AI consulting from an operator who has built with it.
Every board and every investment committee is asking the same two questions about AI: where does it actually pay, and who has actually done this? I answer the second question with a career, not a slide. I founded and ran an AI company as CEO, shipped a generative AI product that Fortune 500 teams use, and led cloud and DevOps transformation across more than a thousand engineers before that.
Then we answer the first question inside your P&L, not in a deck. I turn AI into EBITDA for portfolio companies: workload by workload, starting with the boring wins. If you run a PE portfolio, a company, or a team whose board wants an AI answer, this page is for you.
Where AI pays, workload by workload.
No platform worship, no doom, no hundred-page strategy. A ranked answer to where AI pays in your business, proven one workload at a time, so each win funds the next one.
I am Ken Gavranovic, an AI consultant who founded and ran an AI company as CEO. The people I do this for are private equity operating partners, the CEOs and CFOs of their portfolio companies, and the CEOs of companies whose board wants an AI answer: AI out of pilots and into production, with a number the CFO can audit. The situations I have worked in are written up below: a mid-market CEO with no enterprise budget, an enterprise whose AI bill had no owner, and a private equity-owned company that put each function on the hunt.
Start with the boring wins
The fastest returns are rarely the flashy ones. We inventory the work your people re-type, re-key, and re-read all day: calls, emails, documents, browser work. That list, ranked by what each flow costs you, is your first AI roadmap.
Prove it before you scale it
One workload, one owner, one number to beat. You get a clear roadmap in weeks and proof inside a quarter, so the second workload is funded by the first and nobody is defending a faith-based budget.
People and process first
AI transformation fails the way digital transformation failed: on adoption, not technology. I wrote the playbook on the people side of AI transformation, and I run it with your team rather than around them.
Own the tokenomics
AI spend drifts when nobody owns it. Model choice, inference placement, and workload fit decide whether your bill is a rounding error or a budget line. Matching models to tasks has cut workload spend by 80 percent.
I build with this. I do not just advise on it.
As CEO of Product Genius I built and launched an AI engagement platform and took it to national rollout in under six months. As COO of Blameless I shipped a generative AI incident-response product used by Fortune 500 engineering teams. As EVP and GM of the New Relic platform I launched New Relic AI, anomaly detection across two billion events a minute, so customers heard about a failure before it happened. Before that I led the cloud and DevOps transformation at Cox Automotive, one of the largest automotive platforms in the country. And I use this technology daily in my own work, so the advice tracks what the tools can do this quarter, not what they could do last year.
Building and running software companies since 1995, and advising Fortune 500 corporations and private equity firms, the pattern I see is consistent: the companies that win with AI treat it as an EBITDA program, not an innovation program. That is the program I run.
Ken led our $160M+ cloud, engineering, and DevOps modernization across a $3B+ portfolio. He brought clarity, velocity, and accountability to every initiative.
Outcomes, not pilots.
The $250,000 answer
Mid-market · automationA $25M-revenue CEO asked how to use AI without an enterprise budget. We skipped the platform conversation and automated the flows his team re-typed all day: calls, texts, emails, browser work into systems. A one-time $50,000 investment took $250,000 of annual cost out.
The AI bill nobody owned
Enterprise · AI spendAn enterprise went all in with a frontier model vendor, and the tokens burned with no proof of value. We mapped the workloads honestly: open models were plenty smart for a large share of the tasks, which cut that spend by 80 percent, and the speed-critical paths moved to faster inference, coming back roughly 18X faster. Same outcomes, a fraction of the bill.
The whole company hunts
Private equity · operating modelA private equity company stood up a whole-of-company AI initiative: every function, not just engineering, charged with finding what in its own work could be modeled and automated, and bringing the idea or the working automation back. That flips AI from an IT project into an operating cadence, and it's where the durable EBITDA comes from, because the people closest to the process find the wins no consultant's sweep would.
What the data says. And what I measure.
