Vertical SaaS gross margins were 78 to 82 percent in 2023. The same companies adding AI features in 2026 are reporting blended gross margins of 63 to 68 percent; and the trade is worth it, because TAM expanded two to three times. This piece decomposes the AI-feature COGS shift, walks three vertical-SaaS examples (legal, healthcare, fintech), and names the trade-off math that turns 15 points of margin compression into a strategically dominant move. The argument runs counter to the SaaS investor reflex that defends 80 percent gross margins as a moat. Defending 80 percent margins in a vertical where AI features are emerging is a strategy that wins the next two quarters and loses the next two years.
The piece extends the AI project economics manifesto into the SaaS gross-margin domain. Manifesto principles say “evaluation is the unit of account”; this piece says “the COGS line that funds evaluation is what reshapes vertical SaaS gross margin profiles.”
Why vertical SaaS margins are about to reset
Vertical SaaS earned 78 to 82 percent gross margins by combining low marginal hosting cost with high pricing power inside a defensible niche. The pricing power came from owning the workflow; the cost structure came from delivering software where the marginal cost of one more user was rounding error.
AI features change the cost structure on the inputs side without changing the pricing power on the output side, at least in the short run. An AI-enabled contract-review module that costs $0.04 per processed page is a real per-customer cost line that did not exist before. A clinical-summary feature that costs $0.18 per encounter is a real per-customer cost line that did not exist before. The pricing power that supported 80 percent margins is now spread across a cost base that is 35 percent richer per customer.
The naive response is to refuse the AI features and defend the margin. The strategically correct response is to embrace the margin compression as the cost of an expanded product surface that grows TAM, defends against AI-native entrants, and earns price increases that recover most (not many) of the compression. The math, walked below, says the gross margin will land 12 to 17 points lower and total gross profit dollars will be two to three times higher within 36 months.
Decomposing AI-feature COGS
The COGS shift is composed of four lines that do not exist in pre-AI vertical SaaS COGS.
Inference cost. The dominant line. For a vertical-SaaS product where AI features are core to the workflow, inference cost runs 3 to 8 percent of revenue at scale, depending on call-graph design and model selection. This is the line that needs the cost-per-action discipline to stay defensible.
Eval engineering team. A standing eval-engineering function; typically two to four people for a mid-market vertical-SaaS product; whose work is treated as COGS rather than R&D because it is structurally required to operate the AI features in production. Annual cost roughly 1 to 2.5 percent of revenue at scale. The team owns the eval suite, regression detection on model upgrades, and prompt versioning.
Observability scale-up. AI workloads produce 5 to 10 times the trace volume of equivalent non-AI workloads (most retrieval, most generation, most self-critique step is a trace). Observability cost lands at 0.8 to 1.6 percent of revenue, up from roughly 0.2 percent for pre-AI SaaS. The line is small in dollar terms but structurally significant because skipping it makes the eval discipline impossible.
Customer-success uplift for AI features. AI features have higher initial-failure rates and higher customer-education cost than equivalent non-AI features. Customer-success spend rises 0.5 to 1.5 percent of revenue. This line is the one most often missed in modeling because it gets absorbed into the existing CS budget rather than tracked separately.
The four lines sum to roughly 12 to 17 percent of revenue at scale. The traditional SaaS COGS lines; hosting, support, payment processing; drop slightly as AI features replace some manual workflows on the customer-success side, but not nearly enough to offset. The net is the 12-to-17-point margin reset.
Example 1: legal-tech vertical SaaS
A legal-tech vertical-SaaS product targeting mid-market law firms had 81 percent gross margin in 2023 on a $42M ARR base, selling document management and time tracking. In 2026, the product added AI-powered contract review, deposition summarization, and discovery-document classification.
Pre-AI COGS lines. Hosting 4 percent of revenue. Support 7 percent. Payment processing 1 percent. Customer success 7 percent. Total 19 percent of revenue, gross margin 81 percent.
Post-AI COGS lines. Hosting 3.5 percent (slightly lower because some workflows shifted server-side). Support 6 percent (lower because contract review handles the questions support previously fielded). Payment processing 1 percent. Customer success 7.5 percent. Inference cost 6.2 percent. Eval team 1.8 percent. Observability 1.1 percent. CS uplift 1.0 percent. Total 28.1 percent of revenue, gross margin 71.9 percent.
TAM expansion. AI features moved the product from a tool that supported lawyers’ workflows to a tool that performed lawyer-adjacent work. Solo practitioners and small firms who could not afford the full SaaS subscription at $400/seat/month became viable customers at $90/month for the AI-only tier. The buyer base expanded from roughly 18,000 mid-market firms to roughly 92,000 firms across many segments. Realistic 36-month penetration captured roughly $130M ARR against a 2023 base of $42M.
Gross profit dollars. $42M × 81% = $34M (2023). $130M × 72% = $93.6M (2026 baseline). Gross profit dollars grew 2.75x while gross margin compressed 9 points. The strategically correct trade.
