Every data and GTM leader has had the same conversation this year. The backlog is growing, the requests are stacking up, the roadmap slips a quarter. So you buy capacity: another warehouse tier, another seat of an AI coding assistant, another contractor to clear the queue. It is the fastest decision available and it is defensible in a budget meeting. It is also, increasingly, the wrong one — because the thing holding your roadmap back has quietly moved from production to consumption. The models ship. The pipelines run. Nobody trusts them, nobody can read them, and nobody can prove they changed a number. That is not a capacity problem. That is an adoption problem, and it is solved by different hires.
The money is going to throughput, not to trust
Look at where budget actually landed. 57% of teams report increased warehouse and compute spend, against 36% reporting increased team budgets — capacity is being bought as infrastructure almost twice as fast as it is being hired as people. That makes sense when the constraint is throughput. Compute scales in an afternoon; a senior analytics engineer takes a quarter.
The same pattern shows up inside the AI tooling decision. 72% of teams prioritize AI-assisted coding while only 24% prioritize AI-assisted pipeline management — testing, observability and quality controls. Three times as much attention on producing more output as on making output trustworthy enough that someone will act on it. If your bottleneck were authoring speed, that ratio would be correct. If your bottleneck is a stakeholder who opens a dashboard, sees two revenue numbers that disagree and goes back to their own spreadsheet, that ratio is actively making things worse. You are accelerating the production of artifacts nobody has a reason to believe.
Faster build velocity against an adoption bottleneck does not clear the backlog — it lengthens it, because every new asset is another thing to reconcile, document and defend.
The evidence that the constraint moved
Three signals, from three different places in the stack, all point the same direction.
First, consumption. 36% of data teams name stakeholder data literacy as an active barrier, down only slightly from 39% the year before. More than a third of teams are shipping into an audience that cannot use what arrives. No amount of compute changes that figure.
Second, trust in AI output. 46% of developers say they don’t trust the accuracy of what AI tools produce, up sharply from 31% the previous year. Tool rollout is near-universal; confidence in the result is falling. That is the textbook shape of adoption failing after procurement succeeds — the license is deployed, the behaviour never changed.
Third, governance. A Gartner survey of 223 D&A leaders found cultural resistance outweighs funding constraints as the primary reason governance initiatives fail, 60% versus 40%. When the most common post-mortem on a failed initiative is people would not change how they work, you are not underfunded. You are under-adopted.
And the payoff for fixing it is not soft. Companies that implemented AI governance pushed 12x more projects to production. Not 12x more projects built — 12x more shipped into production, where they count. The scaffolding that makes work adoptable is what moves it across the line, not extra build throughput upstream of the line.
Capacity hires and adoption hires are not the same people
This is where the instinct quietly fails. “We need more help” gets translated into “more of who we already have,” and the job description that goes out is a clone of the last one. But the skills that clear an adoption bottleneck barely overlap with the skills that clear a build bottleneck.
| Capacity hire | Adoption hire | |
|---|---|---|
| Success metric | Tickets closed, models shipped, pipelines live | Assets actually used, decisions changed, projects in production |
| Core skill | Authoring — SQL, Spark, orchestration, dbt | Contracts, testing, lineage, semantics, enablement |
| First 90 days | Burn down the queue | Find the three numbers that disagree and kill two of them |
| Fails when | Requirements are ambiguous | They have no executive air cover |
| Typical titles | Analytics engineer, data engineer, platform engineer | Data product manager, governance lead, analytics enablement, observability/quality engineer, forward-deployed analyst |
| What they fix | Throughput | Trust, literacy, measurement |
The adoption column is thinner in most orgs because it is harder to justify. A capacity hire has a legible output. An adoption hire’s best quarter looks like fewer dashboards, fewer metrics, a retired report and a semantic layer nobody argues with any more. That is a deletion-heavy job, and deletion-heavy jobs do not survive a hiring committee that counts shipped artifacts.
