What Does 78% Net Revenue Retention Mean for a Deal?

A company reporting 78% net revenue retention (NRR) is retaining less revenue from its existing customer base than it started with—even before adding new customers. For a buyer, investor, or board member, that is not a minor operating metric. It is a warning about the durability of recurring revenue, the cost of replacing churned dollars, and the assumptions behind the deal’s valuation multiple.

But 78% does not tell the whole story by itself. The number needs context: how it is calculated, what is driving the decline, whether the customer base is concentrated, and whether management has a credible plan to improve it. In an AI business, the analysis may also need to distinguish between customers who are simply accessing a model and customers who depend on a workflow that reliably delivers a business outcome.

This guide explains how to interpret 78% NRR in a deal, what questions to ask, and why product architecture—including multi-model orchestration—can matter to retention economics.

First, What Does 78% NRR Actually Mean?

NRR measures how recurring revenue from a defined starting group of customers changes over a period, commonly 12 months. It typically includes expansion, contraction, and churn, while excluding revenue from new customers acquired during the period.

A simplified formula is:

NRR = (starting recurring revenue + expansion − contraction − churn) ÷ starting recurring revenue

If a company begins the year with $10 million in recurring revenue from its existing customers and ends with $7.8 million from that same customer group, its NRR is 78%. The lost $2.2 million reflects the net effect of cancellations, downgrades, and any expansions.

That means the company must replace 22% of its starting revenue each year just to keep the existing customer cohort at the same dollar level. New customer bookings may still drive reported growth, but they do not erase the underlying weakness in the installed base.

Why 78% NRR Creates NRR Risk

At 78%, the business has meaningful NRR risk: revenue from existing customers is shrinking quickly enough that customer acquisition must work hard just to offset the decline. This changes the quality of growth. A company growing its total revenue through new logos may appear healthy on the surface while spending heavily to replace revenue that is leaking out the back door.

The effect compounds. If the same cohort contracts at that rate each year, and there is no expansion, only about 37% of its original revenue remains after five years. That illustration is not a forecast—the actual result depends on cohort composition and business changes—but it shows why retention economics matter over a multi-year investment horizon.

Question Why it matters in a deal Is 78% gross or net retention? Gross retention excludes expansion and shows how much revenue is lost before upsells offset it. Is the metric logo-based or revenue-based? A company can retain many small customers while losing a few large accounts—or the reverse. Is 78% a recent result or a long-term pattern? A single period can be distorted by renewals, one-time events, or changes in the customer mix. What is driving the decline? Product gaps, weak onboarding, pricing, customer concentration, and poor-fit acquisition have different remedies.

What It Can Do to the Valuation Multiple

A valuation multiple is not determined by NRR alone. It also reflects growth, gross margin, market opportunity, customer concentration, profitability, and the confidence buyers have in the numbers. Still, 78% NRR usually puts pressure on the multiple because it makes future recurring revenue less dependable.

Two companies with the same current revenue can have very different economic profiles. One may have stable renewals and strong expansion; the other may need a steady stream of new customers to replace churn. The second company’s growth is more expensive to sustain and may be more vulnerable if acquisition costs rise, the sales cycle lengthens, or the market becomes more competitive.

For a buyer, that often means modeling a lower-quality revenue base, higher customer acquisition requirements, or additional investment in customer success and product. For an investor, it means testing whether the growth plan depends on an optimistic improvement in retention that has not yet shown up in cohort data.

Diagnose the Number Before Negotiating Around It

Separate contraction from churn

Churn means a customer leaves; contraction means a customer stays but spends less. The distinction matters. Churn may signal poor product fit or a failed implementation. Contraction may reflect seasonal usage, seat reductions, budget pressure, or a product whose value does not scale with customer needs.

Ask management for a bridge from starting revenue to ending revenue, with churn, downgrades, and expansion broken out separately. Then compare that bridge with customer counts. A small number of large cancellations can look very different from broad-based, low-level contraction.

Look at cohorts, segments, and customer concentration

Company-wide NRR can hide important patterns. Review cohorts by customer start date, segment, product, geography, and contract size. A newer cohort that retains well may indicate that the product or onboarding has improved. Conversely, a strong average can mask a deteriorating enterprise segment or an older product line that is losing relevance.

Also identify how much of the metric depends on a handful of customers. If several large accounts dominate revenue, one renewal decision can move NRR sharply. That is a concentration risk as well as a retention issue.

Check the calculation and the evidence

Confirm the measurement period, treatment of price changes, currency conversions, acquisitions, and customers that pause or resume service. Ensure that the revenue definition matches the company’s recurring-revenue claims. Then reconcile reported NRR to billing records, renewal data, and customer-level revenue. A polished chart is not a substitute for a reproducible calculation.

For AI Deals, Retention Depends on the Workflow—not Just the Model

AI companies introduce a specific question: is the customer paying for access to a model, or for a dependable workflow that produces a valuable result? Access alone can be easier to compare and replace. A workflow embedded in review, research, customer support, or another operating process may be harder to remove—provided it is accurate, usable, and trusted.