73 percent of private equity executives expect AI to lift portfolio value in the next year. 8 percent say it is materially moving EBITDA today, per Alvarez and Marsal's 2026 survey of 100 PE executives. That gap is not a technology problem. It's an execution problem, and it's where I work.
My operating rule: any process you can model can be mostly automated, with the exceptions routed to humans. Applied to customer service, customer success, support, and claims processing, that pattern typically takes 30 to 70 percent of the work off human queues. The public record brackets the same range: Klarna's assistant resolved two thirds of support chats in its first month, strong deployments run 70 to 75 percent, and honest independent tests of off-the-shelf agents land near 38 percent. The spread is execution. For one back-office workload you already have in mind, custom AI agents are built in weeks.
In software teams, output runs 3 to 17x with AI depending on the codebase and the guardrails, based on my own builds, operator surveys, and direct conversations with engineering leaders at Fortune 500 enterprises. The number nobody puts on the slide: when coding stops being the constraint, the bottleneck moves to product and QA. Google's DORA research shows the same shape, with throughput up 20 percent and incidents per change up 23.5 percent. Plan for the new bottleneck or the multiplier eats itself.
Three shapes of engagement.
Same thesis as all my consulting work: advice when you need judgment, ownership when you need the outcome delivered. We pick the shape together on the scoping call.
AI Assessment
A fixed-fee assessment of where AI pays in your business, ending in a written plan you can run with me, run yourself, or shelve: the workloads ranked with a number on each, build, buy and skip calls, the first workload with an owner and a baseline, a 90-day plan, and a readout for the board. How it runs: a small business, under 50 people, takes two weeks, remote. A mid-market company, 50 people up to $1B in revenue, or any PE-backed company, takes four weeks. An enterprise is quoted after the scoping call. The fee credits in full toward an Adoption Sprint that starts within 90 days.
AI Adoption Sprint
A scoped implementation with your team: one or two workloads to production, owners and dates, adoption measured. Your people build the muscle; the win funds the next one. The board-facing version of this plan is written up in The AI Answer Your Board Actually Wants. How it runs: one quarter, one or two workloads, a baseline measured before the switch is thrown and adoption read against it at the end, with the owners in the room throughout.
Fractional AI leadership
When the mandate needs a seat, I act as a fractional Chief AI Officer: executive-level ownership of the AI agenda across a company or a portfolio, without the full-time hire. How it runs: an ongoing seat on a monthly retainer against an agreed maximum of hours, set by what the mandate needs, with a monthly line to the board or the operating partner, until someone inside passes the test.
What the assessment answers
- Where does AI pay in this business, ranked, with a number on each workload?
- Is the data ready: where does it live, and who owns it?
- Can the systems take it, and what has to connect to what?
- Which processes can be modeled, at what volume and unit cost, and where do the exceptions go?
- Are the people ready, and who owns each workload after I leave?
- What are the risks: governance, compliance and security?
- Build, buy or skip, and which vendors, if any?
- What is AI already costing you, in tools and spend without an owner?
- What happens in the next 90 days, and what does the board hear?
The output is a ranked plan, not a maturity score. The full checklist is in the AI readiness assessment guide. The forty-five minute scoping call before it is free. If you are comparing firms first, here is how to choose an AI consulting firm, with seven tests for one call.
For a sponsor the same three shapes map onto the portfolio: an assessment in each portfolio company, priced per company, as the sweep; one portfolio company's sprint as the proof the others follow, or the fractional seat held across the portfolio.
What an AI transformation has to solve, in the order it bites.
Call it AI advisory, AI consulting or AI transformation: the work is the same. I help private equity portfolio companies and CEOs get AI out of pilots and into production, with a number the CFO can audit, and the companies that get there pass through the same seven stages, in roughly this order. Most stalls happen at a stage that was not named. An AI transformation is a business transformation with a new tool in it: the work is the operating model, the decision rights and the habits, and the model is the smallest part. The companies that get there run it the way they would run any other transformation, with an owner, a baseline and a date.
