Reading time: 6 minutes | Issue #41 | Book a Discovery Call
Happy Tuesday. Mark here.
On August 27, Clearlake Capital and Google Cloud announced a deal to push full-stack AI across Clearlake's portfolio: Gemini Enterprise for agentic workflows, TPUs and GPUs underneath, Vertex AI for model choice, wired into Clearlake's O.P.S. operating framework and its in-house AI Labs. At more than $185 billion in assets under management, this is serious people making a serious move, and I read the release twice.
It also doesn't touch the hard part. A portfolio can have Gemini Enterprise, the silicon, and a data platform in hand and still be stuck at the question every portco I talk to is stuck at: which workflow, and proven how. The stack is the easy purchase, and the transformation is the loop you run on top of it, one workflow at a time. So this issue is that loop, written down, the way you would hand it to a portco CEO instead of a deck.
Inside the Issue
The five lenses that turn "we don't know what we don't know" into one workflow worth proving
The ten-step loop you can start Monday, plus the seven-rung ladder every workflow climbs
Clearlake buys the Google stack, AI agents delete production databases, PE's AI job-cut math, and the operating-partner role that didn't exist two years ago
Five Lenses to Find the One Workflow Worth Proving
The specifics of those conversations are confidential, so you get the pattern, not the names. In the last six weeks I spoke with 20 PE operating partners about AI inside their portfolio companies. Most of those portcos use AI for email drafts and meeting notes, and the few pushing into claims, underwriting, or revenue operations keep hitting the same wall: a demo takes a week, and proving it runs at production quality takes six months.
The survey data says the same thing from the outside. Gallup asked 22,573 US employees in May what they use AI for: writing and editing came first at 51%, search at 49%, and automating a process at 16%. Harvard Business Review Analytic Services found that only 18% of organizations have AI integrated into their workflows rather than sitting beside them as a standalone tool. Most of the building is still drafting emails.
The reason a portco can't hand you a use case is not that the team is behind. The people running a portfolio company were running it before AI showed up. They have clients to service and relationships to protect, and those clients are not asking for AI by next quarter, so the week looks like last week. Ask the team for a list of AI use cases and you get a shrug, because you have asked them to stop running the business to answer you.
The numbers show how early this still is. Accordion surveyed 150 operating partners in May about AI in their portcos' finance functions. 17% of the portco CFOs had claimed AI as a strategic mandate with budget and board visibility. 31% were piloting specific workflows, 34% were still evaluating, and 41% of the partners said they were scaling AI across the portfolio with no operational playbook. The barriers they named were data infrastructure at 71%, talent at 63%, and legacy technology at 58%. Gartner's April report, from 782 infrastructure and operations leaders it surveyed late last year, found only 28% of AI use cases fully succeed and meet ROI expectations. The 77% of leaders who landed at least one success credited two things: integrating AI into existing workflows, and full support from business executives. Among the failures, the reason leaders gave most often was that they expected too much, too fast.
Our take: the operators know where the process hurts, and the sponsor has to own the accountability. That is what the five lenses are for. You go looking in one of two ways: you cut cost around the slow people, or you make the smart people faster. Both stay inside the building, though, and neither tells you where to start, so start with the customer and work inward.
1. The customer interaction. Find the moment where the customer waits, complains, or leaves, and ask whether AI would make it better for them, not just cheaper for you. Onboarding is the interaction most companies underrate. Encompass's 2026 survey of corporate treasurers found 96% had walked away from a bank application over how long it took, and 97% were weighing a switch over KYC friction. OnRamp's 2026 report found 57% of customer-success leaders say onboarding friction hits revenue directly, and the teams that digitized it cut time to value by a quarter or more.
2. The work behind it. Walk the process and sort the steps: deterministic, judgment-heavy, exception-driven, or relationship-dependent. Machines take the deterministic steps and prepare the judgment ones, while people keep the decisions, the approvals, and the exceptions. If most of the work is judgment or relationship, save it for a later workflow.