Example 2: healthcare vertical SaaS
A healthcare vertical-SaaS product serving multi-specialty clinics had 79 percent gross margin in 2023 on a $58M ARR base, selling EHR-adjacent workflow management. In 2026, the product added AI clinical-note summarization, prior-authorization drafting, and clinical-coding suggestions.
Pre-AI COGS lines. Hosting 5 percent of revenue (HIPAA infrastructure premium). Support 8 percent. Payment processing 0.5 percent. Customer success 7.5 percent. Total 21 percent of revenue, gross margin 79 percent.
Post-AI COGS lines. Hosting 4.5 percent. Support 7 percent (clinical-note summary cuts charting questions). Payment processing 0.5 percent. Customer success 8 percent. Inference cost 7.4 percent (longer documents, mandatory eval-pass thresholds, mandatory PHI handling overhead). Eval team 2.2 percent. Observability 1.2 percent. CS uplift 1.5 percent (clinical-safety education is high-touch). Total 32.3 percent of revenue, gross margin 67.7 percent.
TAM expansion. AI features turned a workflow product into a productivity product. The buyer base remained roughly the same in count (clinics do not multiply), but average revenue per account grew 2.4x because the product moved from “the EHR adjunct” to “the clinical-productivity layer.” Realistic 36-month outcome was a $135M ARR base against a 2023 base of $58M.
Gross profit dollars. $58M × 79% = $45.8M (2023). $135M × 67.7% = $91.4M (2026 baseline). Gross profit dollars grew 2.0x while gross margin compressed 11 points. The strategically correct trade, with a tighter compliance overlay than legal-tech because of HIPAA and PHI handling; see AI agencies for healthcare with HIPAA compliance for the regulatory layer that adds to the eval-team cost line.
Example 3: fintech vertical SaaS
A fintech vertical-SaaS product targeting mid-market accounting firms had 82 percent gross margin in 2023 on a $31M ARR base, selling general ledger and bookkeeping automation. In 2026, the product added AI-powered transaction categorization, anomaly detection, and audit-prep document generation.
Pre-AI COGS lines. Hosting 3.5 percent. Support 6 percent. Payment processing 1.5 percent. Customer success 7 percent. Total 18 percent of revenue, gross margin 82 percent.
Post-AI COGS lines. Hosting 3.2 percent. Support 4.5 percent (AI categorization cuts the questions support fielded). Payment processing 1.5 percent. Customer success 7.5 percent. Inference cost 5.8 percent. Eval team 2.0 percent (financial accuracy thresholds require dense eval suites). Observability 1.0 percent. CS uplift 0.8 percent (existing CS bandwidth absorbed most of the AI training). Total 26.3 percent of revenue, gross margin 73.7 percent.
TAM expansion. AI features turned the bookkeeping product into a near-CFO-substitute for small businesses. The buyer base expanded from roughly 12,000 mid-market accounting firms to roughly 41,000 firms plus 280,000 individual SMB businesses on a self-serve tier. Realistic 36-month outcome: $98M ARR against a 2023 base of $31M.
Gross profit dollars. $31M × 82% = $25.4M (2023). $98M × 73.7% = $72.2M (2026 baseline). Gross profit dollars grew 2.84x while gross margin compressed 8 points. The strategically correct trade, with the cleanest economics of the three because the eval suite is naturally bounded by accounting-rule precision.
The TAM expansion math
Across the three examples, gross margin compressed 8 to 11 points and gross profit dollars grew 2.0 to 2.8 times. The pattern is consistent because the same underlying mechanic produces the result.
The mechanic. AI features in vertical SaaS turn workflow software into productivity software. Productivity software addresses a wider buyer set (more segments, more sizes) and earns higher per-account revenue (the product does more). The combined effect; wider denominator, higher numerator; drives the 2-3x gross profit dollar expansion that more than offsets the 8-11 point margin compression.
Why the trade favors the AI investment. A vertical SaaS that defends 80 percent margins by refusing AI features keeps a smaller TAM at higher gross margin and is structurally vulnerable to AI-native entrants who land at 65 percent margins on the expanded TAM. Defending 80 percent looks responsible for two quarters and looks like a strategic mistake by quarter eight.
Where the trade does not work. It does not work for vertical-SaaS products with small TAM (where 2-3x expansion of a small base is still small), with regulatory environments that block AI features (where the COGS lines hit but the TAM expansion does not), or with buyer bases that have already paid for the full workflow value (where there is no headroom for AI to capture additional revenue per account). These exceptions exist; they are the minority.
Investor narrative and the Rule of 40 reset
Public-market investors have been benchmarking SaaS at the Rule of 40; the sum of revenue growth and gross margin should be 40 or higher. The shift to AI-feature COGS resets that benchmark in vertical SaaS.
The new shape. Pre-AI vertical SaaS: 25 percent growth + 80 percent gross margin = 105 (well above 40). Post-AI vertical SaaS: 55 percent growth + 67 percent gross margin = 122 (well above 40, by a wider margin, with much stronger growth). The Rule of 40 still works; the composition of the 40 has shifted from “high margin, slower growth” to “lower margin, faster growth.”