1. Data and systems readiness
What has to be true: the process you want to automate has a record you can trust, in a system you can reach, with one owner. What goes wrong: a workload that looks simple on a process map sits on three systems, two of them a vendor's, with no clean history to learn from, and the model gets blamed for a data problem. The question that sets the timeline is where the data actually lives and who owns it, and I ask it before a date is set. Covered by the AI Assessment, which scores each workload on its data before it scores it on its savings. The pattern repeats across companies: the date moves when the record turns out to live in a vendor's system with no clean history, and it moves back when one owner is named for it.
2. The workload sweep and the number behind each one
What has to be true: a ranked list of the recurring processes in the business, each with its volume, what a unit of it costs today, and whether it can be modeled. What goes wrong: the most visible use case gets funded, a chatbot for the board slide, while the boring high-volume flows that carry the margin sit untouched, and afterward the return cannot be stated. Covered by the AI Assessment: a workload-by-workload map with build, buy and skip calls and the number behind each one, ranked by cost times confidence, so the ROI of each workload is stated before a dollar is spent.
3. From pilot to production, with a number to beat
What has to be true: one workload, one internal owner, a defined share of volume, an exception path to a named person, and a baseline measured before the switch is thrown. What goes wrong: a pilot has no owner and no exception path, it ran on curated data with a person watching, and production has neither, so it stays a demo. A workload in production at a measured share of volume, with an owner, beats a pilot that handled a demo. Covered by the AI Adoption Sprint: one or two workloads to production with your team, owners and dates, adoption measured.
4. Governance, risk and security with a named owner
What has to be true: before a model touches customer, financial or employee data, one person owns three things: which models are running and on what data, what each one is allowed to read, and a log of what it decided. What goes wrong: the AI is running and the governance and compliance are not, which goes unnoticed until a buyer's diligence team, a regulator or the board asks a question the CEO cannot answer. A lean portfolio company does not need a risk committee for this; it needs three owned documents and an exception rule for the decisions that matter. Covered by the Adoption Sprint, which puts the three controls on the first workload, and by fractional AI leadership, which holds them across the company.
5. Training, enablement and the Tuesday after the rollout
What has to be true: the people who do the work helped redesign it, were trained on the new flow rather than on the tool, and see their leaders using it. What goes wrong: licenses get bought, a kickoff email goes out, and three months later the work has gone back to the old routine, with adoption reported in logins instead of in volume off the queue. AI transformation fails the way digital transformation failed, on adoption and change management, and I wrote up the people side of AI transformation before the current tools existed; the rule has not changed: people first, then process, then technology. Covered by the Adoption Sprint, which builds the muscle inside your team, and by fractional AI leadership, which runs the program.
6. Who owns the mandate
What has to be true: one person owns the AI agenda at the level where trade-offs get made, and that person has shipped an AI product, has run a P&L of real size, and can get an operating team to adopt something it did not ask for. What goes wrong: a committee, or the CTO with AI as a side project, and the plan comes out technically elegant and commercially beside the point. Covered by fractional AI leadership: the Chief AI Officer seat across a company or a portfolio, without the full-time hire, until someone inside passes the test or the right talent is hired.
7. The running cost
What has to be true: you can see what each class of task costs to run and what it returns, and the routing sends the expensive model only where it earns its cost. What goes wrong: the pilot bill does not predict the production bill; an agentic workflow that loops, retries and verifies burns many times the tokens the demo did, and the first signal is the invoice. Covered by the AI Assessment, which sets the model and inference choice per workload, and by fractional AI leadership, which owns the bill as a line with a name on it. The call I make most often is to route the high-volume, low-judgment work to a small model and reserve the expensive one for the cases that earn it.
