62% of large organizations were building, deploying or developing AI agents in September 2026, KPMG found. After a year and a half of agentic AI in production on LTFI, our lessons are about approval lines, logs, written rules, cost, drift and people.
KPMG reported in September 2026 that 62% of large organizations are now building, deploying or developing AI agents. We have run agentic AI in production at Kief Studio for a year and a half, on our own system, LTFI. What it taught us after the demo is mostly about approvals, logs, written rules, cost, drift and people, and very little about which model is smartest.
Agentic AI in production means AI teammates doing real steps of real work every day, such as researching, drafting, updating records and preparing messages, inside the systems a company already runs, under rules and approvals, with a record of everything they did. A demo shows what an agent can do once. Production shows what it does on the four hundredth Tuesday.
This is a field report rather than a design guide. The design side, building agents that survive malformed data at 3am, lives in AI agents that run at 3am versus agents that demo well. What follows is what changed in how we work once agents had real jobs.
After a year and a half, agentic AI in production looks ordinary: a person asks one chief of staff what needs them today, the work goes to the AI teammate that owns it, drafts go ahead and anything that leaves the company waits for a yes. The novelty wore off within weeks. The routine is what stayed.
The routine runs like this. A person asks the chief of staff a question in plain words. The chief of staff hands the work to the teammate that owns it: outreach, writing or research. That teammate pulls the context it needs from the company's own records, does the work and either finishes it or puts it in front of a person for approval. Briefs, summaries, research, drafts and record keeping go ahead on their own.
The agents live inside the same system as the CRM, projects, content and sites, so they work from the company's real records rather than from whatever someone pasted into a chat window. That one choice explains most of what follows. An agent that can see the project history writes a follow-up that sounds like it knows the client. An agent that cannot see it writes a confident guess.
Here is a small example from my own week. Brian had shipped a batch of updates while I was deep in client work, and I needed the high level without reading every change. I asked the system what had happened on that project and had a clean summary in seconds, including the fact that we had a LinkedIn event coming up, because it was already in the project history. Brian described the same moment in our October webinar. Nothing about it was clever. It was simply faster than asking a person to stop and explain.
Larger companies are arriving at the same place. KPMG found that nearly half of leaders (49%) have defined high-risk use cases where they do not allow autonomous decision-making by agents. The companies that move fastest are the ones drawing lines, and that is where our first lesson starts.
The approval line decides which work an agent finishes alone and which work waits for a person, so it sets both the speed and the risk of the whole system. Get it right and people stop worrying about what the agents might do.
Agents are eager. Brian put it plainly in the October webinar: they are very eager to complete their task, and every now and then you need to tell them to chill out. The approval gate is how you do that without slowing the work that should be fast. These are the default lines we have run for a year and a half. Each company sets its own.
| Goes ahead on its own | Waits for a person's yes | Person-only |
|---|---|---|
| Briefs and summaries | Sending email | Rules and permissions |
| Research | Posting and publishing | Who approves what |
| Drafts | Messages to clients | Passwords and keys |
| Record keeping | Spending and scheduling | Deleting and money |
Some companies choose to let low-stakes sends go without a yes. We kept ours, because a double-booked meeting costs more trust than the approval costs time. Brian said the same thing in the webinar: some teams like to move fast and accept the occasional double booking, and we prefer not to.
The surprise was how little friction the gate adds when the draft is good. In the live demo, the outreach teammate pulled the client's context, wrote the follow-up and waited. Brian read it, approved it and it went. He did not write the email; he decided whether it should go. That split, where the agent does the work and a person owns the decision, is the whole design. It also matches the guidance we keep pointing people to in the tasks you should never automate: the judgment calls stay human.
The log shows exactly which instruction, context or step produced an output, so it tells you what to fix rather than only that something went wrong. In our experience it became the main tool for improving the work, not just a record for accountability.
We started logging for accountability. Every action records who asked, which AI teammate did the work, which model it used and who approved it. Within weeks the log had a second job. When an output missed, we could trace it to the cause: a vague instruction, a stale record or a step that ran on the wrong model. As Brian put it in the webinar, the log is more than proof of work; if you did not like the last output, it shows you exactly where to tweak it.
The wider record explains why this matters. Stanford's 2026 AI Index found that the AI Incident Database recorded 362 documented AI incidents in 2025, up from 233 in 2024. Most businesses will never appear in that database. Every business running agentic AI in production will have small misses, though, and without a log each one becomes an argument about what happened instead of a short fix.
The log also answers questions nobody planned to ask. Which tasks get repeated every week and should become a standing job? Which desk is struggling with the system? Where are people still doing by hand what an agent already drafts? The review habit is the same one described in how to audit what AI is actually doing, only with the evidence already collected.
Write each rule once as a company rule so every AI teammate follows it, then have a second AI teammate check the work against those rules before it counts as done. Rules prevent the repeat mistakes; the check catches the new ones.
Early on, the same corrections came up again and again: the house style, the facts every press release must include, the claims we never make. Brian used the press release example in the webinar because it is so common. The fix was to write each rule once, in one place, and have every AI teammate follow it from then on. A rule written in one place is a rule you can change in one place, and the next draft follows the change.
Rules alone are not enough for agentic AI in production, because models still invent things. In a new accuracy benchmark reported in the same Stanford Index, hallucination rates across 26 top models ranged from 22% to 94%. So one AI teammate's work can go to another to check against the written rules before it counts as done, often on a different model, with a person reading anything that leaves. A bad draft caught by a checker costs a few seconds. A bad draft in a client's inbox costs a relationship.
