Not a survey of what revenue teams say they're doing with AI. A ground-truth reconstruction of what 40+ high-growth B2B software companies have actually built, shipped and switched on inside their go-to-market — drawn from four months of delivery logs, 634 recorded customer calls and 16,000 shared-channel messages, then graded workflow by workflow on whether it really runs.
Almost everyone is running AI in go-to-market. Almost nobody bought it. The winning pattern is a general-purpose model wired into the systems a company already owned.
Every "state of AI in GTM" report is a survey of intentions. This one counts only what we could watch being built, deployed and used — and it separates the workflows that run from the ones that are still a good idea in a meeting.
Finding 01 · Build, not buy
57% of companies run at least one AI workflow their own team assembled from a general-purpose model plus the systems they already had. Only 10% run a purpose-built AI GTM product — an AI SDR, AI dialer, AI scoring engine. The packaged-AI category is losing to a chat window and an API key.
Finding 02 · The new CRM interface
40% have a model connected directly to their CRM or warehouse through a tool-calling protocol. Reps and marketers query, report on and update records through an assistant instead of the application UI. In the most mature case the model has its own scoped integration user — a licensed seat that isn't a person.
Finding 03 · The deployment gap
Research and enrichment is the only workflow type where nobody is still just talking about it — 47% run it in production and zero companies have it stuck at the intent stage. Conversational access to data is the mirror image: 50% are working on it, but only 20% have it live. That spread is the single widest gap in the study.
Finding 04 · AI builds the system
26% use AI to build the revenue system itself — CRM schema, pipelines, routing logic, deployment code, dashboards — and every single one of them has it in production. No other workflow type converts from experiment to production at 100%. In one engagement the entire CRM schema was authored with an assistant, directly in production.
Finding 05 · The AI SDR didn't land
Autonomous outbound is where the market spent its marketing budget and where our panel has the least to show: only 3% run it in production. One company had bought an AI SDR, run it, and switched it off — its records survive only as polluted lead-source values a team is now cleaning up. Where outbound AI does work, it drafts for a human.
Finding 06 · Governance is the bottleneck
30% show an active governance problem: shadow assistants bought by individual reps, leaders shipping their own tooling with no inventory, projects blocked for weeks on admin rights. The mature answer showing up in the data is architectural — a scoped service identity for the model, or a central data mirror so access can be revoked in one place.
Finding 07 · AI-native companies are different
Among companies whose own product is AI-native, 73% have a model wired into their GTM systems versus 21% of everyone else — a 3.4× gap. They're not running bigger stacks; they reach for the model where others reach for a feature request.
The discipline that makes this different from a survey: every workflow is graded on whether it actually runs. Where the record was thin, we say so.
The sample. Every company here is — or recently was — a customer of LeanScale, a revenue-operations engineering firm. That is the study's strength and its bias: these are B2B software companies serious enough about go-to-market to hire a specialist partner, and they have a partner in the room actively pushing on AI. Read the findings as "the AI frontier among companies investing in revenue operations," not a random market draw.
The evidence. Four months of primary material across three sources: roughly 7,400 delivery-project records, 634 recorded customer calls, and 16,000 messages from shared customer channels. A keyword pass over the corpus produced candidates; every candidate was then read in context before it was counted.
The shipped-or-said rule. Each workflow was graded on two axes a keyword count cannot see. First, deployment state — running in production, actively being built, piloted, under evaluation, or merely discussed. Only production and building count as adoption. Second, operator — does the customer run it, does the partner run it on their behalf, or is it joint. The partner-run numbers are held separate throughout.
The trap this avoids. A recorded call where a vendor demos its AI to a customer looks identical, to a keyword scan, to a customer describing a system they run. It is graded evaluating and never counted as adoption. This is the largest single source of false positives in transcript mining, and correcting it moves several headline numbers by double digits.
The caveats. Adoption here is a floor, not a ceiling — we count what surfaced in delivery work, so AI a company runs entirely outside its RevOps program is invisible to us. Figures are point-in-time and several companies are mid-build. Percentages are computed against the companies with enough delivery signal in the window to judge; roughly a third of the roster was too quiet to grade and is excluded rather than assumed.
A survey tells you what a revenue leader believes their team is doing with AI. Delivery logs tell you what someone actually provisioned, connected, and paid an engineer to debug at 11pm. This is the second thing.
40+ B2B software companies, weighted toward Series B–C, 50–500 employees and North America — with a heavy AI-native, security and fintech tilt.
By Sector
By Headcount
By Funding Stage
By Revenue Model
By GTM Motion
By Product & Region
Share of companies with each AI tool inside a workflow that is running or being built. Evaluations, demos and wish-lists are excluded.
Most-adopted AI in go-to-market
% of companies · production & in-build only
Counted per company, not per workflow. A general-purpose assistant leads by a wide margin — the story of AI in GTM is not a new category of tool, it's one tool pointed at everything. Tools the partner runs on a customer's behalf are included here and broken out separately in the Partner Lens below.
Build vs buy
How the AI capability arrived · % of companies
"Assembled" means a general-purpose model plus existing systems. "Purchased" means a product sold as AI. Most companies do both — but the assembled side is where the distinctive work is.