It also means you cannot screen for them the same way. A take-home that measures how fast someone writes a transformation tells you nothing about whether they can get a VP of Sales to retire their personal pipeline spreadsheet. When teams brief us on the data engineering network, the useful signal is usually in the story: did this person inherit a distrusted warehouse and leave with a single agreed definition of revenue? That is an adoption résumé wearing a data engineering title — and the specialists who have done it tend to still be employed somewhere doing it, which is exactly who an anonymized network surfaces.
The GTM and AEO version of the same mistake
Go-to-market has its own flavour of this, and it is worse because the measurement gap is structural. AI search now sends real demand, but attribution barely registers it: even after ChatGPT’s May 2026 update increased the number of citations, only about 2.5% of downstream visits carried a trackable AI-referral parameter. Effectively all of it arrives looking like direct or branded traffic.
Watch what that does to budget. If a channel is invisible in your dashboards, it cannot win an argument against a channel that reports cleanly. So teams keep funding the build work they can measure — more pages, more sequences, more enrichment logic, more automations — instead of the demand they cannot. The constraint is not production capacity. The constraint is that nobody has built the measurement and the internal narrative that would let the organization adopt a channel it cannot yet see.
That is an adoption hire: someone who stands up branded-search lift tracking, self-reported attribution, citation monitoring and a shared definition of what an AI-sourced opportunity looks like before anyone writes another landing page. It is the same role shape as a data governance lead, pointed at demand instead of at the warehouse. If you are building that function, the AI search network and the GTM engineering network sort for exactly this split — the people who can wire the system versus the people who can get the company to believe its output. Most briefs we see ask for the first and need the second.
How to tell which constraint you actually have
Four diagnostics, all answerable this week without a consultant.
1. Count usage, not delivery
Pull the last 20 things your team shipped. How many have a weekly active user who is not on the data team? If the answer is under half, more build capacity will not help you — you are already producing faster than the business consumes.
2. Ask for the number twice
Ask two departments for last quarter’s pipeline number. If they disagree, your bottleneck is semantics and ownership, not compute. Buying a bigger cluster makes both wrong answers arrive faster.
3. Track the reopen rate
What share of “done” requests come back as this doesn’t look right? A high reopen rate is a trust signal, and it maps directly onto that 46% figure on AI output — the work exists, the confidence does not.
4. Follow one decision backwards
Pick a decision a leader made last quarter and trace what informed it. If the trail ends at someone’s local spreadsheet rather than a governed asset, you have your answer, and it is not capacity.
What to do about it without stalling delivery
You do not stop building. You rebalance, and you make the rebalance explicit rather than hoping an existing team absorbs it between sprints.
Give adoption a named owner, not a shared responsibility. Governance that belongs to everyone belongs to no one — which is exactly how you end up in the 60% that fail on culture rather than funding. One person, with a mandate to retire assets and arbitrate definitions, outperforms a committee.
Fund the unglamorous 24%. Testing, observability and quality controls are the cheapest trust you will ever buy, and they are currently the least-prioritized AI use case in the field. Point your AI budget there for a quarter and watch the reopen rate.
Hire for the deletion quarter. Interview for whether a candidate has ever killed a widely-used-but-wrong report and survived it. That is the competence, and it is rarer than any pipeline skill.
Be honest about whether this is a hire or a project. Some adoption work is a standing function; some is a six-month cleanup that a blended team should carry. That trade-off is worth thinking through properly — we unpack it in GSI or in-house for blended delivery — because hiring a permanent governance lead to do a one-off migration is its own kind of capacity error.
Measure the channel before you scale the channel. On the GTM side, no more content or outbound volume until the measurement scaffolding exists. Otherwise you are adding supply to a system that cannot tell you whether any of it worked.
The uncomfortable part is that an adoption-first quarter looks slower on every dashboard your leadership currently reads. Fewer launches, more retirements, more meetings about definitions. Then the 12x shows up, because things finally reach production and stay there. That is the trade: you can buy throughput in an afternoon, or you can hire the people who make throughput matter.