The order we settled on is simple enough to copy:
Running cost and drift are the two expenses demos never show: model spending grows quietly unless it is capped and visible, and the same model can behave differently from one month to the next. Both need a budget and a baseline.
Model spending behaves like a utility bill nobody reads until it spikes. KPMG's June 2026 survey found that only 26% of organizations report full, real-time visibility into what their AI systems cost to operate. By September, 74% included cost reviews in AI approval processes and 43% had implemented usage or token budgets. We cap spending per person and run each job on the model that fits it, a fast one for quick checks and a stronger one for planning or research. That is the practical version of owning the workflow and renting the model.
Drift is the quieter cost. Researchers at Stanford and UC Berkeley found that one widely used model identified prime numbers with 84% accuracy in March 2023 and 51% in June 2023. Nothing in a customer's workflow had changed. The provider had updated the model. For agentic AI in production, the defense is a baseline: the log, the written rules and the second-model check tell you when output moves before a client notices. We go deeper on this in how to catch AI agent drift.
Drift also comes from inside. A new rule, a changed record or a new tool can shift behavior as much as a model update. That is one more reason rules live in one place and every change is logged: when output moves, you can see what else moved that day.
People decide whether agentic AI in production sticks, because a system nobody trusts or enjoys using gets quietly routed around. In our experience, adoption followed three things: letting people keep the AI tools they like, scoping what each seat can see, and sharing lessons across desks.
The technical lessons were the easy ones. A system everyone has to learn from scratch gets used by two people. A system that works with the AI tools each person already prefers gets used by everyone. Brian made this point in the webinar: one person likes one assistant for writing, another prefers a different one for technical documentation, and the team does not have to pick, because both reach the same company context and the same rules.
Permissions turned out to be a people question too. A contractor who needs the files does not need the AI system, and the person running social posts does not need the financials. Scoping each seat made people more willing to use the system, not less, because nobody had to wonder what it could see on their behalf. If you hold Massachusetts residents' personal information, the access rules in your written information security program under 201 CMR 17.00 already point the same way.
The best surprise was how lessons travel. When one desk learns something, it becomes a lesson every AI teammate can use. Brian's example in the webinar was a lobby lighting job where one type of bulb did not work: once it is recorded, nobody repeats it. KPMG found that significant workforce adoption increased to 44%, up from 23% last quarter and 10% a year ago. Adoption follows usefulness, and usefulness follows shared context.
That matters most for the businesses we work with across Worcester and Central Massachusetts, where one person often covers three desks. A newsletter that needs input from ten people can take a week of passing drafts around. When the context from those ten desks is already in the system, one person can produce a first draft quickly and spend the week on approvals instead of chasing input.
LTFI handles agentic AI in production as one system: the CRM, projects, content, sites, security and AI staff together, run by Kief Studio on a dedicated server for each company, with approval lines, a full log and written rules built in from the start. Every lesson above is a feature because we learned it running our own company on it first.
LTFI is an agentic business operating system. There is no shared hosting. What that means in practice:
Lookalikes are arriving now, many of them a chat window or a wrapper around someone else's agent. LTFI was meticulously built from scratch by Kief Studio as one whole system, it is run for each company by the team that built it, and it has been in live production for a year and a half. Your data stays yours, and if you leave, we export it. The governance side of that difference is the subject of AI governance for agentic AI is not a wrapper.
If you want to see it working, there is a walkthrough at ltfi.ai, and the October 23 session is on the webinar replay page. We have worked remotely with clients around the world for more than 13 years, so LTFI can be set up wherever you are, and for businesses in the Worcester and Shrewsbury area we are glad to meet in person.
Agentic AI in production is AI that takes steps on real business work every day, such as drafting, researching, updating records or preparing an email, inside the company's systems rather than in a demo. It runs under approval lines, a log and written rules, with people owning anything that leaves the company.
Work that stays inside the company and is easy to undo: briefs, summaries, research, drafts and record keeping. Sending email, posting, client messages, spending and scheduling should wait for a person, and rules, permissions, keys, deleting and money should stay person-only.
The first useful work, such as briefs and drafts, shows up within days once the agents can see real records. The larger gains build over months, as written rules, logs and shared lessons accumulate and fewer outputs need correcting.
Cap spending per person, run each job on the least expensive model that does it well, and review the log for repeated work. KPMG found in June 2026 that only 26% of large organizations had full, real-time visibility into AI running costs, so visibility comes first.
They should not have to. In our experience the agents work best when they connect to the tools a company already pays for and read from its real records, so the change is in how work moves between desks, not in which software the business runs.
Website maintenance cost depends on which of three jobs you buy: hosting, plugin updates or real website operations. Patchstack counted 11,334 new WordPress vulnerabilities in 2025. Published 2026 ranges, Worcester labor figures and the drivers that move the price.
Ninety days is a realistic plan to offboard a managed service provider. Compress it and the revocation step is what gets cut, leaving agents installed and credentials live for months.
Six tasks get dropped in nearly every co-managed IT arrangement, starting with deprovisioning on someone's last day. Assigning them explicitly takes an hour and closes most of the seam.
Work With Us
Kief Studio builds, protects, automates, and supports full-stack systems for businesses up to $50M ARR.
Newsletter
Strategy, psychology, AI adoption, and the patterns that actually compound. No spam, easy to leave.
Subscribe
