the AI transformation loop

In the last six weeks I've spoken with 20 PE operating partners about AI inside their portfolio companies.

The specifics are confidential, but the pattern repeats.

Most portcos use AI for email drafts and meeting notes, and most of the partners couldn't describe what comes after that.

The few pushing into claims, underwriting, or revenue operations hit 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.

Gallup asked 22,573 US employees in May 2026 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 in April 2026 that 18% of organizations have AI integrated into their workflows. The rest are drafting emails.

The companies that failed at this ran AI like a consulting program.

They ran a long diagnostic, drew a target architecture, projected ROI on assumptions, and spent 18 months on integration before anyone measured a dollar. The last step on that roadmap is a final presentation.

The companies that shipped ran it as a loop.

They picked one workflow with real users, real data, and an outcome the business tracks, proved the smallest useful slice in two to four weeks against a measurement they built before the agent existed, and then made a decision.

This is the loop, written down so an operating partner can hand it to a portco CEO instead of a deck.

Why the portco can't hand you a use case

The people running a portfolio company were running it before AI.

They have clients to service and relationships to protect, and those clients aren't asking for AI by next quarter, so the week looks like last week and the team has no reason to stop.

Ask them for a list of AI use cases and you get a shrug, because you've asked them to stop running the business to answer you.

Accordion's 2026 benchmark asked 150 operating partners about AI in their portfolio companies' finance functions.

17% of the portco CFOs had claimed AI as a strategic mandate with budget and board visibility. 34% were still evaluating, 31% were piloting specific workflows, 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 2026 survey of 782 infrastructure and operations leaders found the same shape from inside the company.

28% of AI use cases fully succeed and meet ROI expectations, and the 77% of leaders who had at least one success credited integrating AI into existing workflows and securing full support from business executives.

The operators know where the process hurts.

Microsoft's 2026 Work Trend Index, 20,000 workers, put 67% of AI's measured impact down to organizational factors rather than individual skill, and found only 26% of AI users say their leadership is aligned on an AI strategy.

The operators need a way to look at their own operation, and the sponsor needs to own the accountability. That is what the five lenses are for.

How to look at a portco: five lenses

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, and neither tells you where to start.

Start with the customer and work inward.

1. The customer interaction.

Find the interaction where the customer waits, complains, or leaves, and ask whether AI would make it better for them. The answer takes one of three shapes: an insight the customer didn't have, a process that runs in minutes instead of days, or a product that improves between releases. Onboarding is the interaction most companies underrate.

Encompass surveyed corporate treasurers in April 2026: 96% had abandoned a bank application because of how long it took, and 97% were considering switching banks over KYC friction.

OnRamp's 2026 onboarding report found 57% of customer-success leaders say onboarding friction hits revenue directly, and the teams that digitized onboarding cut time to value by 25% 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. 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 80% of an agent's success. If no one in the building can define the fields, defining them is the first deliverable.

4. The evidence.

Check whether you can measure the workflow today: unit of work, volume per period, cycle time, touch time, rework, exception rate, and cost per unit. Without a baseline you can't prove the change, and the sponsor won't fund a second workflow on a change you can't prove.

5. The owner.

Name who runs it after we leave: 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 doesn't qualify.

Skip a candidate if it fails any of these: no real users, no real data, no baseline, no owner, or the ask is "redesign the organization." Two categories are obvious and don't need the lenses: software development, which is construction with a gate at every step, and work so repetitive a script could do it. Past the obvious, run the lenses on three to five candidates, rank them by value, feasibility, risk, and what they'd let you reuse next, and pick one.

How to do it: ten steps

1. Pick one workflow, one sponsor, one outcome, and the non-goals. Choose real users, real data, and a number the business tracks, and write the non-goals down, because the team will add scope in week two.

2. Walk the workflow with the people who run it. Collect representative cases, the exceptions, the permissions, and the expert judgment no one wrote down. Interview before you inventory: ask who owns each dataset and what it means before anyone reads the code.

3. Sign the baseline before you build. Agree volume, cycle time, touch time, quality and rework, exception rate, operating cost, and the measurement window, and have both sides sign it before the pilot, because a baseline written afterward measures nothing.

4. Build the evaluation harness before the agent exists. Assemble a golden set of real cases with outcomes the client's own experts agreed on, and add the failure classes that matter in this operation: wrong numbers, protected data exposed, false confidence about missing data. You can't trust a measurement you build after seeing the answers.

5. 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 old way, and make sure every run leaves a trace someone can read.

6. Measure three numbers. Throughput is units completed per period and cycle time. Quality is first-pass acceptance, rework, overrides, and incidents. Economics is run cost per unit against verified capacity value, avoided rework, incremental output, and risk-adjusted avoided loss, minus cloud, model, support, and change costs. Report a range, and count hours saved as cash only when someone agrees where the cash shows up.

7. Decide together. Continue, adjust, stop, or expand on the evidence. If the numbers don't clear the baseline, stop and keep the budget. That outcome cost four weeks instead of four months.

8. Harden. Add evaluation on every change, a security review, observability, integration, fallbacks, cost caps, and a host that isn't someone's laptop. Put role-based access on every knowledge structure and a lifetime on every agent, so a dead agent doesn't survive as tech debt with a live API key.

9. Transfer ownership. Embed the capability in the workflow, train the operators, hand over the runbooks, and name the owner. Operators stop pressing buttons and start judging, and every correction they make goes back into the golden set.

10. Pick the next workflow with what you now have. The context layer, the harness, the hosting, and the governance all carry over, so each workflow reuses more of the last. That is the ladder from tool adoption to an org that runs on agent output, and you climb it one proven workflow at a time.

