AI for Customer Service in PE Portfolio Companies

1. Support Is the Most Proven AI Use Case

If you want one place where AI in a portfolio company is not speculative, it is customer support. It is the use case with the cleanest academic evidence, the most mature vendors, and the fastest adoption. In Zapier's 2026 survey, customer support teams had the highest AI-agent deployment rate of any department.

That maturity is a double-edged thing. It means the wins are real and you are not a guinea pig. It also means the space is full of vendors promising to deflect ninety percent of your tickets, and some of that promise is a way to lose customers quietly while the cost line looks great.

So the job is not to decide whether AI belongs in support. It does. The job is to capture the cost saving without giving back the customer relationship, and that is a more interesting problem than the vendors make it sound.

2. The Three Ways AI Shows Up in Support

1. Agent assist. AI sits beside the human, drafting replies, surfacing the right answer, and summarizing the case. The human stays in control. This is the lowest-risk, highest-floor option, and it is where most companies should start.

2. Deflection. AI handles the contact end to end, with no human, for the routine cases. This is where the big cost savings live and also where the risk lives. Done well it resolves the simple things instantly. Done badly it traps customers in a loop they cannot escape.

The third way is the quietest: visibility. AI reads every conversation and tells managers what customers are actually calling about, which issues are rising, and where the product or process is failing. Most companies ignore this and it is often worth more than the deflection, because it feeds the rest of the business.

3. The Numbers, and What They Really Say

The foundational study is Brynjolfsson, Li, and Raymond's work on more than five thousand support agents. The headline is a 14 percent average lift in issues resolved per hour. The more useful finding is underneath it.

The gain was largest for the least experienced agents, who improved by about 34 percent, while the most experienced barely changed. AI worked by spreading the patterns of the best agents to everyone else. The study also recorded better customer sentiment and higher employee retention, which matters because it means the productivity did not come at the customer's expense.

Read that carefully before you model savings. The lift is real, it is uneven, and it came alongside happier customers, not in spite of them. That is the result to aim for, and it is not the same as maximum deflection.

4. Cost-to-Serve: The EBITDA Line

For a portfolio company, support is a cost center with a number attached: cost-to-serve, the fully loaded cost of handling a customer contact. AI moves that number two ways. It makes each agent more productive, so the same team handles more volume, and it deflects the routine contacts so they never reach an agent at all.

In a business with high contact volume (consumer services, subscriptions, anything with a call center) this is a direct margin lever. Fewer agents per thousand customers, or the same agents serving a growing base without new hires. In a roll-up of contact-heavy businesses, it compounds across every site, which is why support is often the first shared-services function a platform standardizes.

The discipline is to measure cost-to-serve against the baseline and against CSAT at the same time. A cost-to-serve number that falls while satisfaction holds is value creation. A cost-to-serve number that falls while satisfaction drops is a future revenue problem you have not booked yet.

5. Agent Assist vs Full Deflection

The central choice is how much to automate. The two ends of the spectrum behave very differently, and the right answer is usually a deliberate mix, not a maximum.

The mistake is to jump to full deflection because it shows the best cost number in the model. Start with assist, learn which contacts are genuinely routine, deflect those, and keep a fast, obvious path to a human for everything else. The path to a human is not a failure of the system, it is the thing that protects the brand.

6. CSAT: The Constraint That Stops You Going Too Far

Customer satisfaction is the constraint that keeps support AI honest. Without it, deflection optimizes for cost and quietly destroys the customer relationship, and the damage shows up two quarters later as churn nobody connected to the support change.

There is a real demand signal in favor of AI support when it is good: surveys find younger customers increasingly prefer a fast AI answer to hunting through help articles. But the same customers turn fast when the AI cannot help and will not let them out. The difference between the two experiences is the escape hatch and the honesty about what the AI can and cannot do.

So treat CSAT, or its cousin in your business, as a hard gate on every deflection decision. If satisfaction drops when you automate a contact type, that contact type was not ready, no matter what it did for cost-to-serve.

7. The Vendors and the Build Question

The support-AI vendor market is mature and crowded. The serious categories: dedicated AI support platforms (Sierra, Decagon, and the AI tiers of Intercom and Zendesk), the support features inside the CRM the company already runs (Salesforce, ServiceNow), and custom builds on a foundation model for companies with unusual needs.

For most portfolio companies the answer is buy, not build. Support is a solved enough problem that a specialist platform will beat an in-house project on cost and time, and the workflow is similar enough across companies that you are not giving up a real edge. Build only where support is genuinely your differentiator or your data is genuinely unique.

The cross-portfolio question is different. A platform with several contact-heavy companies may standardize on one support stack to capture the shared-services saving, covered in the multi-site services guide. Vet any vendor against confidential customer data using the security and governance guide.

8. Where Support AI Backfires

Deflection without an escape hatch. The single most common failure. A customer with a real problem trapped in a bot that cannot help and will not transfer. Cost falls, churn rises, and the two are never connected on a dashboard.

Automating the wrong contacts. The complex, emotional, high-value contacts are exactly the ones a human should keep. Automating those to hit a deflection target is how you lose your best customers at your worst moments.

Ignoring the visibility goldmine. Companies chase deflection and never read what the AI learns from every conversation. The pattern of why customers contact you is product feedback, churn warning, and pricing signal in one, and most leave it on the floor.

9. Measuring It

Measure cost and quality together, always. Cost-to-serve, contacts deflected, and agent productivity on one side. CSAT, resolution rate, and escalation rate on the other. A win moves the first set without hurting the second.

Watch the downstream number too: retention. Support is where churn signals first appear, so a support change that quietly raises churn is the most expensive kind of false saving. Connect the support dashboard to the retention number, which ties this lever to the revenue-growth side.

Hold it to the EBITDA standard. Contacts deflected is an input. Cost-to-serve down with CSAT held, and retention intact, is the result that belongs in a value creation plan.

10. Where to Start

Start with agent assist on the highest-volume contact type, and turn on the visibility from day one. Assist gives you a fast, low-risk productivity win and teaches you which contacts are genuinely routine. Visibility tells you where to deflect next and feeds the rest of the business.

Only then move selected, proven-routine contacts to full deflection, with a hard CSAT gate and an obvious path to a human. Measure cost-to-serve and CSAT side by side through the whole thing. That sequence captures most of the saving with little of the risk.

If you want help sizing the opportunity and picking the stack, our ROI Calculator runs the cost-to-serve numbers, and a Discovery Sprint maps support AI against your portfolio company's volume, vendors, and CSAT baseline.