that measurably lifted training completion. Companies that govern AI well do not slow down. They get to speed up safely.

Your team is already using AI. Put it to work for them, on your terms.

Your Meetings Cost $29K a Year, and Most of It Evaporates

/by

 That’s $29k per employee. Per year.

In a recent Augusto leadership meeting, the point was made that our client roundtables generate a massive amount of intelligence worth sharing across the company. Nobody in the room treated that as a nice sentiment. We treated it as an operating requirement, because we had already built the system that makes it true: every conversation we have is captured, mined, and turned into knowledge the whole company can use.

Most companies run the opposite way. Their best thinking happens out loud, in meetings, and then most of it disappears. The decisions, the commitments, the client saying something that belongs in a case study, the early warning sign in a customer’s tone. It all happened, and a week later almost none of it exists anywhere. That gap between what gets said and what gets kept is the whole argument for AI meeting intelligence.

The Evaporation Problem

The research on this is brutal. Without follow-up notes, 70 percent of decisions made in meetings are forgotten within 24 hours, and 47 percent of action items discussed are never captured at all. This is not a discipline problem. People speak faster than anyone can reliably record, and the person taking notes is also trying to participate.

Now put a price on it. The average employee spends 392 hours a year in meetings, at a cost of roughly $29,000 per employee per year, and executives average around 23 hours a week. Estimates of what unproductive meetings cost US companies run from $37 billion to $399 billion annually. Your company is already paying full price for these conversations. The only question is whether you keep what you paid for.

What Meeting Intelligence Actually Means

Most leaders hear this topic and picture an AI note-taker that emails a summary nobody reads. That is the shallow version. AI meeting intelligence means the conversation becomes structured, searchable business data. We know because we built this for ourselves before offering it to clients.

The capture layer comes first, and ours is standardized: every call at Augusto is recorded with Fathom, our company-wide AI note taker. We like it because it does the unglamorous part flawlessly. It joins every meeting automatically, produces an accurate transcript with speaker labels and timestamps, and delivers a summary with action items before anyone has left the call. Just as important for what comes next, every moment carries a deep link, so any quote or decision can be traced back to the exact second it was spoken in the recording.

From there, every call flows into a system we call our Second Brain, and this is where Fathom’s clean transcripts become compounding value. AI reads each Fathom transcript and extracts the action items, with owners. The system identifies ROI signals when clients describe measurable outcomes in their own words, detects relationship signals such as appreciation, concern, or renewal risk, and maps attendees, their connections, and trending topics across every engagement.
Before a client meeting, anyone on our team can pull a prep brief built from the entire history of the relationship. When we write a case study, the client quotes are real and verbatim, because the system kept them. This is the same agent workflow pattern we apply to any document-heavy process, pointed at conversations instead. It even feeds our marketing: the proof points in our content pipeline, described in how Augusto automates content creation, come out of these meetings.

What Changes When Conversations Stop Evaporating

The first shift is follow-through. Teams that adopt AI-generated meeting summaries see action-item completion rates rise from the 50 to 60 percent range to 85 to 95 percent, largely because a searchable record kills the follow-up meetings scheduled to clarify what the last meeting said.

The second shift is institutional memory. When a veteran team member leaves, their client context does not leave with them, because two years of their conversations are searchable. When someone new joins an account, they read the relationship’s history instead of asking five people to repeat it. And leadership stops flying blind between quarterly reviews, because concern and satisfaction signals surface from calls as they happen, not months later in a churned account. None of this requires anyone to change how they run a meeting. The meetings you already hold become an asset instead of an expense, which is the cleanest ROI case we know. Our guide on measuring AI ROI before you invest shows how to baseline it.

How to Start

The pattern is simple. Record by default, with clear consent and sensible exceptions. Pipe transcripts into a system that extracts action items, signals, and people, rather than a folder of files nobody opens. Connect the output to where work actually happens, your CRM, your project tool, your team channels. Keep a human reviewing anything that drives a decision. Start with client-facing calls, where the evaporating value is highest, and expand from there.

Your meetings already contain the intelligence. The only thing missing is the system that keeps it. If you want to see what a Second Brain would look like on your own calls, start a conversation with our team.

Monthly LLM News September 2026

/by

August gave us a first: the most capable model of the year got pulled off the field by its own team, one week after its greatest win. Add the biggest leadership shakeup in Google’s AI history, an 80 percent price collapse, and one billion people now using AI weekly, and the month reads like an industry growing up in public. Here is your LLM news roundup, and what each story means for how you run your business.