The ladder has seven rungs, shared below.

AI Adoption Ladder

There are seven main steps for your company to go from tool adoption to agentic org.

Step one: tool adoption. AI tools exist across the org, humans run each one. Developer closes the lid, automations die. Somebody goes on vacation, the workflow goes with them.

Step two: context layer. Knowledge graphs per department, each behind its own API gateway. Sales queries engineering through an API. Separated domains keep agents from hallucinating across departmental boundaries.

Step three: harness engineering. Build evaluation criteria, routing logic, and quality gates before you pick a model. The LLM comes in last. Your harness is your IP and your capex. The model underneath is a commodity you swap when something better ships. If changing the LLM breaks your system, you built a dependency, not a platform.

Step four: go headless. Host every automation outside a laptop. It fires on schedule, on conditions, without a human logging in at 9 AM. Companies stall here because nobody forces the migration from "script on my desktop" to "service that runs on its own."

Step five: governance. RBAC on every knowledge structure from day one. TTL on every agent. Dead agents are tech debt with active API keys. Past 100 tools, you need lifecycle management. Retire anything that hasn't run in 90 days. Deprecate anything whose eval scores drop below threshold.

Step six: agent-to-agent. Stitch headless instances together within a value chain. Agent A finishes, hands off to Agent B. An LLM sits between them as a quality gate, judging every handoff before it passes through.

Step seven: step functions. Orchestrated, semi or fully automated workflows. Agents handle volume. Humans handle exceptions. The org runs on agent output.

Most companies sit on step one, shopping for step seven technology. They need step four discipline.

What it takes from the portco

The proof month asks for an executive sponsor for 30 minutes a week and a go or no-go at the gate, a process owner for two to four hours a week in walkthroughs and acceptance testing, and a data owner for one to three hours at the start.

Our AI velocity pod is one FDE with agents plus a fractional architect, at $15,000 to $20,000 a month.

On the last operations agent we shipped, the infrastructure ran at $250 a month across four client datasets.

Our terms are the same for every portco: no discovery fee, month to month, and if you're not impressed at the end of the month, you don't pay.

Every activity in the month produces one of four things: a decision, a tested assumption, a working increment, or a reusable asset.

If it can't, it doesn't get your time.

What the loop has produced

A healthcare billing platform showed us how they worked before we touched anything. Their clients ask one question about claim denials more than any other, and the answer has to reconcile to the penny.

Six weeks from kickoff their chief product and technology officer was running acceptance tests against an agent that answers it. It scored 59 of 60 on the eval set at first deploy and 15 of 15 on the denial cases.

After a live schema migration it held 97% correctness, and the flagship question reconciled to the penny. The golden set came from the client's own questions, with their experts as the ground truth.

The agent advises, a person decides, and it writes to no system of record.

On a logistics platform's billing rebuild, two engineers merged 122 pull requests in the first 90 days, 104 of them with fewer than five reviewer comments, and delivered 50% ahead of the client's own estimate. Software is the obvious case, and we ran the loop there too: no code in week one, a spec before every feature, a human gate before every merge.

A hardware manufacturer's CTO ran our four-week pilot next to a second pilot from a firm about ten times our size. Four weeks in, he said ours was doing the better job, and he set the next phases himself: 50/50 with his people embedded, then 80/20 across his other ten Scrum teams.

The honest limits

The 2026 numbers on AI payback are consistent, and they are not kind.

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's June pulse of 2,145 leaders found 49% had scaled back, narrowed, delayed, or paused agent deployments after costs outran value, and 7% had established ROI.

Gartner's April survey named the reason most leaders gave for their failures: they expected too much, too fast.

Bain's 2026 Global Private Equity Report sets the clock. A deal now needs 10 to 12% annual EBITDA growth to return 2.5x over five years, against 5% in the 2010s, and holding periods at exit hover around seven years. In FTI-Andersch's April survey of portfolio AI programs, 38% expected returns within 7 to 12 months and 31% within 13 to 24.

In Bain and StepStone's outlook, 39% of general partners expect no material financial impact from AI on their portfolio companies this year. A business does not transform in one to three months, and a sponsor who plans for that will fund a deck.

Two more limits.

Software is where the evidence runs deepest, because software development is construction: requirements in, tested code out, a gate at every step.

Our own pod on the billing rebuild merged 122 pull requests in 90 days and delivered 50% ahead of estimate.

The same 2026 data shows the price of skipping the gate.

Faros AI's April telemetry across 22,000 developers found task throughput up 33.7% while bugs per developer rose 54% and incidents per pull request rose 242.7%.

Outside software, the obvious case is work so repetitive a script could do it, like support answering the same questions, and even the obvious case plateaus.

Past the obvious cases there are no obvious answers, and anyone promising you one across eleven portfolio companies hasn't done it.

The loop compounds because you keep what each workflow builds.

After each proven workflow you keep the context layer, the harness, the hosting, and the owner, and you start the next one higher.

Models will change, and the vendor will deprecate this quarter's framework next quarter.

The workflow knowledge, the evaluation criteria, and the operating ownership your team builds while proving the first slice stay with you.

Instead of the deck

Hand the portco CEO this page.

It has five lenses to find the workflow, ten steps to prove it, and one decision at the end.

Book a free discovery call

One 30-minute call, not a discovery quarter.

Bring the workflow you'd start with, or the portfolio company you can't get moving. We run the five lenses on it live and tell you whether it qualifies.

If it does, the month starts on our terms: no discovery fee, month to month, and you don't pay if you're not impressed.

Book a discovery call: https://meetings-na2.hubspot.com/mark-ajzenstadt

— Mark Ajzenstadt, Founder @LimestoneHQ