vas@vasuman5d ago

📝 Where to Begin With AI Transformation Where should you start with AI if you run a large company? The short answer: it depends. There's no list of 10 steps to execute in sequential order that results in 90% efficiency gains. If an AI transformation partner gives every enterprise the same answer, they haven't done this very many times. I have a strong opinion on AI transformation because we've seen it work at every company we've been inside, from Fortune 50 enterprises to strong PE-backed companies. At a high level: perform discovery across a department, focus on key workflows, agree on value, and build a company brain. But I want to be clear, discovery is where (close to) all the value lies. We worked with a client that engaged Deloitte, then fired them after 3 months because their discovery was simply awful. Consultants who perform "interviews" but don't know what questions to ask are going to waste your money and, more importantly, your time. And this is all too common: we love to point and laugh at the big consulting shops, but I've seen "applied AI companies" do the exact same thing. Knowing what to ask, how to find the information that matters for AI transformation, and THEN being able to build it successfully, is where all the alpha lies. There are probably fewer than 1,000 people in the world who truly know how to do this correctly: meaning both efficiently and accurately. One caveat: do not be fooled by the AI transformation partners who "specialize" in a department. If you want to bring agents to Finance, they must span pieces of Sales, Operations, HR, and more. The same goes for Sales Ops agents: those workflows involve Finance and Marketing. HR involves IT. There is no such thing as a department-specific agent. That's a relic of a human world, where specialization was the byproduct of limited bandwidth and ability. In the agent world, that paradigm is a constraint. Your agents must be able to span multiple departments, because they are capable of knowing everything at the same time. Make no mistake, AI transformation is an invasive procedure. It's highly custom per client, and done incorrectly, it can cripple your business with wasted months and broken migrations. Done properly, however, it can save you from certain death, even give you the life you always wanted. Yes, I mean double your margins, and make a team of 150 operate like it's a team of 300. Here's how to do it properly, if you'd like to embark on this journey for your own organization. 1: Figure out where you're at right now Most companies we speak with fall into one of 3 categories, and each has its own set of problems You haven't started (this isn't even the worst one) Either security, legal, or a migration to a new ERP stood in your way, and now you're worried you're too far behind. Too frequently, I hear: "we just have to get our data in order". You've been saying this for years. AI makes getting your data in order easier than ever before. It is a part of your AI transformation journey, not a prerequisite to it. Your data will always have issues; the transformation effort should expose and remedy these gaps anyways. The good news is, even if you haven't started, you're likely in a better position than you think. You don't have 20,000 seats of AI software collecting dust, (or worse, billing you $10M/month), and you don't have a "center of excellence" preventing any work from beginning unless it's one of the 40 'best use cases' it rated over the last 6 months. You've deployed copilots but hit the plateau (this is the most common one) You've deployed Copilot or ChatGPT Enterprise in the last 2 years, you've trained your people on how to use it, and you've even had them build hundreds of their own agents. Usage went exponential for a few months, but then plateaued. People say it's helping, but it's not showing up on the P&L. The best example is still summarizing meetings. And according to McKinsey's 2026 State of AI survey, you're not alone: 80% of people using AI say it's made them more productive, but only 37% of respondents said it has had any impact on earnings. That's because a copilot (or sidekick) is going to help people be 10-20% faster in a process when they think to prompt it for help. The issue being, this is at best only 8 hours a day (9-5 but even less if we're being honest), and limited by the users' ability to prompt. But when you have agents working in the background in your systems, only reaching out to people when exceptions come up, that's where we're seeing 60-90% improvements in e. If you have a 12,000 person company, and 1,200 of them use Copilot to be 15% faster for a third of their day, you're not going to see any impact on the P&L, since you're not changing anyone's job or the cost line (it's the equivalent of giving them 24 minutes back per day). But if you have a 40 person accounts payable department, and you use agents to automate 75% of the work, that's a 30 person reduction to the cost line your CFO is worried about. Your pilots are stuck (this is the worst one) 9 months in, you have a ton of pilots, but none of them have made it to production. It's likely because 1) you built them within