How to Value an AI-Enabled Services Company

How to value an AI services company when it is neither SaaS nor an agency

If you are working out how to value an AI services company, start with what the customer actually buys. An AI-enabled bookkeeping firm sells closed books, not seats. An AI-enabled recruiting firm sells hires. An AI-assisted legal or back-office provider sells reviewed contracts or processed claims. Investors call this model "service as software" or AI-native services: models and agents do most of the work, and people stay in the loop to review, fix, and take responsibility.

That position is why the usual shortcuts fail. A SaaS revenue multiple assumes software-like margins and low marginal cost. A traditional services multiple assumes revenue grows only as headcount grows. An AI-native services company can sit anywhere between those two, and the main diligence job is finding out where.

This guide covers the margin question, how to define gross margin, revenue quality, three valuation methods applied to one illustrative company, the risks that call for a discount, and a diligence checklist. It is educational material, not financial, tax, legal, or valuation advice. For the general micro-SaaS framework, read how to value a micro-SaaS business first. This page covers what changes when the product is an outcome.

Where AI-native services sit on the spectrum

It helps to think of a spectrum with four rough stages:

Agency or BPO: people deliver the work, and software helps with coordination. Revenue grows roughly in line with headcount.
Tech-enabled services: in-house tools make people faster, but people still do the core work.
AI-native services: models do the core work. People review exceptions, handle quality, and deal with clients. Revenue can grow faster than headcount.
SaaS: the customer does the work with the software, and the vendor's marginal delivery cost is mostly hosting and support.

Public-market data shows why the stage matters. In Aswath Damodaran's January 2026 U.S. industry datasets, Business & Consumer Services companies had an aggregate gross margin of about 33% and traded at about 2.5x EV/Sales. System & Application Software companies had a gross margin of about 72% and traded at about 11.4x EV/Sales (Damodaran, revenue multiples; Damodaran, margins by sector). These are large public companies, not small private firms, so treat the numbers as directional only.

Investors say outright that the thesis depends on margins. Bessemer's framework for AI-native services asks whether "AI can actually do the work at software-like margins" and whether a company can "keep the surplus rather than competing it away" (Bessemer Venture Partners, July 2026). General Catalyst's AI-enabled roll-up thesis says its companies aim to double profit margins, "often targeting 30-40% margins" (General Catalyst, "The Future of Services," August 2025). Those are investor targets, not sector results.

The key question: is the margin structural or temporary?

A high gross margin can come from two different places:

Structural automation: models really do most of the work, and the cost per outcome falls as volume grows.
Temporary sources: the founder does unpaid review work, human reviewers are paid below market, offshore labor is labeled as "AI," a model provider is subsidizing usage with credits, or the client mix is unusually easy.

Only the first source deserves anything close to a software-style multiple. To show it, ask for these series, monthly or quarterly, going back at least 12 months:

Human-hours per outcome: reviewer and delivery hours per bookkeeping close, per placed candidate, or per contract reviewed. This number should trend down or hold steady while volume grows.
Automation rate: the share of tasks or line items finished without human edits. Define "touched" precisely, because a reviewer who checks every item has touched every item.
Cost per outcome: total delivery cost divided by outcomes delivered, split into model cost and human cost.
Model/API cost per outcome: inference spend from provider invoices, with any free credits removed.
QA and rework rates: the share of outcomes sent back, corrected after delivery, or disputed by clients. Rising automation alongside rising rework is a warning sign.
Gross margin by cohort: margin for clients onboarded in each quarter. Older cohorts should usually be cheaper to serve. If margin holds up only in the newest, smallest cohort, it may not last.

Illustrative trend: human-hours per monthly close falling from 4.1 to 3.0 to 2.3 to 2.0 over four quarters while the client count doubles is good evidence of automation. Hours that stay flat while the founder's own time goes undocumented is not.

Define gross margin the right way

For these businesses, cost of revenue (COGS) should include every cost of delivering the outcome:

model inference and API costs, embeddings, and vector storage
human reviewers, QA staff, and delivery or account staff who touch the work
contractors and offshore teams, however they are labeled
the cost of rework and corrections
delivery tooling charged per client and payment processing fees
a market-rate cost for any founder time spent on delivery

Gross margin = (revenue − cost of revenue) ÷ revenue

Worked unit economics: one monthly bookkeeping close (illustrative)

Suppose an AI-enabled bookkeeping firm charges $900 per client per month. Every figure below is illustrative.

