Gleb Mezhanskiy spent years constructing instruments to make enterprise knowledge clear. In March 2026, the Datafold CEO instructed his viewers the trouble by no means paid off the best way software program monitoring did for firms like Datadog. His argument is now reshaping how knowledge leaders outline high quality going into the again half of 2026.
A Vendor CEO Calls the Business a Disappointment
Mezhanskiy laid out the case in a March 5 publish titled “Information Engineering in 2026: 12 Predictions.” Prediction eleven argues knowledge groups will cease chasing knowledge high quality as a result of AI brokers care about context as a substitute. Years of funding and engineering effort, he wrote, by no means produced a breakout success corresponding to Datadog’s rise in software program monitoring. Information high quality, in his telling, moved from a line merchandise on annual objectives to one thing groups deal with on a best-effort foundation.
The declare deserves a caveat earlier than it travels any additional. Mezhanskiy runs one firm in a crowded area, and his view displays a single vendor’s vantage level, not an trade consensus. Monte Carlo reported elevating $236 million, whereas Bigeye reported $73.5 million in whole funding. Datafold individually introduced a $20 million Sequence A. Collectively, the three firms disclosed not less than $329.5 million in funding, unfold throughout high quality, reliability, and observability slightly than one tidy class. The label of failure relies on which yardstick will get used, and Mezhanskiy picked a yardstick favoring his prediction.
Why Information Resists the Software program Playbook
Mezhanskiy’s strongest level has nothing to do with funding rounds. He argues knowledge is more durable to check than software program as a result of floor fact retains shifting. A login both succeeds or it doesn’t. An “energetic consumer” can imply three various things relying on whether or not advertising and marketing, product, or finance is asking, and no quantity of column-level testing settles a disagreement over definitions. Add extra alerts to a loud pipeline and the worth of every new alert drops quick.
The comparability holds up higher as a spectrum than a tough line. Safety groups chase ambiguous indicators every single day, and loads of AI-driven software program ships with outcomes no person can confirm with certainty. Software program floor fact just isn’t all the time as clear as Mezhanskiy’s framing suggests. What units knowledge aside is the dimensions of the paradox: a single warehouse can maintain dozens of conflicting definitions for a similar enterprise idea, and a software program group not often faces so many forks in the identical afternoon.
What AI Brokers Want
Right here is the a part of Mezhanskiy’s argument price taking severely even with the caveats connected. An agent pulling from a warehouse wants greater than a validated column. It wants lineage exhibiting the place a quantity got here from, the transformation logic behind it, documentation explaining why a fallback desk exists, and an ontology connecting enterprise entities like buyer, order, and product. Mezhanskiy calls the mix a context graph, and Datafold now sells one alongside its typical high quality instruments.
Value saying immediately: Mezhanskiy just isn’t a impartial narrator. His firm income if patrons shift spending from high quality monitoring towards the context layer his prediction describes, and the monetary stake doesn’t vanish simply because the underlying argument is cheap. The argument nonetheless has a restrict: context helps an agent interpret a quantity accurately, nevertheless it doesn’t flip a corrupt, stale, or biased quantity right into a protected one. Lineage tells an agent the place knowledge got here from, not whether or not the information deserved belief within the first place.
What Threat-Primarily based Information High quality Seems Like in Follow
Essentially the most helpful proof on this debate just isn’t a prediction. It’s a sample already exhibiting up in how groups construct knowledge contracts. The Open Information Contract Normal, printed by Bitol beneath Apache 2.0 by means of the LF AI and Information Basis, defines a vendor-neutral YAML format overlaying schemas, high quality guidelines, possession, assist channels, and repair ranges. Groups use the usual to formalize what a dataset guarantees, with out locking into one firm’s platform.
A handful of working habits separate groups getting worth from contracts from groups including paperwork:
- Put duty on the group producing a dataset, not the group consuming it three pipelines downstream.
- Retailer contracts as code in model management slightly than as a slide deck no person opens once more.
- Run checks in CI or within the pipeline itself, so a violation will get caught earlier than it reaches a dashboard or an agent.
Monte Carlo’s steerage tells clients to maintain contracts light-weight and aimed toward pipelines carrying actual enterprise weight, slightly than each desk within the warehouse. Soda and Atlan assist the identical sample by means of YAML information, Git workflows, and rule enforcement, and neither positions its tooling as a requirement: every treats automation as a comfort layered on high of a self-discipline a group may run with a textual content editor and a CI pipeline.
The Counterevidence
Market researchers don’t assist a collapse story for data-quality spending. Mordor Intelligence estimates the data-quality instruments market will develop from $3.27 billion in 2026 to $7.39 billion by 2031, a 17.7 p.c compound annual development fee. Deal with the determine as a industrial estimate slightly than an audited whole. Totally different analysis companies outline the class in a different way and land on numbers disagreeing with one another, which is regular for a market this fragmented and says extra about inconsistent definitions than concerning the underlying pattern.
The safer learn: spending retains rising whereas the definition of high quality will get broader. No proof reviewed for this piece reveals budgets shifting from high quality monitoring into context graphs. Joe Reis’s 2026 State of Information Engineering Survey, drawn from 1,101 practitioners over two weeks in late 2025, makes a easy level: high quality has not slipped down anybody’s record of worries. Thirty-four p.c of respondents named knowledge high quality or reliability as a significant drain on group time, and simply over ten p.c known as it their single greatest organizational bottleneck. Respondents skew senior and concentrated in North America and Europe, and Reis describes the odds as indicative slightly than definitive. Even with caveats, the survey factors towards groups stretched skinny by high quality work, not groups strolling away from it.