one team 2) you didn't map out the process first 3) you didn't set a metric before building. So the demo worked on the happy path, failed on the first exception, and nobody knew what fixing it was worth. Why is this the worst one? Because the entire company feels fatigue. They gave it a shot, it didn't work, and no one wants to do it again, because it feels like a herculean effort, and it's easier to just sit and believe AI doesn't work. It does work, it's working for your competitors, and they're going to crush you. For context, I'm Vas, the CEO of @varickagents, and this is what we do, day in and day out. We partner with some of the world's largest, most complicated organizations, and help them make the most out of AI: from strategy and training, to process mapping and reengineering, to building and deploying agents, to optimizing and expanding them. 2: Start with a department Take a step back and realize regardless of where you are, you need to start your discovery in a department, not a single workflow, and not the entire company. Look at every system of record, process flow, rule, agent, and person in it. We say department, not workflow, because most of the time taken in a process (e.g., procure to pay) is not spent in just one department (e.g., accounts payable). It's likely spent in one or more other departments (e.g., procurement, receiving), and your close each month can be held up by an exception in payroll. So when you work to optimize one workflow at a time (e.g., exceptions in accounts payable), you're likely going to miss the area that would have the biggest impact (e.g., the exception rate in procurement or receiving). Each department feels like it's meeting its SLA, but the process takes 3 weeks because of the handoffs between departments. We call these the upstream and downstream dependencies. This is also why you can't keep agents inside one department: your finance agents will read from and write to a lot of other departments, like procurement, payroll, etc. So if you engage a partner that only builds agents for one department, like finance, they won't be able to map the entire process. Which department should you start with? The one that scores highest on 3 criteria: volume (how many times its processes run a month), cost (do you know what this department costs; most often the CFO will have this number), and reach (how many departments its processes touch). Most often this is finance, then sales operations, then HR or procurement or operations, and finally legal and marketing. For context, this rings true even in our client work; around 40% of our projects land in the office of the CFO, then 25% in Sales operations. 3: 7 steps Let's walk through these steps with a global distribution company that came to us a year after deploying a copilot to 12,000 employees and hitting a plateau. They had 3 pilots going, but these were stuck inside the organization, and the CEO wanted results to show the board in 16 weeks. 1. Pick the department that scores highest Finance scored highest because of the volume of processes (the month-end close took 20 days, and there were 700 exceptions a month in accounts payable) and the reach of the processes (the close involved 4 departments). Sales operations scored second, and legal scored lowest because of its low volume of processes. 2. Agree on the value and write it down The CEO, the CFO, and the controller agreed on a set of KPIs in 3 categories: time and/or cost, revenue (after token spend), and risk. Then they wrote down where each stood. The month-end close took 20 days, there were 700 exceptions a month in accounts payable, and it cost $31 to process an invoice (everyone there assumed it was $5). That one-pager is what the CEO took to the board in week 16. 3. Map the as-is process This step is the discovery, and it's where most partners fail. The as-is process was documented in 11 steps, but by reading the event logs in the ERP and talking to the people, it turned out to be 20 steps in the ERP and 4 more in email. Half of the extra steps existed because intercompany invoices were sent without a purchase order, and none of that was in the process documentation. So it takes both the event logs in your systems and talking to the people to map the process, and you need someone who knows what to ask. 4. Design the to-be process For each step in the as-is process, the question is: can we delete this step, or can we build a piece of code or an agent to do it? In this example, of the 20 steps the finance function used to do, 7 were deleted, 4 are done by code, and 6 are done by agents. 3 are still done by a person because the auditors want a person to do them. So instead of a person doing 20 steps, they do 3, plus all the exceptions the agents send them. The biggest mistake people make is that they don't delete steps and just automate the as-is. In McKinsey's 2026 survey, almost 75% of the companies making money with AI had fundamentally changed their processes, compared to only about 25% of the others. [Figure 3 here] 5. Build on the systems you already have The agents are built in your current systems, so your people don't have to migrate to or log into a new system. The agents are all code, except for a small part. A model is only used where the input is unpredictable, for example, reading a new invoice layout or coding an invoice to the correct general ledger account based on the previous 1,000 invoices. Everything the agents learn about your company is stored in your company brain, so when you start in a second department, the knowledge is already there. 