Model and API cost: $18
Reviewer time: 2.0 hours × $45 loaded = $90
Senior QA review (currently done by the founder, unpaid): 0.5 hours × $70 = $35
Per-client tooling: $22
Payment processing: $27
Rework: 8% of closes need 1.5 extra reviewer hours, so 0.08 × 1.5 × $45 = $5.40

The founder's "reported" cost leaves out QA and rework: $18 + $90 + $22 + $27 = $157, which gives a gross margin of 82.6%. The normalized cost is $197.40, which gives a gross margin of 78.1%.

The same logic applies to recruiting. If a placement earns an illustrative fee of 20% on a $120,000 salary, or $24,000, divide total sourcing, screening, model, and recruiter cost by placements, not by candidates screened. Include the cost of searches that never fill.

Notice that model cost is a small part of this example. When people costs dominate COGS, the margin story depends on human-hours, not on falling token prices.

Revenue quality

Two companies with the same revenue can deserve very different values.

Recurring retainers or subscriptions (a monthly close, ongoing compliance work) are the most predictable revenue. Check contract length, notice periods, and auto-renewal terms.
Per-outcome or contingent fees (recruiting placements, success fees) are lumpy and cyclical. Value them on trailing-twelve-month revenue and look at the pipeline, not one strong month.
One-off projects (catch-up bookkeeping, migrations) should be separated from recurring revenue and usually earn a lower multiple.
Retention: calculate gross revenue retention (GRR) and net revenue retention (NRR) by cohort. See net revenue retention.
Concentration: the shares of revenue from the largest client, the top five clients, and the largest referral channel.
Gross vs net: if the company passes through third-party costs such as job-board spend or filing fees, see gross vs net revenue.

Pricing power when AI costs fall

Inference prices have fallen quickly. An a16z analysis estimated that, for a model of equal capability, "the cost is decreasing by 10x every year," while noting that the method is imperfect (a16z, "Welcome to LLMflation," November 2024). Falling costs help margins only if prices hold. If competitors pass their savings on to customers, outcome prices can fall faster than your human costs do.

Illustrative stress test: in the bookkeeping example, a 15% price cut to $765 with unchanged cost per close reduces gross margin from 78% to about 74%. Across 150 clients, annual recurring gross profit drops from about $1.26 million to about $1.02 million. Nearly all of that loss comes straight out of EBITDA.

Valuation methods

SDE or EBITDA multiple

For a smaller, owner-operated firm, most buyers underwrite cash flow. Use seller's discretionary earnings (SDE) when one owner-operator will take over, and EBITDA when a management team stays. See SDE vs EBITDA. For context, Acquire.com reports that profitable online-business deals on its marketplace in 2025 typically traded at 3x to 5x net income, with averages of 3.7x to 3.9x depending on size (Acquire.com, 2025 multiples recap). That data covers mostly software and online businesses, not AI services specifically. Normalize earnings for founder labor before you apply any multiple.

Revenue or gross-profit multiple

Revenue multiples overstate value when margins differ, so a gross-profit multiple is the better bridge between services and SaaS comparables. Using the Damodaran figures above, dividing EV/Sales by gross margin implies roughly 7.6x gross profit for business services and roughly 15.9x for system software. The gap shrinks from about 4.5 times on revenue to about 2 times on gross profit. That narrower gap is the logic of the bridge: once margins are accounted for, a buyer is paying for durability, growth, and risk. These public aggregates are much higher than small private deals and should not be applied directly. SaaS Capital, for example, estimated private SaaS multiples well below its public index (SaaS Capital, 2025 private SaaS valuations). For recurring-revenue framing, see ARR multiples.

DCF sanity check

A discounted cash flow model makes explicit the assumptions a multiple hides: price compression, falling human-hours, model cost, and churn. With small companies, use it to check a multiple-based price rather than to set the price.

One company, three valuations (illustrative)

Consider "Ledgerline," a fictional AI-enabled bookkeeping firm. All numbers and multiples are illustrative, not market benchmarks.

150 clients at $900 per month: $1.62 million in recurring revenue at a 78% normalized gross margin, which is about $1.26 million of gross profit
$180,000 of catch-up project revenue at a 50% margin: $90,000 of gross profit
Total revenue of $1.8 million and total gross profit of about $1.35 million
Normalized EBITDA of $420,000 after paying a market-rate general manager. SDE for a single owner-operator is $540,000.

1. Earnings multiple: $420,000 EBITDA × 5.0x = $2.10 million. As a cross-check, $540,000 SDE × 4.0x = $2.16 million.

2. Gross-profit multiple: $1.35 million × 2.0x = $2.71 million, which works out to about 1.5x total revenue.

3. DCF: free cash flow of $350,000 in year one, growing 15% a year for five years, discounted at 25%, with a terminal value of 4x year-five cash flow. The present value of the five years of cash flow is about $1.19 million and the present value of the terminal value is about $0.80 million, for a total of about $2.0 million.