One tool, pointed at everything. A frontier assistant is in two-thirds of companies — more than double the reach of any purpose-built AI GTM product in the study. It is used to write Apex, draft business cases, query the CRM, score accounts and build dashboards, often inside the same company in the same week.
Enrichment is the exception that proves the rule. The one category where a packaged AI product genuinely won is programmable enrichment, in half the panel. It succeeded by being infrastructure — a place to run waterfalls and prompts against a credit budget — rather than by automating a job.
The CRM-native AI features are quiet. Vendor-embedded assistants appear in roughly a tenth of companies, and where they appear they are usually being compared unfavourably to an outside model. One company's rebuild explicitly rejected the native option in favour of pointing an external model at their own metadata.
For every workflow type: what share of companies have it running, what share are mid-build, and what share have only talked about it. The distance between those bars is where the AI conversation and the AI reality separate.
Running vs building vs talking
% of companies at each stage, by workflow type
A company is counted once per workflow type, at its highest stage. Read the top rows as settled categories and the middle as the live frontier.
Where every graded workflow landed
All 108 graded workflows by deployment state
Just over half of everything we found is genuinely in production. That is a far higher hit rate than the discourse implies — but it is concentrated. Enrichment, construction work and content drafting convert; conversational data access and autonomous outbound do not, yet.
The widest gap is the most-hyped capability. Talking to your CRM is the second-most common workflow type in the panel and one of the least deployed. The blockers we watched were almost never model quality — they were admin rights, a missing semantic layer, per-user API keys nobody wanted to hand out, and a genuine fear of an agent writing to the wrong field.
Nobody is "considering" enrichment. It is the only workflow type with zero companies stuck at the intent stage. When a category is truly solved, the evaluation phase disappears — teams just run it. Watch for that signature as the other categories mature.
The categories that convert fastest are the ones where a human still checks the output before it leaves the building. The categories that stall are the ones that ask a company to let a model act unsupervised on the record of truth. AI adoption in GTM is not gated by intelligence — it's gated by blast radius.
For each workflow type: the share of companies running or building it, and the tools doing the work. Coverage is the headline number; the bars are what it's built on.
Which function got there first
% of companies with a live AI workflow owned by each function
How much rope the model gets
% of companies with at least one workflow at each autonomy level
"Autonomous" means it runs without a human in the path — overwhelmingly on enrichment and routing, where a wrong answer is cheap and reversible.
The questions this study was built to answer: does size help or hurt? Do the companies selling AI actually use it? Does how you charge change what you automate?
Do AI companies run AI?
% of companies with a model wired into their GTM systems
Selling AI predicts using AI — by 2.5×. The clearest relationship in the data. Companies whose own product is AI-native connect a model to their revenue systems at more than double the rate of everyone else. The mechanism is cultural, not technical: the same instinct that builds an AI product reaches for a model instead of a ticket.
Small is fast, big is deep. Under 200 employees, essentially everyone has something in production — small teams adopt because there is nobody to ask for permission. Above 500 the production rate dips but the depth rises sharply: the biggest companies are the most likely to have a model wired into the system of record, because they're the only ones who can fund the governance to do it safely.
The middle is where it stalls. The 201–500 band has the lowest rate on both measures. Big enough to have controls, not big enough to have a platform team to satisfy them — the classic RevOps squeeze, now showing up in AI.
By company size
% running AI in production vs % with a model wired into the systems
Bands collapsed so each holds enough companies to report; the largest band is the thinnest and should be read as directional.
By revenue model
% running AI in production
Production rates are high across every pricing model — how a company charges turns out to predict far less than how large it is or whether its own product is AI-native. The smallest bands here hold only a handful of companies and should be read as directional.
AI adoption isn't random — it clusters. These five are overlapping lenses, not exclusive bins; a company can live in two. Membership is rule-derived from the graded workflows.
Because every workflow was tagged with its operator, the study reveals its own instrument: how much of this AI arrived through the services partner rather than the customer's own initiative.
Partner-introduced AI
% of engagements where the partner runs or co-runs each tool
The partner is the distribution channel. A frontier assistant appears in slightly more engagements as partner-run or joint than as purely customer-native. For a meaningful slice of these companies, AI arrived in their go-to-market because a vendor brought it — not because they went looking.
That is exactly why the provenance split matters. A naive keyword count of these same transcripts would report near-universal AI adoption. Separating who actually operates the workflow is the difference between a real number and a flattering one.
Read it as a leading indicator. Capabilities that show up first as partner-run — conversational data access, agent libraries, transcript pipelines — are the ones customers are most likely to internalize next. It is the diffusion path, visible a quarter or two early.
The reason to read a study of other companies' AI is to calibrate your own sequence. Based on where adoption is settled, where it's converting, and where it's still failing, the twelve workflow types sort into three tiers.
Tier 1 · Do this now
Tier 2 · The live frontier
Tier 3 · Not yet
Start where the output is checkable and the blast radius is small — enrichment, construction, drafts. Fund Tier 2 as a governance project, not a tooling one; the blockers are permissions and semantics, not models. Leave Tier 3 to the vendors for now — the panel has already tried it.
Want this run on your GTM?
We'll map the AI already running inside your revenue org — what's live, what's stalled, and what to build next — and then go build it with you.
Talk to LeanScale →