For a sponsor the same seven stages run across the portfolio, and the sweep does what a single company cannot: the same win shows up in several companies, so it gets built once and adopted everywhere, with adoption measured, not licenses. That is the work described under AI across the ownership cycle, below. The stage skipped most often is the fourth: governance arrives after the first workload is live, when a diligence team or the board asks who approved the model and what it read. The second is the baseline: the pilot ran before the queue was measured, so the return cannot be stated.
AI across the ownership cycle.
For private equity firms and family offices the work runs the whole cycle: AI diligence on targets, a post-close AI value creation plan that survives contact with the operating team, portfolio-wide workload sweeps that find the same win in six companies, and the AI story your exit will need. I have written up how a private equity firm gets AI into its portfolio companies.
I have sat on your side of the table too: a decade as a Managing Partner, 15+ exits, and 35+ M&A transactions as an operator. For the broader operating work, from post-merger integration to leadership transitions, see strategy consulting.
How a sponsor measures this
Three lines, agreed before the first workload moves and reported on the operating partner's cadence, inside the hold period:
- The EBITDA line: the share of volume that no longer touches a human queue, times what a unit of that volume cost before, against a baseline measured before the switch. A line the CFO can audit.
- Adoption: counted in volume off the queue, not in licenses or logins, read month over month.
- Exit readiness: which models run on what data, what each is allowed to read, and a log of what it decided, kept current from the first workload, so the AI story at exit is a record a buyer's diligence team can test, not a slide.
Based in Alpharetta, Georgia. I work remotely with leaders everywhere, and on-site when the work calls for it: workshops, working sessions, and engagements anywhere in the country. Atlanta metro in person is easy. In person across North Fulton and Cobb too: Johns Creek, Roswell, Sandy Springs and Marietta. (629) 304-8755 · ken@kengavranovic.com
The ones people actually ask.
What does an AI consultant actually do?
An AI consultant separates where AI pays in your business from where it does not yet. That means mapping real workloads, putting a number on each, and getting the first wins to production. The difference between AI consultants is what stands behind the map. Mine is having built and shipped AI products as an operator, and deployed them at enterprise budgets.
What does AI consulting cost?
The first step is a fixed-fee assessment, quoted on the scoping call. The AI Assessment is a fixed fee sized to the company, and the fee credits in full toward an Adoption Sprint that starts within 90 days. The sprint carries a project fee with owners and dates, and fractional AI leadership is priced like the operating commitment it is. The forty-five minute scoping call before any of it is free, no pitch, and the first workload is chosen so the savings are visible against the fee.
Do you act as a fractional Chief AI Officer?
Yes. For companies and portfolios that need executive-level ownership of the AI agenda without a full-time hire, I hold the seat: strategy, vendor and model choices, the adoption program, and the reporting your board actually wants. The seat ends when someone inside passes the test.
How is this different from hiring a big consulting firm?
Firms staff AI projects with people who have studied the technology. I have built with it: an AI platform launched as CEO, a generative AI product shipped to Fortune 500 teams, and daily hands-on work with the current tools. If you want a hundred-page strategy, hire the firm. If you want workloads in production and a smaller bill, we should talk. The engagement is sized to the company in front of me, from a 30-person business to a portfolio company, with a fixed fee for the first step.
What does the first engagement look like?
A free scoping call, then the fixed-fee AI Assessment: where your people spend repetitive hours, what each flow costs, and which two or three are the boring wins. You get the written plan in two weeks for a small business and four for a mid-market company. Then you choose: run it with me in an Adoption Sprint, with the fee credited, run it yourself, or shelve it. Each of the three is a good outcome. The person who scopes it builds it with your team, no handoff, deployed inside your own systems. I sell no software and take nothing from a vendor; if I ever advise a company I would recommend to you, I say so first.
Do you work with private equity firms?
Yes, across the ownership cycle: AI diligence on targets, post-close AI plans, portfolio-wide workload sweeps, and exit readiness. I spent a decade as a Managing Partner on the investing side, so the EBITDA lens is native, not adopted. The people I work with are the operating partner and the portfolio company's CEO and CFO, on the fund's clock. The question we answer together is which AI use cases move EBITDA inside the hold period, before the exit.