That distinction matters when considering model aggregators and multi-model orchestration. A model aggregator can offer access to multiple models through a common interface. Multi-model orchestration goes further: it routes or coordinates models as part of a task flow, potentially combining model selection, tools, validation, and human review. The terms are not always used consistently, so a buyer should examine the actual product rather than rely on the label.

A product that lets a customer choose among OpenAI, Anthropic, and Google models may be convenient. But model choice alone does not prove durable value. If customers can switch providers without changing their workflow, the interface may face price competition and limited differentiation. By contrast, orchestration that solves a repeatable business problem may create stronger product dependency—but only if customers can see the outcome and trust the system.

Disagreement Can Be a Feature for Decision Quality

In some workflows, running more than one model is not just redundancy. It can help reveal uncertainty. If two models produce materially different answers, that disagreement may signal that the task needs additional evidence, a human decision, or a more cautious response. The goal is not to declare one model correct automatically; it is to use disagreement as a cue for review.

For a deal team evaluating an AI product, ask how the system handles disagreement. Does it expose conflicting outputs? Does it resolve them using evidence, rules, or a human reviewer? Can the customer see why a recommendation was made? If the product hides disagreement behind a single confident answer, it may make decisions look cleaner while making their reliability harder to assess.

This has direct retention implications. Customers are more likely to keep paying for a system they can use safely in a consequential workflow. A product that produces useful uncertainty signals and clear escalation paths may earn trust more effectively than one that simply returns an answer quickly.

Sequential vs. Parallel Workflows

Sequential workflows

In a sequential workflow, one step informs the next. For example, a system may retrieve relevant documents, ask a model to summarize them, check the summary against source material, and then route uncertain cases to a person. This can make the process easier to audit and may reduce unnecessary model calls. It can also introduce latency and create dependencies: if an early step fails, later steps inherit the error.

Parallel workflows

In a parallel workflow, multiple models or methods work on the same task at the same time. The system can compare outputs, identify disagreement, or select a result based on a defined evaluation process. Parallel processing may improve coverage and make cross-checking possible, but it can increase cost and does not guarantee accuracy. Several models can make the same mistake, especially when they rely on similar information or assumptions.

The relevant deal question is not whether parallel or sequential is universally better. It is whether the workflow is designed around the customer’s risk tolerance, speed requirements, and economics. More model calls can raise infrastructure costs. Too few checks can weaken reliability. A durable product balances these needs and can explain how the balance affects customer outcomes.

Cross-Checking Can Catch Hallucinations—but It Is Not a Guarantee

Cross-checking can help identify hallucinations by comparing an answer with retrieved sources, structured data, a second model, or explicit business rules. For multi model AI example, a system might flag a claim that lacks supporting evidence in the documents it was asked to analyze. That is useful, but it should not be described as a universal hallucination fix.

Independent checks can still fail. A second model may repeat an error, a source may be incomplete, or a validation rule may be too narrow. Ask for evidence of how the product performs on representative customer tasks, including failure cases. Look at error rates, review rates, escalation behavior, and the cost of checking—not only a polished demo.

For retention analysis, connect these product controls to customer behavior. Are customers expanding usage because the system saves time or improves decisions? Are they limiting deployment to low-risk tasks because accuracy is uncertain? Do renewal conversations cite trust, workflow adoption, and measurable value, or mostly discounts and unused capacity?

Questions to Ask Before You Price the Deal

What does 78% NRR include? Request the calculation, cohort data, and a bridge showing churn, contraction, and expansion.
Which customers are leaving or spending less? Identify patterns by segment, product, tenure, and contract size.
How much does the company spend to replace lost revenue? Test whether sales and marketing efficiency can support the implied growth plan.
What evidence supports an NRR improvement case? Separate implemented changes and observed results from management targets.
For AI products, what is genuinely embedded? Distinguish model access from workflow value, and test switching costs in practice.
How does the product manage uncertainty? Review model disagreement, source-based checks, human escalation, and failure handling.
Do unit economics hold as usage grows? Model inference costs, support burden, and any multi-model checks alongside gross margin.

How to Frame 78% NRR in the Deal

Do not treat 78% as an automatic deal-breaker, and do not wave it away because new-customer bookings are strong. Treat it as a signal that demands evidence. A company with a clear, improving cohort trend, identifiable causes, and a credible path to better retention may warrant a different view from one with persistent broad-based churn and no operational explanation.

In an AI business, the product story should be tested against customer behavior and economics. Multi-model orchestration, parallel comparison, and hallucination checks can contribute to a more trusted workflow. They do not, by themselves, create retention. The proof is whether customers keep using the product, expand their use, and renew at a price that supports attractive margins.

The Bottom Line

78% NRR means the company is losing 22 cents of recurring revenue for every dollar held by its existing customer base over the measurement period, net of expansion. That puts pressure on growth efficiency, increases NRR risk, and can weigh on the valuation multiple. The right response is to trace the metric to its causes, test its durability across cohorts, and price the deal based on evidence rather than aspiration.

For AI companies, examine whether the business sells interchangeable model access or a dependable workflow with clear value. Cross-checking, model disagreement, and thoughtful sequencing can improve decision quality when designed and measured well. The commercial test remains simple: do customers trust the result enough to keep paying for it?