Reconciliation: the earnings and DCF methods agree at about $2.0 to $2.1 million. The gross-profit method comes out higher because it assumes margins will hold. A buyer who accepts the automation evidence might settle around $2.2 to $2.4 million, perhaps with part of the price contingent on revenue retention (see earnouts). Under the 15% price-cut scenario, EBITDA falls to about $180,000, and both cash-flow methods fall sharply. That downside case should drive the structure of the deal.

You can test your own inputs with the online business valuation calculator or the SaaS valuation calculator.

Risks that justify discounts

Model provider dependency: providers retire models on a schedule. OpenAI's deprecation policy, for example, promises at least six months' notice for generally available models and shorter notice for preview models (OpenAI deprecations). Ask whether the company has tested its workflows on other models, and how prompts and evaluations are versioned.
Regulatory and professional liability: accounting, tax, legal, and employment work carry licensing and liability exposure. In recruiting, New York City's Local Law 144 requires a bias audit and candidate notices before employers or employment agencies use automated employment decision tools (NYC DCWP). Check who signs off on the work, what professional liability insurance covers, and how errors are handled.
Overstated AI claims: regulators have acted against overstated AI capabilities. The SEC settled "AI washing" charges against two investment advisers in 2024 (SEC press release 2024-36). The FTC's Operation AI Comply included a complaint against DoNotPay's "robot lawyer," which the FTC said never tested whether its output matched a human lawyer's (FTC, September 2024). Marketing claims are a liability you inherit.
How much of the "AI" is offshore labor: reconcile contractor invoices and timesheets against the automation claims. Offshore delivery is legitimate, but value it as labor.
Data rights and confidentiality: client contracts may limit using client data to train or evaluate models, or sending it to third-party APIs. Review data processing agreements and the provider's data-retention settings.
Key-person risk among reviewers: if one senior reviewer holds the quality bar, losing them hurts both margin and quality.

Due-diligence checklist and data room

Buyers should request, and sellers should prepare, the following. The general startup buyer due diligence checklist still applies.

Monthly revenue by client, split into retainer, per-outcome, and project revenue, reconciled to bank deposits and the general ledger
A COGS breakdown: model invoices (with credits shown separately), reviewer payroll, contractors, tooling, and rework
Human-hours per outcome, automation rate, and QA/rework rates for at least 12 months, with definitions
Gross margin by client cohort and by service line
Client contracts: pricing terms, termination rights, liability caps, data-use clauses
Model and vendor inventory, fallback testing results, and evaluation sets
Licenses, professional liability insurance, and any compliance audits (such as AEDT bias audits)
An org chart showing who reviews what, plus offshore and contractor arrangements
Records of errors, client complaints, and refunds
IP assignments for prompts, workflows, and code

How RevenueBug helps, and where it stops

RevenueBug shows trailing 30-day revenue and a daily revenue history calculated from a connected payment provider such as Stripe. For an AI services firm that bills through a supported provider, that history helps a buyer see whether retainer revenue is steady or lumpy. It does not verify margins, automation rates, costs, contracts, or professional compliance, and it is not an audit or a valuation. The methodology explains what is in scope. Founders can list a business, and buyers can browse AI startups for sale or the wider AI category.

Frequently asked questions

Should an AI services company get a SaaS multiple?

Not by default. A SaaS-style multiple becomes defensible only when normalized gross margin, falling cost per outcome, and recurring retention all look like software. Otherwise, start from earnings or gross profit.

What gross margin counts as "AI-native"?

There is no standard threshold. What matters is that margin is calculated with all reviewers, QA, rework, model costs, and founder delivery time included, and that it holds or improves across cohorts as volume grows.

Are recruiting placement fees recurring revenue?

No. They are contingent and cyclical. Value them on trailing-twelve-month results and look at client repeat rates, rather than annualizing a strong month.

Do falling model costs raise the valuation?

Only if the company keeps the savings. If competitors cut prices, the savings go to customers, and falling inference costs do not help much when human review makes up most of COGS.

How do I tell whether the "AI" is really people?

Compare contractor and payroll costs with outcome volume over time, ask for automation-rate definitions, and watch a live delivery run. Flat human-hours per outcome during growth point to a labor-based business.

Is a gross-profit multiple better than a revenue multiple here?

Usually, yes, for comparing across the spectrum, because it adjusts for different margin structures. Still cross-check it against an earnings multiple and a DCF.