We already bought AI and the bill keeps growing. Can you fix that?
Usually, yes. Runaway AI spend is almost always a workload-fit problem: frontier models doing work smaller models handle, and nobody owning the tokenomics. Mapping the workloads honestly has cut spend by 80 percent on some, with the speed-critical paths moved to faster inference. Same outcomes, smaller bill.
How much of a process can AI actually automate?
My rule: any process you can model can be mostly automated, with exceptions routed to humans. In service, success, support, and claims work that typically means 30 to 70 percent off human queues. In software development, teams I have measured and surveyed run 3 to 17x on output, and the constraint moves to product and QA. The range is wide because execution is the variable, and execution is the work. Automation changes the work of the people on the queue, so they redesign it with us.
How is AI advisory different from digital transformation consulting?
Digital transformation covers the broader technology operating model: ERP, CRM, cloud, security, the org around them. AI advisory is narrower and faster: which recurring workloads can be modeled, which pay first, and getting them into production with an owner and a measured share of volume off the human queue. Both fail on adoption, so I run the people and process work in either. If your digital transformation stalled, that is a different page.
What does an AI readiness assessment cover, and when does a company need one?
Mine covers five questions per workload: where the data lives and who owns it, whether the process can be modeled, its volume and unit cost, what happens to the exceptions, and who inside will own it after I leave. You need one before you commit to a date or a vendor, and especially when the data sits across legacy systems. The output is a ranked roadmap, not a maturity score. It is the AI Assessment: two weeks for a small business and four for a mid-market company, at a fixed fee.
How long does it take?
The AI Assessment takes two weeks for a small business and four for a mid-market company. An adoption sprint puts one or two workloads into production inside one quarter, with owners and dates. Fractional AI leadership is an ongoing seat. Data readiness is what adds time, which is why the assessment asks the data question first, before a date is set.
Why do most AI pilots fail to reach production?
Because a pilot has no owner and no exception path. It ran on curated data with a person watching, and it ran beside the system the team works in rather than inside it; production has neither the curation nor the tolerance, and the first hard case stops the team trusting it. The fix is to skip the pilot stage: put the workload into production at a defined share of volume, route the exceptions to a named person, and measure the share of volume that no longer touches a human queue.
How do you measure whether AI worked?
Per workload: the share of volume that no longer touches a human queue, times what a unit of that volume cost before, against a baseline measured before the switch. Adoption is counted in volume, not in licenses or logins, read month over month on the monthly line to the board or the operating partner. The result is a line the CFO can audit and a buyer can diligence at exit, which is the number a board actually wants.
What happens to our data, and who can see what the models see?
Before a model touches customer, financial or employee data, one person in your company owns three documents: which models are running and on what data, what each one is allowed to read, and a log of what it decided. Your data stays inside the systems you already control; a model reads what the access rule permits and no more. The same three documents answer a buyer's diligence team at exit.
Which AI opportunities are worth pursuing?
Think through which use cases deserve investment, when to build or buy, and what adoption requires.
Free, every week, for CEOs and senior leaders. One operating problem, useful evidence and a practical next step. Topics rotate across leadership, strategy, offsites and AI.
Confirm your email to subscribe. Unsubscribe anytime.
Read now: the AI answer your board actually wants.
Bring me one workload, or the whole portfolio.
A forty-five minute scoping call, no pitch, for operating partners, CEOs and CFOs who want AI in production rather than another deck. Bring the AI question your board or investment committee is asking, or the workload you suspect is burning money. You leave with the shape that fits, an assessment, a sprint or the seat, the first workload we would start on and the number we would set out to beat; the scope and the fee follow in writing.
Or reach me directly:
(629) 304-8755
ken@kengavranovic.com
Prefer to write? Use the note: I reply with a scope and a number, or a straight no.