A Framework for Tiering Information High quality
The sensible transfer is to not choose a facet between Mezhanskiy’s prediction and the survey knowledge. It’s to cease treating each dataset prefer it deserves the identical stage of scrutiny. A four-tier mannequin offers knowledge leaders a place to begin for deciding the place strict contracts belong and the place lighter documentation will do.
Tier 0 covers income and regulatory important knowledge: billing programs, monetary reporting feeds, and compliance submissions. Every dataset right here will get a proper contract, automated checks working in CI, a named proprietor, and an on-call web page when one thing fails.
Tier 1 covers buyer and product important knowledge: dashboards clients see immediately, metrics executives report externally, and machine studying options feeding customer-facing predictions. Every dataset nonetheless will get a proper contract, with scheduled checks and an alert routed to an proprietor, although with out paging anybody at 2 a.m.
Tier 2 covers inner and operational knowledge: advert hoc reporting, inner analytics, and experimentation tables. Light-weight documentation and preserved lineage matter greater than a proper contract right here, since a mistake stays contained inside one group.
Tier 3 covers exploratory knowledge: one-off exports, scratch tables, and prototype datasets. No contract applies, no high quality assure exists, and every dataset carries a transparent label saying so.
Three questions place most datasets accurately.
Would a incorrect quantity set off monetary loss, authorized publicity, or a regulatory submitting drawback?
Tier 0. Does the dataset feed a customer-facing floor or a metric reported exterior the corporate?
Tier 1, except the monetary or regulatory publicity already pushed it to Tier 0. Does a couple of group depend on the dataset for selections, with none exterior or regulatory stakes connected?
Tier 2. Something left over, one-off exports and prototypes included, defaults to Tier 3.
As soon as a dataset earns a contract, the doc wants six fields, no matter format:
- Schema and knowledge sorts for each area a shopper would possibly contact, with nullable fields and anticipated ranges spelled out.
- Freshness and availability targets acknowledged as a quantity, not an outline: up to date inside 4 hours, obtainable 99.5 p.c of enterprise days.
- High quality thresholds and the checks imposing them: completeness, uniqueness, and any enterprise rule particular to the dataset.
- A named producer group, a named shopper group, and an escalation path for when the 2 disagree.
- A change administration course of describing how schema adjustments get introduced and the way lengthy shoppers get to adapt.
- A assist channel, acknowledged by identify, the place a shopper stories an issue and will get a response time dedication.
For an illustrative instance, image a subscription firm assigning its month-to-month recurring income desk to Tier 0. The six fields would possibly learn:
- Schema: customer_id (string, not null), mrr_amount (decimal, zero or larger), billing_period (date).
- Freshness: up to date inside 4 hours of every billing run.
- High quality checks: completeness at 99.9 p.c or larger, uniqueness enforced on customer_id plus billing_period.
- Possession: the Billing Platform group produces the desk, Finance Reporting consumes it, and disputes escalate to the on-call knowledge engineer inside quarter-hour.
- Change administration: schema adjustments get introduced two weeks forward within the #data-contracts channel.
- Assist: a named inbox commits to a response inside one enterprise day.
A scratch desk feeding a one-off cohort evaluation wants none of this. The price of writing six fields for each desk within the warehouse is strictly why most contract applications stall, and tiering exists to maintain the associated fee pointed on the knowledge the place it pays for itself.
Information leaders monitoring this mannequin ought to watch incident influence, detection time, false-alert quantity, and contract violations by tier, slightly than a single company-wide high quality rating hiding the place the actual injury occurs. A Tier 0 violation and a Tier 3 violation aren’t the identical occasion, and a dashboard treating them equally will bury the sign leaders want most.
The place Tiering Breaks Down
Two failure modes present up virtually instantly as soon as a group adopts a tier mannequin, and neither entails the framework itself.
The primary is tier creep. Each group believes its knowledge issues most, and a mannequin with no enforcement mechanism drifts towards labeling every thing Tier 0 inside a yr. The repair is procedural slightly than technical: route Tier 0 nominations by means of finance, authorized, or whichever perform owns the regulatory publicity, and require a acknowledged greenback determine or compliance quotation earlier than a dataset earns the label.
The second failure mode connects immediately again to Mezhanskiy’s argument. A tier task lives in an individual’s head or a wiki web page except somebody writes it into metadata an agent or a question engine can learn. An AI agent pulling from a warehouse has no technique to know a desk is a Tier 3 scratch dataset except the task travels with the desk itself, by means of tags, a catalog entry, or the context graph Mezhanskiy’s firm sells. Skip the tagging step, and an agent can pull from an unreviewed prototype desk to reply a query belonging to Tier 0. The outcome turns a low-stakes dataset right into a high-stakes choice, and no person notices till one thing breaks.
Tiers additionally want a evaluation cadence, since a dataset’s danger profile not often stays fastened. A cohort evaluation constructed for one board assembly can flip right into a recurring metric a CFO quotes externally inside two quarters, at which level the dataset has quietly moved from Tier 3 to Tier 1 with out a contract ever connected to it. A quarterly re-tiering evaluation, owned by whoever runs the information platform, catches the drift earlier than a metric goes exterior.
Information groups aren’t giving up on high quality. They’re admitting common protection was all the time a fiction, and the fiction grew dearer as soon as AI brokers began working on the identical warehouses people used to babysit by hand. Sturdy applications in 2026 will determine, in writing, the place unhealthy knowledge does actual injury, defend the bottom exhausting, and go away sufficient context behind for folks and machines to deal with every thing else with open eyes.