What investors should ask. Three questions that separate disciplined AI-margin compression from undisciplined AI-margin compression. First: is the cost-per-action discipline in place at the feature level? Second: is the eval-engineering team structured as a standing function or as a per-project add? Third: are the TAM-expansion claims supported by actual cohort data showing new segments converting, not just larger numbers in the deck? The first two filter out vendors who have not internalized the cost discipline. The third filters out narratives that have not yet produced revenue.
The board conversation that goes wrong. A board pushing back on AI-feature investment by citing 80 percent gross margins is asking the company to optimize for the wrong objective. The right metric is gross profit dollars three years out, weighted by competitive defensibility against AI-native entrants. Boards that fixate on the margin line and ignore the dollar line make the predictable strategic mistake; see why AI projects fail the CFO before they fail the CTO for the broader pattern of finance-side mistakes that wreck AI investments before engineering ever gets a chance to.
Frequently asked questions
Why do AI features compress vertical SaaS gross margins?
AI features add four COGS lines that did not exist in pre-AI vertical SaaS: inference cost, eval-engineering team, observability scale-up, and customer-success uplift for AI features. The four lines sum to roughly 12 to 17 percent of revenue at scale. Existing COGS lines (hosting, support) drop slightly but not enough to offset, producing a net 8 to 12 point margin compression for most vertical-SaaS products.
How much does gross margin compress?
For most vertical SaaS products with meaningful AI features, gross margin moves from 78 to 82 percent (pre-AI) into the 65 to 73 percent range (with AI). The exact number depends on the inference-cost ratio (a function of call-graph design and model selection), the eval-team scale, and how aggressively customer-success-side cost reductions are realized.
Is the margin compression worth it?
In nearly many vertical SaaS where AI features are technically feasible, yes. The TAM expansion that AI features enable typically grows gross profit dollars 2.0 to 2.8 times within 36 months. A vertical-SaaS company that defends 80 percent margins by refusing AI features ends up with a smaller TAM at higher gross margin and is vulnerable to AI-native entrants. Defending margin in this market is a strategy that wins two quarters and loses two years.
What is the largest single new COGS line for AI features?
Inference cost, typically 3 to 8 percent of revenue at scale. The eval-engineering team is the second largest at 1 to 2.5 percent. Observability and customer-success uplift round out the four-line addition. Together they sum to 12 to 17 percent of revenue, which is the AI-feature COGS load that produces the gross-margin reset.
Should the eval-engineering team be in COGS or R&D?
COGS, because it is structurally required to operate the AI features in production. An eval suite that stops running is not a working AI feature. Treating eval engineering as R&D produces under-funded eval functions whose first regression is invisible until a customer reports it. The accounting framing flows downstream from the production-criticality framing.
Does this analysis apply to horizontal SaaS as well?
The COGS-line decomposition applies to horizontal SaaS too. The TAM-expansion math is weaker because horizontal SaaS often already addresses a wide buyer base and has less room for productization gains. Horizontal SaaS adopting AI typically sees 6 to 10 points of margin compression with 1.3 to 1.8x gross-profit-dollar growth; still favorable, but not as dramatic as vertical SaaS.
How does this affect the Rule of 40?
The composition of the 40 shifts. Pre-AI vertical SaaS often hit 25 percent growth + 80 percent gross margin = 105. Post-AI vertical SaaS hits 55 percent growth + 67 percent gross margin = 122. The Rule of 40 still works as a benchmark; the path to clearing it has shifted from “high margin, slower growth” to “compressed margin, faster growth.” Investor frameworks need to reset accordingly.
What if my product cannot easily add AI features?
Three reasons it might not. The TAM is too small for 2-3x expansion to matter. The regulatory environment blocks AI features. The buyer base has already paid for the full workflow value. In these cases, the margin compression is not justified by TAM expansion and the product should focus on workflow defensibility rather than AI-feature breadth. These cases are the minority of vertical SaaS.
Key takeaways
- AI features compress vertical SaaS gross margins from the 78-82 percent pre-AI range into the 65-73 percent post-AI range, driven by four new COGS lines: inference, eval engineering, observability scale-up, and customer-success uplift, totaling 12-17 percent of revenue.
- The compression is offset by TAM expansion. AI features turn workflow software into productivity software, expanding both the buyer base and per-account revenue. The combined effect grows gross-profit dollars 2.0 to 2.8 times within 36 months.
- Three vertical-SaaS examples (legal, healthcare, fintech) many land in the same band: 8-11 points of margin compression, 2.0-2.8x growth in gross-profit dollars, and a buyer base 2-3x wider than the pre-AI footprint.
- The Rule of 40 still works as an investor benchmark, but the composition shifts from “high margin, slower growth” to “compressed margin, faster growth.” Investor frameworks built on 80 percent gross margins as a moat will mis-evaluate vertical SaaS by 2027.
- The strategically incorrect response is to defend pre-AI margins by refusing AI features. That strategy wins two quarters and loses two years to AI-native entrants who land at 65 percent margins on the expanded TAM.
Arthur Wandzel