OpenAI’s Astra solved 80-year-old math problems, then got benched

In the first week of August, OpenAI announced that an internal version of Astra, its next major model, solved ten previously open problems in mathematics and theoretical computer science, publishing formally verified proofs for roughly $2,000 in compute. These were not benchmark tricks; some had resisted mathematicians for decades.

Five days later, the same model earned a very different distinction. Astra became the first system OpenAI classified as potentially Critical for cybersecurity under its Preparedness Framework, meaning it may be capable of finding zero-day exploits or executing novel attacks on hardened systems on its own. OpenAI paused every internal Astra activity that did not meet a strengthened set of security controls, and Axios later reported that its largest planned frontier training run remains on hold.

The smartest model of the year is currently benched by its own maker. That is the whole story of AI in 2026 in one sentence.

What it means for your business: capability is still climbing steeply, and so is the caution around it. Do not build Q4 plans around a promised model release, because safety pauses are now part of the calendar. And take the dual-use warning seriously: if a frontier model can attack hardened systems, your security posture is the differentiator. Our guide on AI governance for executives is the place to start.

Frontier AI prices collapsed 80 percent, but read the fine print

On July 30, OpenAI cut the price of GPT-5.6 Luna by 80 percent, from $1 and $6 per million tokens down to $0.20 and $1.20, with Terra down 20 percent. A week later Luna became the free ChatGPT default. The pressure is real: Chinese open-weight models captured roughly 46 percent of US enterprise token usage on OpenRouter, and pricing is the counterpunch.

One caution before you celebrate: some of the new low prices are promotional. Google’s newly released Gemini 3.7 Flash, for example, carries pricing that is locked only through 2026, with the list price set to double on January 1, 2027. Teaser rates are now part of the AI market.

What it means for your business: re-run the math on every high-volume AI workload, because tasks too expensive in June may be nearly free now. Then diary a pricing review for every model you depend on, so an expiring promo does not quietly double a line item. AI budgeting is where enthusiasm meets the CFO, and this month moved the numbers in your favor if you act.

 

Google’s AI brain trust walked out the door in one day

On August 5, Google announced that Demis Hassabis is stepping back from running Google DeepMind day to day, becoming its Chair and Alphabet’s Chief Scientist, while Jeff Dean, employee number 30 and a 27-year veteran, left to co-found the startup Discovery Loop with fellow senior researchers. Alphabet fell about 4 percent, and reporting has since tied the shakeup to stalled models, missed deadlines, and staff burnout, with Gemini’s next flagships months behind.

What it means for your business: vendor stability is now a diligence item. July showed what happens when a model disappears overnight; August showed even the biggest labs can wobble organizationally. The answer we build with clients is model-agnostic plumbing, so critical workflows reroute between providers without a rewrite.

One billion people now use AI weekly, and they work for you

Alongside the price cuts, OpenAI disclosed that its models now reach more than one billion active users and over two million businesses, with usage deepening over time: six months in, people send about 50 percent more messages and use AI for twice as many kinds of work. The shift OpenAI describes, from asking to doing, matches what we see in mid-market companies weekly: employees are not waiting for an AI strategy, and leaders in our workshops still call their environment a Wild West.

What it means for your business: your customers now experience AI-grade responsiveness everywhere else, so their expectation bar for you just moved. And your employees are already among that billion, which makes governed, sanctioned AI a retention and security issue at the same time.

Meta doubled down on open weights, and that helps you

In mid-August Meta shipped its Muse Glimmer model alongside a 6,500-word Zuckerberg manifesto arguing for more open-weight releases plus a $1 billion fund for data center communities. The practical effect: credible open-weight options keep improving, strengthening your hand at every renewal and keeping self-hosted AI viable for cost-heavy or privacy-sensitive work. The client teams that standardized on governed platforms with an open-model fallback are sleeping through months like this one.

Four moves for September, with owners

  1. Re-run your model pricing (CFO and CTO, 30 days): Test Luna-class models on your highest-volume tasks and diary every promotional rate’s expiry date.
  2. Kill your single point of failure (CIO, 60 days): Add fallback routing so no one vendor pause, shakeup, or shutdown can stall a production workflow.
  3. Raise your customer experience bar (CEO, this quarter): A billion weekly AI users are recalibrating what responsive feels like. Pick one customer-facing workflow and close the gap.
  4. Harden before the agents arrive (Security leader, now): Astra’s Critical rating is a preview of the threat curve. Verify your AI use is governed and your defenses assume AI-assisted attackers.

The pattern across all five stories is the same one we see in the field: the technology is sprinting, and the winners are the companies whose foundations let them change direction without falling over. Treat AI as infrastructure, deploy it against high-value workflows, and scale what delivers measurable returns.

Put AI to work for your people, and keep your footing while the industry finds its own.