Operating partners have a math problem. A typical hold period is five years. If your AI program spends year one on strategy decks, year two on a stalled data-lake project, and year three on a pilot that never leaves the lab, you exit with a slide about "AI initiatives" and zero EBITDA to show for it. Buyers have seen that slide. They discount it to zero.
The firms getting real value out of AI agents in portfolio companies are doing something less impressive-sounding and more effective: they pick one workflow inside one company, automate it under governance, prove the number moved, then run the same play at the next company. BCG's work on AI-first companies in private equity makes the point bluntly: value comes from redesigning how work gets done in the portfolio, not from adding AI features at the edges.
This is the 100-day plan we use for that. It assumes a mid-market portfolio company, a lean ops team, and an operating partner who needs numbers before the next board meeting.
Why Portfolio Companies Beat Deal Teams as the Starting Point
Most PE firms start their AI journey at the fund level: deal sourcing agents, diligence summarization, CIM screening. It feels natural because the deal team is closest to the decision makers.
It is usually the wrong first move, for three reasons.
First, the money is in the portfolio. A sourcing agent that surfaces 10% more targets improves an already-functioning process at the margin. An agent that cuts invoice-dispute resolution time by 60% at an industrial services portco shows up directly in working capital and headcount, and it survives diligence at exit because the numbers are in the financials.
Second, portfolio workflows repeat. Your fund does one deal process. Your portfolio companies collectively run thousands of instances per month of order intake, support tickets, collections calls, and vendor invoices. Agents get better with volume and feedback; deal workflows give them neither.
Third, replication is the structural advantage PE has over everyone else. A standalone mid-market company that automates its AP workflow gets one win. A firm with twelve portfolio companies that packages the same workflow gets twelve, and each one after the first is cheaper. Glean's analysis of the AI value gap in private equity finds the gap between AI leaders and laggards in PE comes down to exactly this: operational deployment discipline, not model access. Everyone has the same models now.
The takeaway: treat the fund-level use cases as a later phase. The first 100 days belong to one portfolio company and one workflow.
Days 1-30: Baseline the Workflows and the Data
The first month produces no automation. That is deliberate, and it is the part most programs skip. You cannot claim a 40% cycle-time improvement at exit if nobody measured the cycle time before you started.
The deliverable for days 1-30 is a ranked workflow inventory for the target company, with baselines. For each candidate workflow, document:
- Volume: transactions or cases per month
- Fully loaded cost: hours spent times loaded labor rate, plus error and rework cost
- Cycle time: request to resolution, measured from system timestamps, not from what the team estimates
- Data condition: where the inputs live, how structured they are, and whether the systems have usable APIs
- Owner: one named person accountable for the workflow today
Focus the inventory on three functional areas, because that is where mid-market portcos concentrate manual cost:
| Function | Typical candidate workflows | Why it ranks well |
|---|---|---|
| Finance operations | Invoice processing, dispute resolution, collections follow-up, expense audit | Digital inputs, checkable outputs, direct working-capital impact |
| Customer operations | Tier 1 support triage and drafting, order status, returns handling | High volume, measurable deflection and CSAT baselines |
| RevOps | CRM hygiene, lead enrichment, follow-up sequencing, pipeline data quality | Bad data here quietly taxes every downstream initiative |
Two filters matter more than any scoring rubric. The workflow's inputs must already be digital (if the process starts with a phone call and a sticky note, fix that first), and its outputs must be checkable by a human in seconds, not minutes. We wrote up the full scoring approach in our AI workflow readiness scorecard, and the interview technique for finding workflows the org chart hides in workflow discovery for AI automation.
Expect an uncomfortable finding: at most mid-market portcos, nobody can answer "what are your three most expensive manual workflows" on day one. The finance leader knows headcount by department, not cost by workflow. Producing that answer is the month-one value, independent of anything you automate later.
One more thing belongs in this month: pick the governance scaffold. For companies without a compliance function, the NIST AI Risk Management Framework is a practical starting point. You do not need all of it. You need its core habit: for each deployed system, document what it does, what can go wrong, who owns it, and how you would know it is failing.
Days 31-60: One Governed Pilot, With Exit Criteria in Writing
Month two is where discipline gets tested, because the temptation after a good inventory is to launch four pilots. Launch one.
A single pilot concentrates the scarce resources that actually determine success: the attention of the workflow owner, the patience of the operators whose work is changing, and the ops team's capacity to review edge cases weekly. Four parallel pilots split that attention four ways and typically produce four demos and zero production systems.
The pilot needs four things agreed in writing before anything is built:
- A baseline number from month one that the pilot is trying to move
- Pass/fail criteria - the specific thresholds on accuracy, cycle time, and escalation rate that trigger a production decision, a redesign, or a kill
- Human approval points - which agent actions execute automatically and which queue for a person; anything touching money, customer commitments, or external communications starts in approval mode
- A review cadence - a standing weekly session where the owner and the build team walk failure cases
We covered how to set the thresholds in pass/fail criteria before AI rollout. The short version: define "good enough to ship" before you see any output, because after you see output, motivated reasoning takes over.