3. The context. Find where the truth lives, who owns it, and what it means. Every dataset in the workflow needs an owner, a definition, and a sensitivity level before an agent sees it. Our delivery team puts context at roughly 80% of an agent's success. On a healthcare billing build this year, the first deliverable was a definition, not a model: more than 300 denial codes that two people had been carrying in their heads. If no one in the building can define the fields, defining them is the work.
4. The evidence. Check whether you can measure the workflow today: unit of work, volume, cycle time, touch time, rework, exception rate, cost per unit. Without a baseline you cannot prove the change, and the sponsor will not fund a second workflow on a change you could not prove.
5. The owner. Name who runs it after the delivery team leaves: a process owner, a data owner, and an executive sponsor who will make a go or no-go call on evidence. A workflow with no owner does not qualify, however good the demo looks.
Run the five on three to five candidates, rank them by value, feasibility, risk, and what each one lets you reuse next, and pick one. Skip any candidate that fails a lens: no real users, no real data, no baseline, no owner, or an ask that amounts to "redesign the organization." Two categories skip the lenses because the answer is already yes: software development, which is construction with a gate at every step, and work so repetitive a script could do it.
Our read on the timeline, because someone on the board will ask. The payback data this year is consistent, and it is not flattering. McKinsey's August State of AI puts enterprise EBIT impact at 37% of companies, flat on the year, while 80% report individual productivity gains. KPMG found 49% of leaders had scaled back, narrowed, delayed, or paused agent work after costs outran value, and 7% had established ROI. That gap between individual gains and enterprise impact is the argument for a loop, not against one. Bain now says a deal needs 10 to 12% annual EBITDA growth to return 2.5 times over five years, where 5% did the job in the 2010s, and holding periods sit near seven years. FTI's 2026 PE AI Radar found 38% of firms expect returns inside 7 to 12 months. A business does not transform in one to three months, and a sponsor who budgets for that will fund a deck.
Who should be uncomfortable reading this: any operating partner whose AI plan is a portfolio-wide platform license and a target architecture, with no named workflow, no baseline, and no one who owns the outcome.
The Ten-Step Loop, and What It Costs the Portco
Once you have the one workflow, this is the loop that takes it from demo to a measured decision in weeks. The discipline is in the sequence, not the tooling.
Frame it. One workflow, one sponsor, one outcome the business already tracks, and the non-goals written down, because the team will try to add scope in week two.
Walk it with the people who run it. Collect real cases, the exceptions, the permissions, and the expert judgment no one wrote down. Interview before you inventory.
Sign the baseline before you build. Volume, cycle time, touch time, quality, rework, exception rate, cost per unit, and the measurement window, agreed and signed by both sides. A baseline written afterward measures nothing.
Build the evaluation harness before the agent exists. A golden set of real cases with answers the client's own experts agreed on, plus the failure classes that matter here: wrong numbers, protected data exposed, false confidence about missing data. You cannot trust a measurement you build after seeing the answers. This is the step teams skip most, and it is the step that decides whether the numbers in step six mean anything.
Ship the smallest useful slice on real data in two to four weeks. The agent proposes, a named person decides, and it writes to no system of record yet. Run it in parallel with the current way, and make every run leave a trace someone can read.
Measure three numbers. Throughput (units cleared without a human), quality (first-pass acceptance, rework, overrides, incidents), and economics (run cost per unit against verified value, minus model, cloud, support, and change costs). Report a range, and count hours saved as cash only when someone agrees where the cash shows up.
Decide together. Continue, adjust, stop, or expand, on the evidence. If the numbers do not clear the baseline, stop and keep the budget. That decision cost four weeks, not four months.
Harden it. Evaluation on every change, a security review, observability, fallbacks, cost caps, and a host that is not someone's laptop. Role-based access on every knowledge structure, and a lifetime on every agent, so a dead agent does not survive as tech debt with a live API key.
Transfer ownership. Embed the capability, train the operators, hand over the runbooks, name the owner. Operators stop pressing buttons and start judging, and every correction they make goes back into the golden set.