6. Run in shadow mode, then go live Run the agents in shadow mode for the first 2 weeks (at full volume) without allowing them to take action. Review every case where the agents come to a different conclusion than the people. From week 3, let the agents act on the cases where you've coded all the checks and are 99%+ sure of the action (normally around 50% of cases). By week 12 they handle about 90%, and your people keep the rest. [Figure 4 here] 7. Train the top slice alongside During all of the above, identify the people (normally 5-10% of your organization) who are already using AI to make themselves better at what they do. Give them a day's training, then run a leaderboard on how much time they're saving with their skill files (not the number of tokens they're using). Then get them to build the agents for the next department you want to tackle. Everyone else just gets the agents in the background. In this example, after 4 months, the company closes its books in 8 days at month-end, and its AP team has fewer than 50 exceptions a month to work through. Next, it goes through the same steps for sales operations. What changes depending on where you start Whichever of the 3 places you start from, you then go through the same 7 steps. If you haven't started, go to step 1, but don't roll it out across the organization. Get security, legal, and finance involved in week 1. As of October 2026, 79% of organizations have AI agents running in production, but only 2% are using identity security tools built for them, according to SailPoint. If you've rolled out Copilot seats across the organization, keep the seats (everyone is quicker at what they do with them), but run a report on who is using them most, identify the top 10% (these are your step 7 people), and then pick a department to start with. If you've been running pilots, put a number on each one as in step 2, and kill the ones you can't. Roll the rest into the department you've picked and map their processes as in step 3. It's likely each pilot covers just 1 step of the process (so it's saving 8 minutes on a process that takes 22 days to complete). Why most companies don't start here Why don't more companies start here? Everyone wants to sell you something else. Seats are easy to sell, and easy to show the board as progress. Or one of the big consulting shops will want to come in, run workshops, carry out interviews, and identify all the use cases you could have (the 40 you could come up with yourself and then run through a 1-5 scoring grid), and they usually don't know what to ask. And the labs want to sell you their engineers to build the agents for you. OpenAI launched its deployment company in May, backed by 19 partners who gave it $4 billion, and Microsoft launched its Frontier Company in July, backed by $2.5 billion and around 6,000 embedded engineers and industry experts. As the CIO of Vistance Networks wrote in CIO.com on October 13, 2026: "We are spending millions sending AI experts to learn business. We should be sending business operators to learn AI." He's half right, because the people who know your business can't do all of the above on their own (especially running agents in shadow mode and then getting them through a SOX audit), so you need both to get the best out of starting with a department. And no one can sell it to you as a product, because it isn't one. https://x.com/vasuman/status/2105083554631942541 What does this look like at Varick? We do this at Varick. We start at the department level with a pod of operators and engineers, do the discovery, get agreement with the CFO on the numbers, and build agents on top of your current systems. It takes about 4 weeks per agent from build to production. If you want to do this at your company, book a call. If not, subscribe to our newsletter, where I share what we learn doing this inside very large companies. TLDR You're probably in one of 3 places: you haven't started, your copilot rollout has plateaued, or your pilots are stuck. Most are in number 2. A copilot can make people 10-20% faster when they prompt it, while agents that work in the background can do 60-90% of the workflows you give them. So what should you do next? Regardless of where you're at, pick a department and rebuild it end to end. You'll save the most time in the handoffs between departments, which is why the agents will span multiple departments, and you win or lose it on discovery. Pick the department based on volume, cost (the CFO will have it), and the number of departments it touches. It's usually finance. At a global distribution business, in 16 weeks, a person now does 3 of the 20 steps in the close, the close takes 8 days instead of 20, and AP exceptions are down to under 50 a month from 700. https://x.com/i/article/2111222023745945600