On KPIs, resist dashboard sprawl. Dynamiq's guide to AI agent KPIs for private equity lists dozens of plausible metrics; for a 30-day pilot you need about five:
| KPI | What it tells you |
|---|---|
| Task accuracy vs. human baseline | Whether the agent's output is usable |
| Escalation rate | How often it correctly hands off instead of guessing |
| Cycle time delta | The number that shows up in operational reviews |
| Cost per transaction | The number that shows up at exit |
| Override rate | How often humans reject approved-mode actions; your early-warning signal |
A concrete example of what this looks like when it works. At one industrial services company (sub-$200M revenue, the classic mid-market profile), the pilot was invoice dispute handling: an agent that reads the disputed invoice, pulls the contract terms and delivery records, drafts a resolution, and queues it for AP approval. Baseline: 11 days average resolution, roughly 2.5 hours of staff time per dispute. After 30 days of piloting with every resolution human-approved: 4 days average resolution, 20 minutes of staff time, 8% of drafts rejected at approval. That 8% override rate was the most valuable number produced, because each rejection was a documented edge case that fed the next iteration.
Notice what the pilot did not do: it did not act autonomously, did not touch payment execution, did not email customers. Scope discipline is what makes the production decision easy.
Days 61-100: Production, Then Portfolio Replication
If the pilot cleared its gates, month three has two workstreams that matter equally: hardening the pilot into production, and packaging it so the next portfolio company does not start from zero.
Production hardening is unglamorous and specific:
- Move from "human approves everything" to graduated autonomy: low-risk, high-confidence cases execute automatically; everything else still queues
- Wire monitoring and alerting to the KPIs from the pilot, so drift shows up in a dashboard rather than in a customer complaint
- Document failure handling: what happens when the ERP API is down, when the agent's confidence is low, when volume spikes
- Train the operators and the reviewers, and name a permanent owner inside the portco who does not work for the PE firm
That last point deserves emphasis. An agent owned by the sponsor's ops team is a liability at exit; buyers want to see that the capability lives inside the company. Ownership handoff is a deliverable, not an afterthought.
The replication workstream is where the firm-level return comes from, and it is what makes this plan a private equity strategy rather than a single-company IT project. Package three artifacts from the deployment:
- The workflow pattern: process map, agent design, integration approach, and the prompts and tool configurations, written for reuse
- The controls kit: the governance doc, approval-point design, KPI definitions, and pass/fail gate template, which transfer almost unchanged between companies
- The baseline method: the month-one measurement approach, so the next company's numbers are comparable
Here is the counter-intuitive part: the reusable asset is mostly the controls, not the code. Portfolio companies run different ERPs and different helpdesk tools, so integrations get rebuilt. What transfers cleanly is the harder-won material: knowing which approval points matter, which KPIs predict failure, and which edge cases show up in week three. In our experience the second deployment of a packaged workflow runs 40-60% cheaper than the first, and the discount comes almost entirely from not re-learning governance and evaluation.
This mirrors what the broader market data shows about why AI programs stall. McKinsey's State of AI research has consistently found that a small minority of adopters capture most of the value, and the differentiator is workflow redesign and governance maturity rather than model choice. A PE firm that packages governance as a portfolio asset is building exactly the muscle the laggards lack.
At day 100 the honest scoreboard looks like this: one production workflow with a measured EBITDA or working-capital impact, one replication kit, and a ranked backlog of the next two workflows at the first company plus the first workflow at a second company. That is a smaller claim than "AI transformation." It is also one you can defend in an exit process, which Bain's Global Private Equity Report suggests is becoming table stakes as buyers start diligencing AI claims the way they diligence revenue quality.
What Kills These Programs
Three failure modes account for most of the wreckage we see when firms bring us programs that stalled.
The platform-first purchase. The firm signs a portfolio-wide license for an AI platform, announces it at the annual portco summit, and discovers that a license is not a deployment. Twelve companies get logins; two get value. Tools follow workflows, not the other way around.
The unowned pilot. A pilot sponsored by the deal team, tolerated by the portco, and owned by nobody. When the champion at the fund rotates to a new deal, the pilot dies. Every workflow needs an owner whose bonus depends on the number the agent is supposed to move.
Autonomy before evidence. Skipping approval mode because the demo looked clean. The first customer-facing mistake then burns the political capital the whole program runs on, and the portco CEO quietly shelves the initiative. Graduated autonomy is slower and it is the only path that survives contact with a bad week.
None of these are technology failures. All of them are sequencing failures, which is why the 100-day structure matters more than the model or vendor you pick.
How OpenNash Can Help
OpenNash builds production AI agents for exactly the workflow types this plan targets: finance operations, customer operations, and RevOps inside mid-market companies. For PE operating teams, we map directly onto the plan's phases: the audit produces the day 1-30 workflow inventory and baselines, design defines the approval points and pass/fail gates before anything is built, the build phase runs the governed pilot with weekly edge-case reviews, and deployment ends with full ownership handoff to the portfolio company, including documentation a buyer can diligence.
To be fair about the alternatives: if a portfolio company's need is a standard, well-served workflow on a system it already runs (Zendesk AI for basic deflection, an ERP vendor's built-in AP automation), a platform feature is often the faster answer. Custom agents earn their keep when the workflow crosses systems, needs firm-specific controls, or needs to be replicated across a portfolio with an auditable pattern. And if a company cannot pass the readiness baseline (no digital inputs, no workflow owner), the right move is to wait and fix the prerequisites first.
If you operate a portfolio and want to know which company and which workflow should go first, book a call for a portfolio workflow opportunity scan. We will walk the inventory method against two or three of your companies and give you a ranked starting point, whether or not you build with us.