Pick the next workflow with what you now have. The context layer, the harness, the hosting, and the governance carry over, so each workflow reuses more of the last.
What it costs the portco. The proof month asks for an executive sponsor about 30 minutes a week plus a go or no-go at the gate, a process owner two to four hours a week in walkthroughs and acceptance testing, and a data owner one to three hours at the start. Our AI velocity pod is one forward-deployed engineer with agents plus a fractional architect, at $15,000 to $20,000 a month. On the last operations agent we shipped, infrastructure ran $250 a month across four client datasets.
One test keeps the month honest: every activity produces a decision, a tested assumption, a working increment, or a reusable asset. If it cannot, it does not get your time.
01 AI agents are deleting production systems. New research from StackGen, drawn from nearly 178,000 public status-page records, finds AI now factors into more than 1 in 10 reported tech incidents, six times its share in 2023, and documents at least nine cases of autonomous agents wiping live data or whole systems on their own. Median time to resolve has not improved since 2023. This is rung five of the ladder as a headline: role-based access and a lifetime on every agent are not governance theater, they are how you keep a helpful agent from becoming an unsupervised one.
02 The AI operating partner is now a job title. Korn Ferry's institute has been tracking a role that did not exist at most firms two years ago: a partner whose whole job is AI value creation across the portfolio, separate from the technology operating partner. It maps to the fifth lens. The firms getting results named an owner for it, rather than adding "AI" to someone's existing plate.
03 The other side of the portfolio-AI trade. The PE Stakeholder Project, a watchdog group, tallied this year's wave of PE and AI-vendor deals and put the reach at more than 2,000 portfolio companies and millions of jobs, arguing the incentives point straight at headcount. I run an AI delivery firm, so read my bias into this. The honest version of our pitch still includes the part where "automate the deterministic steps" is a real person's week. The loop keeps humans on the decisions and the exceptions by design, and that is a choice each operator has to make on purpose.
04 Even the obvious case has a gate. Faros AI's 2026 report, two years of telemetry across 22,000 developers, found task throughput up 33.7% as AI adoption deepened, while bugs per developer rose 54% and incidents per pull request rose 242.7%. Software is the clearest place to run the loop, and the clearest proof of what skipping the human gate costs.
05 GPs are resetting their own expectations. In Bain and StepStone's 2026 GP Outlook, 39% of general partners said they expect no material financial impact from AI on their portfolio companies this year, with most of the rest expecting cost savings ahead of revenue growth. Read next to the firms selling portfolio-wide transformation, that is a useful gap to sit in.
The AI Adoption Ladder. Seven rungs from tool adoption to an org that runs on agent output. Find the rung your portco is actually on before you buy anything built for the rung you wish it were on.
Tool adoption. AI tools exist across the org, a human drives each one. Someone goes on vacation and the workflow goes with them.
Context layer. Knowledge per department, each behind its own gateway. Separated domains keep agents from hallucinating across boundaries.
Harness engineering. Evaluation criteria, routing, and quality gates built before you pick a model. The harness is your IP. The model under it is a commodity you swap when a better one ships.
Headless. Every automation runs off the laptop, on a schedule or a trigger, with no one logging in at 9am to start it. Most companies stall right here.
Governance. Role-based access on every knowledge structure, a lifetime on every agent, and retirement for anything that has not run in 90 days. Dead agents are tech debt with live API keys.
Agent-to-agent. Headless instances stitched into a value chain, with a model between them as a quality gate on every handoff.
Step functions. Orchestrated workflows where agents handle volume and humans handle exceptions. The org runs on agent output.
Most portcos sit on rung one, shopping for rung seven. They need rung four discipline.
If you have one workflow where the outcome is measurable and the current answer is still a demo, that is the one to bring. On a 30-minute discovery call I will run the five lenses on it live and tell you whether it qualifies as a first slice and what its evaluation harness would look like. If it does, the month runs on our terms: no discovery fee, month to month, and you do not pay if you are not impressed at the end of it.
Until next Tuesday,
— Mark Ajzenstadt, Founder @ Limestone Digital
